As businesses incorporate machine learning into their commercial strategies, the race to innovate and stay ahead of the market has highlighted some new challenges.

Traditionally, software development (DevOps) and machine learning (MLOps) teams have operated with separate workflows, tools, and objectives. In today’s environment, this leads to inefficiencies and redundancies that can hamper software delivery.

What are the risks of sticking with the status quo? How can the two teams be bridged together?

Siloed pipelines create blockers

DevOps pipelines are built around continuous integration and delivery (CI/CD), emphasizing speed and reliability. MLOps, on the other hand, introduces stages such as data preparation, training models, and validation. When these operations are managed separately, the handoff from data science to engineering can be slow and error-prone. Data scientists may work in one environment, while engineers work in another, often requiring manual steps that disrupt the overall software lifecycle.

Different toolchains only exacerbate the problem. DevOps and MLOps both require automation, reproducibility, and version control. However, keeping two systems running at the same time is a waste of resources when they’re designed to achieve the same goal.

Channel providers who serve these infrastructures typically have to deal with many of these cases, adding complexity without delivering extra value. Silos between teams further complicate matters, with broken communication and misaligned objectives.

Unlike traditional code, ML models often rely on dynamic, data-driven outputs that can change depending on their training data or hyperparameters used. As a result, they don’t always fit neatly into existing DevOps pipelines, meaning standard testing, validation, or security checks can be skipped or inconsistently applied.

These problems increase the time it takes to get AI-powered features to market. Limited traceability of model versions, training data, and hyperparameters makes troubleshooting and auditing cumbersome, raising concerns about governance, compliance, and accountability.

The case for unification

The solution, as many organizations are discovering, is to combine DevOps and MLOps into a single, cohesive software supply chain. This approach doesn’t overlook the unique requirements of machine learning, though. Indeed, it treats artificial intelligence (AI) as if it were any other software component, creating a system of consistent protocols, whether for code snippets or trained models.

DevOps and MLOps share some of the same goals: rapid delivery, automation, and reliability. Aligning around these goals helps organizations and channel partners to operate more efficiently, reduce redundant work, and foster better collaboration.

The way to achieve true unification is to treat ML models as first-class software artifacts. Like binaries, libraries, and configuration files, models should be versioned, tested, and distributed through the same automated pipelines. This ensures unified visibility, so teams can keep track of which model version aligns with which release, reducing confusion and ensuring reproducibility.

Integrating models into these workflows extends automation across the entire lifecycle, from preparing data to deploying it, which cuts down on human handoffs and speeds up delivery from start to finish.

This approach also improves collaboration between data scientists, engineers, and operations teams. Sharing infrastructure and using the same processes makes communication simpler and allows for smoother handoffs.

Governance is also strengthened by subjecting ML models to the same quality assurance, security scanning, and compliance checks as other software components. For channel partners tasked with safeguarding software supply chains, this consistency is essential.

Opportunities for the channel

For the IT channel, bringing DevOps and MLOps together is both a challenge and an opportunity.

Organizations want to use AI, but they often lack the skills or infrastructure to do so. Partners who help customers set up these pipelines enable them to deliver faster, more reliable, and compliant solutions. Channel providers are at the vanguard of AI-driven innovation when they bridge the gap between DevOps and MLOps.

Companies need to be able to quickly and safely migrate models from testing to production to realize AI’s potential. There has to be a single software supply chain where ML models are handled like first-class assets and workflows are automated from start to finish. For channel partners, this method helps customers grow their AI projects while making sure that quality, security, and governance are maintained throughout the software lifecycle.

As organisations race to adopt more software and models, the industry needs holistic governance. Currently, only 60% of companies have full visibility into software that is running in production. Combining DevOps and MLOps into one software supply chain can help companies achieve their shared goals of rapid delivery, automation, and reliability. This will create an efficient and secure environment for building, testing, and deploying the entire spectrum of software, from application code to machine learning models.

2024 could see a concerted shift toward a 'local' machine learning development and training approach in a bid to cut prohibitive costs, according to a leading industry figure. 

With 2023 dominated by headlines on powerful large language models (LLMs) and the steep energy and cost requirements needed for AI development, Hugging Face CTO Julien Chaumond suggested “local machine learning” could become an emerging trend in the year ahead.

Chaumond’s comments came in a year’s end prediction post on LinkedIn in which he outlined a key area to watch.

“Local ML is going to be huge,” he said. “It will be in part driven by the adoption of Apple Silicon and other innovative hardware, but also on raw CPU and mobile devices.”

“In many cases except for the largest of LLMs, local inference will become a viable alternative to hosted inference.”

Chaumond’s comments regarding Apple Silicon specifically follow the recent launch of a batch of machine learning tools designed primarily for use on Apple hardware.

In December 2023, the tech giant quietly announced the launch of Apple MLX, a machine learning framework developed by its internal ML research group. The framework will enable Apple device users to harness in-house silicon for AI inferencing.

The framework is available through open source libraries such as PyPI and GitHub, and could represent a step change in how developers build AI tools and platforms.

Ben Wood, chief analyst and CMO at CCS Insight echoed Chaumond’s comments on local machine learning inferencing, adding that he expects this to be an “essential step” for developers moving forward.

A key factor here, he told ITPro, will be lingering privacy concerns surrounding the development and deployment of AI tools and services. Organizations such as Apple appear to be framing a more localized approach to inferencing as a vital method that curtails the risk of data exposure either during model training practices or by active exploitation.

In early 2023, OpenAI-owned ChatGPT came under intense scrutiny due to a flaw that exposed user conversations. A slew of organizations globally implemented rules to prevent staff from using the platform due to the risk of data exposure.

“Privacy will be a major focus with companies such as Apple extolling the virtues of using on-device AI to keep sensitive data on the device rather than having to send it to the cloud,” he explained.

Local machine learning could cut costs

AI training and development costs have been a key hurdle for many organizations over the last year, with analysis in May 2023 showing that the task of training an LLM such as GPT-3 could surpass $4 million.

Research from CCS Insight in October, for example, warned that generative AI could experience a “cold shower” in 2024 due to these prohibitive development costs.

Wood suggested that developers could circumvent cumbersome costs and computing requirements through on-device training methods.

“The intense computing requirements for cloud-based AI will mean the cost of deployment will be a factor in the shift towards more on-device AI processing,” he said.

In recent weeks, Google has emerged as a leading champion of on-device AI capabilities, Wood added, with the announcement that its Gemini Nano model will be available on certain devices.

This has been specifically designed for on-device tasks, he added, and marks a “surefire sign” that the tech giant views the practice as the way forward.

Major industry players such as Google and Apple aren’t alone in this sharpened focus toward local, on-device machine learning inferencing. Pointing toward Qualcomm, Wood noted that the manufacturer has been “working hard” to support on-device AI through its latest Snapdragon platform generation.

Samsung, too, is rumored to be exploring such capabilities on its upcoming Galaxy S24 Ultra smartphone.

This, Wood said, is tipped to be powered by Qualcomm silicon and will have “numerous on-device AI functions”.

Machine learning (ML) has been a field of research for more than 50 years, and as a subset of artificial intelligence has also been the focus of great innovation in that time. 

Many businesses use machine learning throughout their stack and some will have relied upon ML frameworks without realizing it. With the rise in popularity of newer forms of AI developments such as generative AI, however, some have questioned the extent to which traditional approaches such as machine learning algorithms still have a place in the tech industry. 

In this episode Jane speaks to Sascha Heyer, senior machine learning engineer at DoiT, to explore whether ML still has a role to play in a world that is more interested in conversational AI.

Highlights

“The future is always hard to predict, no one was expecting ChatGPT and GPT in general and I don't know what's happening next year. But I think we will have some more traditional machine learning projects for quite some time, especially because you can combine those traditional machine learning approaches with natural language model approaches.”

“Without large language models engineers had to consider various models, from simple linear regression to more complex ensemble models or deep learning models. And they had to adjust them, according to their use cases, which required a lot of manual steps.”

“If you're just going a couple of months back before ChatGPT and then after ChatGPT it’s a completely different world on how you solve challenges with machine learning. But that doesn't mean that traditional machine learning is obsolete. It still excels when a data set is small, use cases easy, interpretability is crucial, or when the use case is too individual to be solved with large language models.”

Footnotes

• What is machine learning and why is it important?
• Machine learning vs data science: What’s the difference?
• Machine learning vs deep learning vs neural networks: What’s the difference?
• Machine learning vs AI vs NLP: What are the differences?
• Seven reasons why machine learning is a good career
• Machine learning vs statistics: What’s the difference?
• What is generative artificial intelligence (AI)?
• What is ChatGPT and what does it mean for businesses?
• AI in software engineering – Six ways the profession is changing

Subscribe

• Subscribe to The IT Pro Podcast on Apple Podcasts
• Subscribe to The IT Pro Podcast on Spotify
• Subscribe to the IT Pro newsletter
• Join us on LinkedIn

Apple has launched a series of open source tools as part of a move to provide machine learning capabilities on Apple Silicon. 

Known as Apple MLX, the machine learning framework has been developed by the tech giant’s ML research division and will enable users to build AI tools harnessing its in-house silicon.

The framework is available through open source libraries such as PyPI and GitHub, Apple confirmed.

“MLX is designed by machine learning researchers for machine learning researchers,” Apple said on GitHub. “The framework is intended to be user-friendly, but still efficient to train and deploy models.”

“The design of the framework itself is also conceptually simple. We intend to make it easy for researchers to extend and improve MLX with the goal of quickly exploring new ideas.”

Apple said the design of MLX is “inspired by frameworks like PyTorch, Jax, and ArrayFile”, yet boasts notable differences. Key among these is the use of its unified memory model, the firm said, with arrays hosted in shared memory.

This means operations on MLX arrays can be “performed on any of the support device types without performing data copies”.

The MLX examples repo, accessible via GitHub, enables a range of capabilities, including the ability to conduct language model training, large-scale text generation using Meta’s Llama model, and speech recognition via OpenAI’s whisper model.

Users can also generate images with Stable Diffusion, the firm noted.

In a post on X, Apple ML researcher Awni Hannun said MLX is designed specifically for use on Apple Silicon, such as MacBooks, pointing toward the company’s potential goal of broadening access to tools by harnessing its own silicon.

Apple MLX: Key features for users

Key features of Apple MLX include the ability to leverage familiar APIs and dynamic graph construction, as well as multi-device capabilities. 

“MLX has a Python API that closely follows NumPy,” the firm’s GitHub repo explains. “MLX also has a fully featured C++ API, which closely mirrors the Python API.”

Dynamic graph construction has also been simplified, Apple said, providing developers with more optimized capabilities to streamline debugging.

“Computation graphs in MLX are built dynamically. Changing the shapes of function arguments does not trigger slow compilations, and debugging is simple and intuitive.”

Apple’s discreet AI movements

The roll-out of Apple MLX coincided with the launch of Google’s flagship Gemini AI model, which appears to have stolen the limelight. 

However, this move does show the tech giant has been quietly innovating on the AI front amid the furor of the generative AI explosion over the last year.

In stark contrast to Google, Microsoft, and other industry heavyweights, Apple has been keeping its cards close to its chest during this period.

Reports emerged in July 2023 that the company was developing its own internal large language model (LLM), dubbed ‘Apple GPT’, but as of yet nothing has been confirmed.

CEO Tim Cook has been relatively subdued on the topic also. However, in an August interview with CNBC said the tech giant does intend to integrate generative AI capabilities across its product offerings.

“We view AI and ML as fundamental core technologies. And they are virtually embedded in every product that we build,” he said.

Repeated speculation over the company’s approach to AI has been further fueled by the fact that Q3 earnings showed the firm increased R&D spending by roughly $3 billion up to that point compared to the same period in 2022.

In August, ITPro reported that Apple has been keen to emphasize the heightened need for locally-run AI models by leveraging its in-house silicon capabilities.

IBM has announced a host of new enhancements for its watsonx platform, which include new generative AI foundation models and new capabilities. 

Designed to help enterprises scale and accelerate the impact of artificial intelligence, watsonx is IBM’s AI and data platform that combines AI capabilities with analytical software.

The offering’s upcoming enhancements include a technical preview for watsonx.governance, new generative AI data services coming to watsonx.data, as well as a planned integration of watsonx.ai foundation models across select software and products.

“As demonstrated by the ongoing rollout of the watsonx platform within just a few months since launch, we are here to support clients through the entire AI lifecycle” said Dinesh Nirmal, senior vice president of Products at IBM Software. 

“As a transformation partner, IBM is collaborating with clients to help them scale AI in a trustworthy way – from helping to institute foundational elements of their data strategies to tuning models for their specific business use cases to helping them govern models beyond that.”

Generative AI models

The new IBM and third party generative AI models coming to watsonx.ai include its Granite series models, planned for later this month. These use the Decoder architecture of current large language models (LLMs) to predict the next word in a sequence, and can support enterprise natural language processing (NLP) tasks such as summarization, content generation, and insight extraction. 

IBM said it plans to provide a list of the sources of data as well as a description of the data processing and filtering steps that were performed to produce the training data for the Granite series of models – with estimated availability in Q3 2023.

As for third-party models, the company is also offering Meta’s Llama 2-chat 70 billion parameter model, as well as the StarCoder LLM for code generation in watsonx.ai on IBM cloud.

IBM said it has established a training process for its foundation models that’s centred on principles of trust and transparency for enabling responsible deployments of models and applications.

New watsonx capabilities

Freshly revealed capabilities for watsonx.ai include the first iteration of IBM’s Tuning studio. Expected to be available in Q3, the tool will include prompt tuning for adapting foundation models to unique downstream tasks. 

Available now, there’s also a new synthetic data generator for assisting in the creation of artificial tabular data sets from custom data schemas or internal data sets for extracting insights for AI model training.

IBM revealed it plans to “infuse” watsonx.ai capabilities into watsonx.data to help users process data for AI through a “self-service experience” that’s powered by a conversational natural language interface. The company is also planning to integrate a vector database capability into watsonx.data to support watsonx.ai retrieval augmented generation use cases. Tech previews are planned for Q4.

The firm is also launching a tech preview for watsonx.governance, which will enable users to explore capabilities for automated collection and documentation of foundation model details and model risk governance capabilities. This will allow stakeholders to view relevant metrics in dashboards of their enterprise-wide AI workflows with approvals so humans are engaged at the right times, IBM said.

AI assistants

Additionally, the watsonx platform is gaining a new set of AI assistants. For application modernization, watsonx Code Assistant products - coming later this year – will use tailored foundation models to convert code and generate code recommendations for developers. 

Meanwhile, the new watsonx Assistant will help users deliver consistent and intelligent customer service solutions with conversational AI, IBM said, while watsonx Orchestrate aims to help HR professionals automate repetitive, high-friction tasks and back office processes via a conversational interface.

IBM added it plans to embed its watsonx.ai innovations across its hybrid cloud software and infrastructure products, including intelligent IT automation, and developer services for watsonx.

Developers will be able to try out many of these new capabilities and models at IBM’s TechXchange technical learning event in Las Vegas, from 11-14 September.

OpenAI has revealed a privacy flaw in its popular chatbot ChatGPT temporarily allowed users to see the conversation titles of other users in their chat history.

Users on Reddit and Twitter started reporting the bug on 20 March and shared screenshots indicating that their ChatGPT web history contained titles they didn't recognise.

While the contents of the chats do not appear to have been accessible while shared in this fashion, OpenAI pulled the chat history altogether while the bug was addressed.

On the same day, major ChatGPT outages were reported and those with access noted inconsistent service. OpenAI noted the outages on its status page, and restored service within hours of the initial reports

Sam Altman, CEO at OpenAI, tweeted that while the issue was “significant”, it has now been resolved.

Altman did not provide the name of the open source library in question, nor provide an exact percentage of affected users.

With millions of daily visitors, a privacy flaw affecting even a small percentage could have resulted in widespread data sharing, and Altman’s promised “technical postmortem” should address these concerns.

Each ‘chat’ a user has with ChatGPT is saved as its own instance in a user’s history, with a title decided based on the contents of the conversation.

On Monday, users found that their history contained titles pertaining to unfamiliar topics or functions, as well as titles written in other languages indicating that the flaw was a worldwide issue.

OpenAI may have to carefully outline its data protection policies and procedures, and reassure users that its open source supply chain is secure and will not lead to similar issues down the line

Some on Reddit have reported seeing other types of information, but did not provide verifiable evidence to back up these claims.

“I see someone else's phone number as the phone number tied to my account. I'm concerned but not concerned enough to quit the app,” stated one user.

Another alleged that they had signed up for ChatGPT Plus, the $20 (£16) per month subscription plan for the platform, under another email that had become linked to their account and as a result were not granted access to the service.

The bug came at a crucial time for the AI firm, which has just released its GPT-4 model. Users who are subscribed to ChatGPT Plus have access to the GPT-4 variant of the chatbot already, with OpenAI having promised “human-level performance”.

AI competitor Google launched its own chatbot Bard in the UK and US this week, with users able to sign up for open access through a waiting list.

After reportedly upending internal teams to compete with ChatGPT, Google has expressed hope that Bard, which is powered by an optimised version of its 540 billion-parameter large language model (LLM), LaMDA, will be quickly improved through user feedback.

Onlookers are already drawing comparisons between Bard and ChatGPT, which along with Microsoft’s GPT-4-powered Bing chatbot have emerged as the standout competitors of the new AI era.

While Microsoft and OpenAI have collaborated, Google has relied on its market dominance to buoy Bard’s chances.

The search giant has also been outspoken about the shortcomings of generative AI. In February, Alphabet chairman John Hennessy warned that Google had been “hesitant” to release Bard as it was still in development, and the blog post that announced Bard described it as an “experiment”.

In ITPro’s internal testing, the chatbot has operated similarly to earlier versions of Bing Chat, with fast response times for text generation but a tendency to engage in false or fictional outputs known as ‘hallucination’.

As the popularity and reach of generative AI grows, many are starting to sound alarm bells over the potential for misuse and exploitation. Models like ChatGPT can already generate text that can be passed off as human-generated, and AI-generated media continues to develop rapidly.

If threat actors use the chatbots of tomorrow for phishing emails, to write malware, or for misinformation – and there’s every indication they will – public and private entities must be equipped with tools to flag AI-generated content. The same also applies to detecting plagiaristic content made with chatbots, a rapidly-emerging concern amongst academics.

In January, OpenAI released its classifier, a tool intended to pick out text that was generated by large language models (LLM), whether GPT-4 or a similar programme. The problem? It’s highly inaccurate. OpenAI’s classifier could only label 26% of AI-written text as “likely AI-written” in tests, while it incorrectly identified human-written text as machine-generated 9% of the time. Strings under 1,000 characters are especially tricky to unpick, and the classifier can’t process languages other than English at nearly the same degree of accuracy.

It’s clear OpenAI believes there’s scope for these tools in academia, with the company noting it’s “an important point of discussion among educators” in a blog post. But businesses could also gain a great deal from tools of this kind.

What are the dangers of undetected AI content?

One immediate concern is identifying the sources generative AI tools use. Many popular systems are trained on models that scrape information from the internet. In a black box development scenario with little insight into the data, it can be difficult to reassure companies their intellectual property (IP) hasn’t been used to ‘create’ something else.

There are already concerns that AI threatens the livelihoods of artists, and if systems such as OpenAI’s deep learning model DALL·E 2 or the CompVis Group’s Stable Diffusion draw on licensed works for their output, then firms from all sectors might find it in their interests to establish a method for the granular analysis of AI content.

At PrivSec London, a panel of AI experts urged firms to adopt greater transparency over how their models work to avoid regulatory difficulties down the line. Panel host Tharishni Arumugam, global privacy technology and operations director at Aon noted that if firms aren't already doing so, now's the time to think about what goes into their AI contracts. Companies should, for example, make a decision over whether or not to have clauses around the use of their data for training AI models.

OpenAI addressed this concern with its recently-announced ChatGPT API, which doesn’t use the data it processes for training purposes unless a company opts in.

“Broadly, there will be a space for a set of technologies that interrogate generated results,” Artem Koren, co-founder and CPO at Sembly AI tells IT Pro. “Because IP compliance is not the only issue with generated results. There are also other things like inappropriate content, violent aggressive content, illegal content, incorrect content. AI today loves to lie, it’s really good at it.”

This capacity to lie isn’t just problematic for its potential role in generating pro-Russian propaganda, misinformation, and convincing lies that could damage a company’s reputation. Chatbots like ChatGPT are a source of worry due to their extremely wide user base, the ease with which one can access it, and the large volume of content it can generate at once.

The impending generative AI vulnerability nightmare

These factors could result in a large and well-meaning public using generative AI with damaging consequences. For example, code generated with large language models such as OpenAI’s Codex may contain logic errors that could cause damage to a business’ IT stack down the line. Stanford University researchers, for example, found developers using AI assistants introduced more security vulnerabilities into their code compared to those who wrote code from scratch. Those in the study who used AI assistants also disproportionately rated their code as more secure, indicating misplaced trust in the ability of such AI tools.

Even when errors in code don’t open up software to zero-day exploits, they can lead to disruptions to normal operations, as seen in Windows Defender’s false reporting over ransomware.

The widespread use of AI assistants could also open companies up to legal challenges given the potential for models to have been unlawfully trained on licensed content. Challenges are already underway, with GitHub Copilot being sued for “software piracy on an unprecedented scale,” in a charge that also contests the legality of OpenAI’s Codex model for generating programming languages.

Due to the ease with which AI-generated code can be obtained, it can be shared on forums, and quickly overwhelm moderators who have little way of reliably identifying whether code is written by humans or machines. This has led to the popular coding forum StackOverflow banning ChatGPT responses entirely.

“I think the reality is that eventually, it will not be possible to tell whether an AI generated something,” Koren elaborates. “Or, to put it another way, there will be large language models that specifically generate things in a way that you won't be able to tell that it was a model that generated it.”

With this in mind, the development window for creating tools to detect machine-generated code is fast closing. If models develop become sophisticated faster than countermeasures can be designed, businesses could find they have few mechanisms to identify potentially troublesome sources for code such as AI models.

Increased testing, or changing how we test?

Beyond legal concerns, there’s also widespread alarm over the potential for AI to plagiarise or too easily generate responses to exams and essay questions. Offensive Security has banned the use of ChatGPT in cyber security certification exams, stating the model does not allow for an accurate assessment of an applicant’s skill level.

Dr Usama Fayyad is the executive director of the Institute for Experiential AI at Northeastern University, which aims to develop responsible AI solutions and foster better practical understanding of AI models. He agrees, in future, there’s a “huge” need for tools that can detect AI generated content, but also contends academia will need to accept change.

“We've done a successful transition from slide rules to calculators, to computers to, hey, the whole search engine on your mobile world. In each of those cases, we adapted how we teach, how we train, and how we assess,” he tells IT Pro.

“The good news is the technology isn't magic, and people have already come up with quick tricks to detect; you can use GPT on GPT. All of those [tools] are red herrings in my mind. The real issue here is it's not going to be too hard to change the way we assess and the way we teach to adapt to the fact: hey, this thing is here, it helps people, so part of your smarts is how well do you leverage it?

“I think people will come up with ways to make it really hard and very obvious that you're using GPT.”

While detecting unlawful content is more of a black and white endeavour, Fayyad suggests that academic concerns over the potential for students to use models like ChatGPT to write essays and exams are the result of a culture shock. It necessitates a change in procedure, whether that’s exams or hiring processes.

“In job interviews, we need to get to a level where we can say, ok, here is the right way to assess and here is what you need to focus on. Because this other stuff, even if this candidate is not good at it… I can't multiply 27 digit numbers in my head, but I can use a calculator and that's not a skill.

“And I think people will elevate to that level and say, “hey, you can use these engines and then your proficiency, your ability to utilise them correctly becomes a skill that's highly valued. Much like what happened with the graphic arts when Adobe first became a big tool.”

Davio Larnout, CEO at AI firm Radix states there are two main camps in this debate. They’re torn between rejecting generative AI entirely for fears “we’ll all become stupid if we start using that” and embracing these tools.

“For me, [arguments against] don't have a lot of value, because we should embrace it. The way that ChatGPT works today, it makes mistakes and even requires you to be more critical. Your critical thinking needs to go a level up, because you need to read it, you need to understand it and then think “where might this go wrong?”

Here, Larnout and Fayyad are in agreement; tools to detect AI content may be key in the near future, but in the long term firms will have to adapt to the prevalence of generative AI and approach critcial assessment from a new perspective.

Text generators have been around for years, but they’ve finally gotten good. Where previous iterations required a heavy edit, the latest version of ChatGPT from OpenAI writes with better grammar than I – sorry, me. But both chatbots and generative artificial intelligence (AI) tools like the newly launched GPT-4 also lack that creative spark.

This is largely down to how they work. ChatGPT’s machine learning system, for example, was trained on 570GB of text from Wikipedia, the open internet and books. Churning through those 300 billion words taught the system how sentences are structured, but it still doesn’t understand what it’s saying – and that means the outputted paragraphs read naturally, but might not be accurate.

While AI models can be trained on vast amounts of data and can recognise patterns and generate text that is coherent and coherently styled, they may not fully grasp the deeper meaning or implications of the words they are using. Additionally, AI is not capable of creativity in the same way that humans are. It can generate text based on patterns it has seen in data, but it cannot come up with new ideas or perspectives in the way that a human writer can. This can lead to a lack of originality and depth in the writing.

No wonder, then, that teachers are concerned, with the New York City Department of Education even banning access to ChatGPT via its network and an Australian school returning to in-person, paper-and-pencil exams. School is the one place where people are judged by their writing on well-trodden subjects. The answers to questions about themes in 19th-century novels and symbolism in Shakespeare are all already online, so ChatGPT can easily reproduce them. So, of course, can students.

It’s easy to buy pre-written essays, find example texts to nick bits from, or otherwise get answers on the internet – chuck “themes Pride & Prejudice” into Google and you’re set. Reading SparkNotes, Wikipedia or other literature explainers isn’t cheating; as any student should know, copying and pasting is plagiarism, but reading, thinking about the information, before paraphrasing, is the actual job.

For some students, writing in their own words is the hardest part, but there’s plenty of algorithmic assistance beyond Google’s search to lend a hand. Spell checkers are a lot smarter than they used to be, while tools such as Grammarly spot not just mistakes but awkward phrasing, helping to improve your writing as you bash your keyboard.

Teachers too have tools to hand to stave off AI cheating, beyond blocking websites on campus networks or forcing students to write pre-digital style in classrooms. Plagiarism checkers can spot some AI-generated text, as it’s largely been lifted from the internet and contains noticeable patterns and repetition – and that’s according to ChatGPT itself, as I asked it for advice hunting down text it’s written.

I also asked ChatGPT to write my column for me. Sadly my editor wasn’t keen on paying me to copy and paste several hundred words, but ChatGPT’s argument for why AI isn’t a good way to write a column was actually sound, but the system lacks nuance and can’t understand context. Plus, it’s unethical, as it puts me out of a job, and can’t generate new ideas.

Lastly, this AI isn’t thinking, it’s regurgitating. That we ask students to do the same so often that it can be replicated by AI is a fair criticism of our curriculum.

By introducing more randomness, AI can be used to smash ideas together to come up with something new. That’s what Janelle Shane does with AI Weirdness, a fabulous blog and newsletter that reveals plenty about the inner working of AI systems while also coming up with paint colours such as “turdly” and recipes for a chocolate brownie that includes a cup of horseradish.

Randomly placing letters one after the other into an approximation of a word, though, while often hilarious, isn’t actual creativity on the part of AI. Humans have a capacity for originality that machine learning lacks. AI can only reassemble what we’ve already done.

My favourite example is from Mark O’Connell’s annoyingly brilliant book To Be a Machine, where an unnamed futurist suggests to the author that much journalism can be easily replaced with AI. In retribution, O’Connell includes in his book a description of the futurist “retrieving a dropped pistachio from inside his expensive shirt – an act of petty and futile vengeance, and the kind of absurd irrelevance that would certainly be beneath the dignity and professional discipline of an automated writing AI.”

Of course, ChatGPT can include that example if it’s been trained on O’Connell’s book, though it can’t come up with its own similar absurdities. But just like the AI, I’m borrowing from others’ writing to pull together my argument, so while writers like O’Connell are perhaps irreplaceable, perhaps I’m not. That’s a tough conclusion to reach regarding my life’s work, so prove me wrong: spot the paragraph in this column that was written by ChatGPT. There’s no prize, but if you’re correct it might help shift this sinking feeling that I should retrain as a yoga instructor or park ranger or other job less easily replaced by AI.

On September 26 last year, without most of us noticing, humanity’s long-term odds of survival became ever so slightly better. Some 11,000,000 km away, a NASA spacecraft was deliberately smashed into the minor planet-moon, Dimorphos, and successfully changed its direction of travel. It was an important proof of concept that showed that if we’re ever in danger of being wiped out by an asteroid, we might be able to stop it from hitting the Earth.

But what if the existential threat we need to worry about isn’t Deep Impact but Terminator? Despite years of efforts from professionals and researchers to quash any and all comparisons with apocalyptic science fiction and real-world artificial intelligence (AI), the threat of this technology going rogue and posing a serious threat to survival isn’t just for Hollywood movies. As crazy as it sounds, this is increasingly a threat that serious thinkers worry about.

AI development is akin to survival of the fittest

“AI could pose a threat to humanity's future if it has certain ingredients, which would be superhuman intelligence, extensive autonomy, some resources and novel technology,” says Ryan Carey, a research fellow in AI safety at Oxford University’s Future of Humanity Institute. “If you placed a handful of modern humans, with some resources and some modern technology into a world with eight billion members of some non-human primate species, that could pose a threat to the future of that species.”

The idea of treating AI as its own species, potentially replacing another, is consistent with evolution by natural selection. “There's no reason why AI couldn't be smarter than us and generally more capable and more fit in an evolutionary sense,” adds David Krueger, an assistant professor in Cambridge University’s Computational and Biological Learning Lab.

“Look at how humans have reshaped the Earth and changed the environment leading to the extinction of many other species,” he explains. “If [AI] becomes the more significant re-shaper of the environment than humans have been, then it could make the environment of the world something that's very difficult or impossible for humans to survive on.”

Today’s AI might not be ready to take on its human masters – it’s too busy identifying faces in our photos or transcribing our Alexa commands. But what makes the existential risk hypothesis scary is there is a compelling logic to it.

The three ingredients to AI dominance

1. The logic of AI

The first ingredient is to do with how machine learning systems “optimise” for specific objectives, but lack the “common sense” that (most) humans possess. “No matter how you specify what you want from the system, there's always loopholes in the specification,” says Krueger. “And when you just create a system that optimises for that specification as hard as it can, without any common sense interpretation, then you're going to get usually really weird and undesirable behaviour.”

In other words, an AI tasked with something seemingly simple could have unintended consequences that we can’t predict, because of the way AI ‘thinks’ about solving problems. If, for example, it was told that the number of books in a child’s home correlates with university admissions, it wouldn’t read to the child more – it would start piling up boxes of books instead. Even if we think we’re designing AI that operates only within very strict rules, unusual behaviour can still occur.

2. Over-optimisation

The second ingredient is how an AI with one specific task will naturally engage in “power seeking” behaviour in order to achieve it. “Let's say you have an algorithm where you can, by hand, compute the next digit of Pi, and that takes you a few seconds for every digit,” says Krueger. “It would probably be better to do this with a computer, and probably be better to do this with the most powerful computer that you can find. So, then you might say, ‘actually, maybe what I need to do is make a ton of money, so that I can build a custom computer that can do this more quickly, efficiently and reliably’.”

An AI might think the same way if it has such a singular goal. An AI optimised for a specific task will use every tool at its disposal to achieve that goal. “There's no natural end to that cycle of reasoning,” says Krueger. “If you're doing that kind of goal-directed reasoning, you naturally end up seeking more and more power and resources, especially when you have a long-term goal.”

3. Handing over power

Krueger’s third ingredient answers the obvious question: why would we humans ever give an AI system the ability to engage in this maniacal ‘power seeking’ behaviour? Ultimately, even if we have a clever AI, won’t it be stuck inside a computer?

“The reason why you might expect people to build these kinds of systems is because there will be some sort of trade-off between how safe the system is and how effective it is,” says Krueger. “If you give the system more autonomy, more power, or the ability to reshape the world in more ways, then it'll be more capable.”

It’s not like this isn’t already happening – we let AI software trade stocks or navigate traffic without any human intervention, because AI is much more useful when it is directly in charge of things.

“You can build a really safe system if it basically doesn't do anything,” Krueger continues. “But if you give it more and more ability to directly affect the world… if you let it directly take actions that affect the real world or that affect the digital world, that gives it more power and will make it more able to accomplish your goals more effectively.”

Cooking up an AI-powered disaster

Putting these ingredients together leads to one possible outcome: that your goals may not be the same as the AI’s goals. Because when we give an AI a goal, it may not be our actual goal. Theorists call this the “alignment” problem.

This was memorably captured in a thought experiment by the philosopher Nick Bostrom, who imagined an AI tasked with creating paperclips. With sufficient control of real-world systems, it might engage in power-seeking behaviour to maximise the number of paperclips it can manufacture.

The problem is, we don’t want to turn everything into paperclips – even though we told the machine to make paperclips, what we really wanted was merely to organise our papers. The machine isn’t blessed with that sort of common sense, though. It remains maniacally focused on one goal, and works to make itself more powerful to achieve it. Now substitute something benign such as paperclips for something scarier, such as designing drugs or weapons systems, and it becomes easier to imagine nightmare scenarios.

How worried should we be?

The risk of a paperclip-hungry AI wiping out human existence might seem fantastically low, but it does paint a picture of how a threatening AI could emerge by following the remorseless logic that underpins AI technologies today.

“If you have a very intelligent system, it could be used for social engineering of anyone using the internet, or you could develop extraordinary technologies like bioweapons, nanotechnology or smarter AI,” says Carey.

If we’re lucky, though, we might see the problem coming. “It could be more of a gradual escalation of resources, as you're giving an AI the keys to more and more important corporate or military institutions as it gets more capable,” he continues, perhaps thinking of how over the past several decades we’ve put more of our trust in AI and computer systems, because it makes things more useful.

But, of course, we might not be so lucky. Other theorists argue that if you have an AI that is capable of redesigning and improving itself, then the sophistication of AI could scale up very quickly indeed. “If that's the case, then we're in some trouble,” Carey adds.

So just how worried should we be? Scarily, Carey points to the work of philosopher Toby Ord, who wrote The Precipice in 2020, attempting to weigh the risks of various existential risks, such as asteroids, pandemics and nuclear war. In the book, he estimates a 10% chance of an AI-linked catastrophe in the next century, compared to only a 6% chance of all of the other risks combined.

“The logic for that is that these natural catastrophes would have gotten us over the past million years already, if it was going to be a volcano or something,” says Carey. “He thinks AI, if anything, is most likely to be the big one.”

Microsoft and Google are both making the generative AI core to their search engine propositions, with each company having made major announcements to this effect in February. But, Google’s approach so far has been hasty, suggesting a lack of confidence in its ability to fend off ChatGPT - something that could hurt its reputation in the long run.

Microsoft had been rumoured to be negotiating significant funding for OpenAI since November and has since announced the full integration of ChatGPT into Bing search in a future update.

The firm reportedly went to the extreme lengths of upending the work of internal teams in December, reassigning departments to focus on a response to the newfound competition in ChatGPT. The result is Bard, which was reportedly pulled out of development and fast-tracked for a public reveal by Sundar Pichai.

Whether the public unveiling of Bard in February followed the original timeline of the product or was hastened by the rapid progress OpenAI was making with ChatGPT, is unknown.

What we do know is that the launch of Bard was sudden and unexpected - wider reports speculated a May release date - and it's something Google's top executives have since alluded to.

Alphabet chairman John Hennessy has stated that prior to the change in strategy, Google's position was that Bard was too immature a product for public demonstration.

“I think Google was hesitant to productise this because it didn’t think it was really ready for a product yet, but, I think, as a demonstration vehicle, it’s a great piece of technology,” Hennessy said at a venture capital summit, as reported by CNBC.

OpenAI’s GPT-3 model had been thought ‘too dangerous’ for the public when first beta tested in 2020. The company spent the following years tweaking the model to nullify its tendency to produce hate speech.

It’s possible that Google has been putting out similar fires internally, and that Bard represents a marked improvement over previous models. But, a lack of public insight into this development cycle, as opposed to ChatGPT's controversial yet more transparent history, makes it difficult to benchmark Bard outside of its immediate accuracy.

Bard is only available to designated testers at the time of writing, but its cracks have famously already materialised.

Flaws in the system were telegraphed by Google itself, with the tool having provided false information on the James Webb Space Telescope in one of the announcement demos. Google’s stock plunged 7% in response, worth $100 billion (£83 billion) in market value.

If the market’s reaction seems overstated, it’s worth remembering that Google has spent over a decade as an industry leader in AI.

With its acquisition of Deepmind and the development of LaMDA and TensorFlow, for years the firm has made efforts to rebrand itself as an AI company first and foremost.

It’s significant, then, that in a showdown with the Microsoft-backed ChatGPT, Google has been the first to blink and seemingly exposed its own tech as substandard. For all its talk of taking time to perfect the system, reports stated that Google felt pushed to rush development to get the system into shape in time for the reveal.

As with all cutting-edge technology, there will always be one vendor that wins the race to launch. OpenAI won this round. However, Google's delay in realising the potential for generative AI, forced to scramble hurriedly as ChatGPT soared in popularity, potentially shows how difficult it was to predict that the technology would take off as it has done in the past three months. So much so that not even one of the world's AI powerhouses could do it.

Alternatively, it may indicate a lack of foresight within Google's ranks, raising questions about why its potential wasn't realised sooner.

In January, an NYT report stated that Google “will ‘recalibrate’ the level of risk it is willing to take when releasing the technology", prompting OpenAI CEO Sam Altman to claim on Twitter that Google is increasing its risk, and that OpenAI will not do the same.

In this sense, Google has suffered a double defeat at this important juncture in the history of AI.

Forced to sideline its own safety principles in order to outflank Microsoft, its prize has been to publicly display a product that doesn’t meet its own quality standards. In response to the criticism, senior figures in the company have sought to muddy the waters of discussion around generative AI entirely.

"This kind of artificial intelligence we're talking about right now can sometimes lead to something we call hallucination," Prabhakar Raghavan, senior vice president at Google and head of Google Search, told Welt am Sonntag.

"This then expresses itself in such a way that a machine provides a convincing but completely made-up answer.”

Google took a seat at the table attempting to sell its own alternative to ChatGPT. It’s now presenting a version of events in which both it and OpenAI are fighting against the common enemy of “hallucination”.

In backpedalling to this degree, Google has shown its weaker hand and will need to demonstrate measurable improvement in Bard to reinspire confidence in the future.

There is still a long way to go for both of these products, and Google has bet on its strong user base as a foundation to maintain its search engine dominance. This could put it in good stead over OpenAI, which despite Microsoft’s funds has been forced to launch a paid tier of ChatGPT in order to meet rising costs and server demand.

Neither Google nor Microsoft has set out firm general release dates for their respective AI upgrades. But, in setting the pace so far, Microsoft has hammered a significant dent in Google’s long-maintained title as the leading AI firm in the industry.

Data scientists and machine learning engineers are in hot demand as more companies look to leverage real-time data and uncover insights and trends that they can then use to gain an advantage in their industry.

The terms ‘data science’ and ‘machine learning’ may often be used interchangeably by those who aren’t experts in either field, but they are two completely separate disciplines.

In short, data science is used to turn large amounts of data into useful insights. Machine learning, on the other hand, falls under the umbrella category of AI and leans on the insights derived through data science to create projections or otherwise improve a system’s performance.

Machine learning vs data science: What are the similarities?

Machine learning and data science have the shared purpose of lifting insight from data. A data scientist will deploy statistical modeling and data vizualisation on this quest, generating structured findings from unstructured data.

Both may use algorithms, but machine learning relies entirely on algorithms to function. Modern data science projects often involve machine learning to speed up manual processes. 

However, despite the similarities and overlaps between these two fields, each has distinct characteristics and uses.

Machine learning vs data science: What are the differences?

While data science brings order to disordered data, or pulls order from it, machine learning aims to learn from data to inform future actions or predictions.

A core part of data science is data cleaning, in which erroneous data, anomalies, and non-defined values are corrected or deleted.

Machine learning relies on some level of data science, as its algorithms cannot properly use data to learn and make decisions unless it has been cleaned and proven to be accurate and reliable.

Which skills are required for data science and machine learning?

Data science requires a wide range of skills, including domain knowledge, programming, statistics, and data visualization.

A career in machine learning requires experience in computer science and mathematics. In particular, skills such as linear algebra, calculus, and probability theory are handy to be able to implement the algorithms machines will need to make predictions.

Both disciplines can come in use in a wide range of roles across various industries, including AI engineer, business intelligence analyst, data analyst and research scientist.

How are machine learning and data science used across different industries?

In technology, data science and machine learning are used by engineers to build intelligent systems and improve user experience (UX). For example, Google uses machine learning to improve its search results, and Amazon uses it to personalize product recommendations.

Financial institutions, on the other hand, would use data science and machine learning to detect fraudulent transactions, predict stock prices, and identify potential investment opportunities.

The technologies both have a clear role in healthcare. For example, machine learning algorithms can be used to identify patterns in medical images and make accurate diagnoses, while data science can be used to identify inefficiencies in medical processes and improve the management of a hospital environment.

Similarly, in retail, companies analyze customer data and make personalized product recommendations, while also working to optimize the supply chain and logistics. In transport, meanwhile, machine learning is used to optimize traffic flow, prejudice maintenance, and demand for a specific route.

Finally, data science and machine learning are widely used in research fields such as physics, genomics, biology, and environmental science to analyze and understand complex data sets.

What are some of the most common applications of data science?

There’s an array of use cases for strong data science skills across various organizations. Among these include the following.

1. Predictive maintenance for manufacturers

Data collated from sensors on a factory floor can be used to predict when a piece of machinery might break down and allow engineers to carry out maintenance before a fault occurs. Coupled with automation, this is the basis for smart manufacturing and smart ports.

2. Fraud detection

By analyzing transaction data, data scientists can build models to identify unusual behavior that may indicate fraudulent activity and help with fraud detection and prevention.

3. Market segmentation

Data scientists can segment a market and identify customer groups with similar characteristics based on enterprise data. This allows businesses to target their marketing efforts effectively.

4. Sentiment analysis

Analyzing text data by carrying out text mining can help a data scientist understand customers' perceptions of a particular product, service, and brand and help businesses make decisions accordingly.

What are the most common applications of machine learning?

Machine learning is used across the entire economy, with varying sizes and types of organizations able to leverage the approach to iteratively improve and automate business aims.

1. Forming the basis for more complex AI systems

Machine learning is used as the building blocks for widespread AI approaches such as generative AI, in which it is used to map the contextual web that connects text, images, video, and audio. At a basic level, this forms the basis for large language models (LLM) to accept user input and produce relevant outputs.

2. Image and speech recognition

Machine learning algorithms can be used to analyze images, such as for facial recognition in security systems or for computer vision systems that can help improve accessibility. 

Visual data is analyzed based on known visual markers, so that a system may recognize audio patterns such as speech or discern visually between objects and people.

3. Recommendation systems

Machine learning algorithms can be used to analyze user data and make personalized recommendations for products or content.

This has become a core part of the user experience across many online platforms, with most shopping and streaming websites using machine learning to make tailored recommendations based on user data.

4. Handling language

Machine learning models can be used for tasks like natural language processing (NLP), language translation, text summarization, sentiment analysis, and much more.

5. Autonomous vehicles

Machine learning can be used to iteratively train self-navigating vehicles, such as autonomous robots on factory floors or self-driving cars. In practice, machine learning is applied to sensor data, to make low-latency decisions at the edge to ensure vehicles stay within markers and don’t endanger people.

Businesses have sought different forms of automation and intelligent data processing for years, but generative AI has opened the floodgates of investment and enterprise interest.

Formed on the back of machine learning and natural language processing (NLP), which have been in the enterprise wheelhouse for many years, generative AI entered the public eye through the popularity of ChatGPT and has since come to dominate discussions in the tech sector.

The relatively new approach to AI is already being used to generate detailed text and image outputs using simple user input and is increasingly integrated within business environments to automate a range of menial tasks. 

Generative AI has sparked interest across the business world because of the degree to which it can be personalized. With the right approach, the technology can radically improve worker productivity and help companies provide customers with far more intuitive user experiences.

How does generative AI work?

Generative AI’ is a term that refers broadly to AI systems that are capable of producing outputs based on a prompt. These come in a few different forms that work in subtly different ways. The most popular models in use today rely on a complex artificial neural network (ANN) architecture known as a transformer.

In simple terms, transformers take a prompt and output a response based on statistics and the exhaustive training process to which they have been subjected.

Transformers convert inputs into context, by breaking words down into mathematical values that are used to inform a model’s output. For example, when a user inputs “Where is Microsoft’s HQ?” into a model, the transformer converts the words into ‘tokens’ of data which are combined to form a coherent ‘vector’ of context used to produce a statistically relevant output.

If the model in the example had been trained well and on a specific company’s data, it could pick out the contextual significant of the words ‘our company’ and ‘city’ to produce a relevant output i.e. “Microsoft is headquartered in Redmond, Washington”.

As opposed to the form of AI that attempts to categorise information, generative AI relies on modelling that tries to understand the dataset structure and generate examples that might relate, or match. The two main forms of neural networks at play here – GANs and transformers – work in slightly different ways. By and large the former is involved in creating visual and multimedia content from images and text, while the latter uses information on the internet to generate textual output.

What are some examples of generative AI?

Although generative AI became popular in 2022, experts have been working on the technology in theoretical and practical applications for many years.

Dutch researchers wrote about the philosophical underpinnings of generative AI as far back as 2012. Indium Software, too, released a white paper [PDF] less than three years ago which highlighted not just how generative AI could be used creatively, but also in high-friction workplaces like healthcare.

Transformer models rose to prominence through the research paper Attention Is All You Need and quickly became prized for the efficient and performant way in which they could be used to produce coherent AI output. This was a significant breakthrough, which fed directly into subsequent products at OpenAI, Google DeepMind, and other firms.

Examples of generative AI

The generative AI market has grown incredibly quickly, to the extent that there are now innumerable LLMs and tools that rely on the technology. 

For many businesses and consumers alike, the standout example will still be OpenAI’s ChatGPT. Microsoft users also now have access to Copilot, the company’s AI productivity assistant, across a wide range of apps in the 365 suite, as well as within Bing search.

Google’s Gemini AI family powers the company’s AI product offerings, including Google Cloud’s new Vertex AI Agents. AWS has also invested heavily in AI tools, having rolled out solutions such its chatbot for enterprise Amazon Q.

AI pair programmers such as Code Llama, Gemini Code Assist, or GitHub Copilot can provide code suggestions based on a company’s private codebase or make existing cod emore efficient. These tools can also analyze code to provide user-friendly explanations of its function and produce comments, or translate it from one programming language into another.

Open-source AI models are also growing in popularity. These include models from AI community Hugging Face, with models in the space rapidly approaching the sophistication of some proprietary models.

In recent months, however, experts have questioned whether these models are truly open, as developers continue to impose usage restrictions on some models in their licenses. For example, open models from developers Meta and Databricks come with clauses that prohibit their use by firms with more than 700 million active monthly users, without express permission from the developers.

The latest flagship models, including OpenAI’s GPT-4, Google’s Gemini 1.5 Pro, Meta’s Llama 3, and Anthropic’s Claude 3 are multimodal. This means they can process text, images, video, or audio as inputs and produce outputs in a variety of formats.

Gemini 1.5 Pro also uses an architecture known as ‘mixture-of-experts’ (MOE), in which multiple ANNs dubbed ‘experts’ are assigned inputs based on which expert will most effectively process the input to produce an output.

The MoE approach has been hailed as a cost-effective route for training compared to prohibitively expensive traditional methods. It allows for models to increase in size and effectiveness, especially for multimodal processing, and is likely to become more widely-used in future models.

What are the benefits of generative AI?

Part of the reason that generative AI has seen such a boom in popularity is its ease of use relative to its power and potential. Generative AI has opened the door to far more detailed responses to natural language inputs, with LLMs able to unpick meaning from user queries and provide informed responses on its own.

Through generative AI-powered chatbots, businesses can provide customers with personalized answers based on questions to improve the overall digital experience of their website.

Using generative AI, enterprises can automate manual tasks such as drafting text, collating data from across different sources, or identifying anomalous details in a file or image.

Summarization stands out as a major benefit of major AI, particularly when used across a company’s estate. The chief appeal of products such as Microsoft Copilot or Google’s Vertex Agents is that they can provide users with accurate answers to specific questions grounded in their company’s data. For example, a user could ask for a summary of an internal medical policy.

AI code generation is another major use case for generative AI. Using an AI pair programmer, developers can speed up their workflows and focus on problems that require expert attention rather than sinking time into simpler work. The capability for these tools to translate code from one language to another has massive applications for legacy codebases, which may be written in obscure languages like COBOL. 

Easy code translation could save companies time and money down the line or even prevent critical outages, as experts with firsthand experience of these old languages are harder and harder to find.

More recent developments have allowed generative AI models to be used for tasks such as live video analysis through computer vision, which has applications from accessibility in tech to more autonomous robots in a manufacturing environment.

What are the concerns surrounding generative AI?

Widespread use of generative AI has been matched with concerns over its potential risks and harms. From the earliest models in public use, it has been clear that AI has drawbacks such as ‘hallucinations’ – the term used to describe incorrect statements confidently reported by an LLM.

Hallucinations are just one problem with generative AI. Leaders will need to keep up to date on the latest issues and remedies in the space as they seek to integrate it within their tech stacks, not least because of an anticipated increase in regulatory oversight of AI similar to the GDPR.

Generative AI legal risks

The rapid growth of the generative AI market has led to widespread discussions over the importance of ethical AI. One of the most basic concerns when it comes to generative AI models is who owns the data used to train the models – with some developers already facing lawsuits from artists, writers, and publishing houses over the alleged use of copyrighted material for training LLMs.

This is the tip of the iceberg for the legal issues of generative AI and governments around the world are progressing AI legislation to control the risks and harms that AI could pose.

The AI Bill of Rights was drafted in the US in 2022, as a framework for shaping future regulation. But the US has fallen behind on AI since, having lagged behind the EU and UK's AI approach.

The EU AI Act seeks to regulate AI models – inclusive of generative AI – according to their assessed risk. Businesses will need to know how their AI is being used and have a good understanding of the data used to train the models they use, or face hefty fines. 

Generative AI job losses

AI-linked job cuts are a major point of concern for employees and the speed and sophistication of generative AI has fanned anxieties in this space.

In 2023, IBM’s CEO Arvind Krishna was forced to go into damage control mode after stating that generative AI would benefit productivity at firms at the expense of human roles. However, Microsoft found that more workers care about AI benefits than its job impacts in its 2023 Work Trend Index, emphasizing how the benefits of generative AI are being increasingly weighed up against its downsides.

Job losses due to generative AI are not a given. Leaders can pursue upskilling to prevent AI cuts, insulating workers with AI skills that keep them in their roles longer.

Generative AI security risks

Generative AI threats, those specifically linked to attackers misusing generative AI to launch more sophisticated attacks on victims, are a major focal point for security teams as the technology becomes widespread. Though the 

One risk of generative AI that’s already been widely discussed is the potential to exacerbate the rise of deepfakes. These images or videos are created in such a way that renders a lifelike imitation of a person, often a celebrity but also potentially a prominent business leader, in such a way that it can trick others.

Real-time deepfakes are now becoming a more serious threat, while voice cloning tools such as Microsoft’s VALL-E or OpenAI’s Voice Engine can reproduce realistic copies of people’s voices using less than a minute of sample audio. Both could be effective weapons in a new era of social engineering.

At the same time as attackers are figuring out how to use AI to enhance their attacks, AI cyber security is being used by defense teams to counter these threats in more sophisticated ways. Tools such as Microsoft Security Copilot or Gemini in Security Command Center can help identify and summarize threats, or suggest actions for effectively responding to threats.

While it’s tempting to see generative AI as a malign force, given the early chaos it has sewn in the creative industries and across parts of the economy, that doesn’t mean it can’t be tamed and channeled into productive use cases. That said, there are many questions that need to be addressed first, with regulation a hot topic at the moment.

Rarely does a month pass without fresh reports of a forward-looking and autonomous system poised to change our future. What we don’t often hear about is the increasing use of artificial intelligence (AI) to examine our past.

Historians, archaeologists, musicians and data scientists are deploying AI to reimagine and recreate historical moments. Like so many tales from the evolution of modern computing, success with AI is grounded in the values of collaboration, opportunity and experimentation.

There are immense human challenges in getting the best results from AI and there’s no magic bullet computing at work The challenges experts face require distinct solutions, whilst sharing striking amounts of commonality. Bias and ethics of restorative AI are also of widespread concern, as is how we should interpret and categorise such works.

Uncovering ancient Athenian secrets

Jonathan Prag, professor of ancient history at Oxford University, has always had a passion for computing. “I got into mapping and visual analysis, which led to trying to build a digital catalogue of all the inscriptions from ancient Sicily,” he says.

Prag is an epigraphist, specialising in the restoration of ancient Greek texts carved into stone. Over the centuries, many carvings have been smashed into fragments, some of which have never been recovered, leaving vast gaps between words. In 2018, Prag’s PhD student Thea Sommerschield and Google DeepMind’s Yannis Assaelto suggested using AI to speed up the laborious process of filling the gaps in ancient texts. “I just sat there and went, that’s cool! Can you do it?”

A successful AI project relies on high-quality source data to ‘learn’ and Prag’s team didn’t have any. “In the eighties, Hewlett Packard mobilised scholars to type up the published Greek inscriptions,” he says. “It’s horribly messy and because it went from beta code into Unicode at a certain point, it’s full of artifacts”.

Sommerschield painstakingly cleaned the data to give the team more than 100,000 texts, from which the team trained Pythia, its first AI machine, to hypothesise the missing words from the Greek texts. However, Pythia’s successor was already in development. “Ithaca pays attention to patterns in the text, with reference about the region that each text came from, and the proposed date,” says Prag.

Ithaca has already resolved a point of conjecture in Athenian texts regarding the Greek letter S, sigma. “The state abandoned the three-bar form of the sigma, which enabled you to put a bunch of texts on one side or the other of 445 BC,” Prag notes. This date marker had been considered gospel within historical scholarship, but the few dissenting voices must have cheered when Ithaca was let loose on the data. “We ran Ithaca over the original data and it came out with new dates, moving them down by about 30 years”.

This shift alters the interpretation of a key period of Athenian imperialism, Prag adds, and makes a big difference to our reading of Greek history.

Using neural networks to seek lost tombs

Dr Gino Caspari, from the University of Sydney, is an archaeologist who’s studied the burial mounds of the Scythians, an ancient tribe of nomadic warriors who lived in parts of Asia more than 3,000 years ago.

Collaborating with a colleague, they built a convolutional neural network (CNN) which used satellite imagery to identify circular structures, seeking lost Scythian tombs. “Arriving in the survey area a year later, I immediately saw how wrong I had been,” he says. “What I had assumed to be buried structures in fact stemmed from the locals corralling sheep overnight in circularly fenced areas”.

Caspari’s long journey to see someone else’s lambs was because of poor data fuelling the AI. “The limitation is the availability of high-resolution satellite data, which is too expensive for archaeological projects to afford,” he adds.

His recent work on tracking 3,000-year-old Native American settlements in southern America was completed with the commercial package, ArcGIS. “In archaeology, we are clearly not at the forefront of development in AI and the number of people actively working on AI is limited,” he says. “To find broader adoption, we will ultimately need a kind of intuitive GUI that allows you to train models without coding”.

It’s important to remember that despite what dystopian headlines may suggest, AI is simply a customisable tool. Dr Caspari combined AI, LiDAR images and multispectral data to find more settlements, including many situated further north than had ever been documented. Could this have been achieved without AI? “Yes, but it would have taken a lot more time,” he argues. “Due to small training datasets, we often have a high number of false positive detections and those still need to be weeded out by hand. We are not really reaching superhuman performance in most cases due to a lack of training data.”

Resurrecting lost films

In the garden of Oakwood Grange in Leeds, Yorkshire, in 1888, Louis Le Prince, the ‘father of cinematography’, is using his invention, the motion-picture camera, to film the Whitely family. Today, only 20 grainy frames remain of what we acknowledge as the oldest surviving film in existence, but this didn’t stop Denis Shiryaev from using AI to create something rather marvellous. “I’ve always been the history junkie, and I decided to apply my knowledge of AI,” he says.

Using the images posted on the Science Museum website, Shiryaev reanimated the stills, applying a CNN to add detail to the faces and upscale the resolution. In total, the tool generated 250 colourised and stable frames. “I am a creator for my own audience,” says Shiryaev. “AI colourisation is not real, and it’s not historically accurate. I generate faces based on old source photos or paintings, it’s also an approximation. I think it’s important to have this small disclaimer that this video is AI.”

His use of AI is not to establish historical accuracy, then, but to inject verisimilitude into old footage, refreshing it for a modern audience. Almost 65 million YouTube hits underline the interest. “I saw this popularity as an opportunity,” he says. “Some companies from Hollywood, huge brand names, contacted me”.

In 2020, Shiryaev launched neural.love, an automatic cloud-based AI service that makes media enhancement accessible to anyone. “We have this beautiful feature called Generate Portfolio,” he says. “You can upload really low-resolution old photos and generate high-quality portraits and people love it”.

Cleaning up old hits

A popular use of AI within the music industry is to enhance old recordings by removing noise and artefacts, often present since the day they were captured. In May 1972, singer Mary Hopkin performed at London’s Royal Festival Hall and a venue engineer captured her sublime performance. By 2005, the technical quality of the recording needed some assistance, as Mary’s son Morgan Visconti explains. “I think it was probably done hastily; I think it was quarter-inch tape,” he says. “The noise flow is heavy, and it gets worse as you got towards the end of the gig, there’s more noise than signal.”

Visconti is a musician and producer with a passion for technology. “I was chatting with my mum about noise removal and suggested she let me have a whack at it, and it was stunning. Just to hear it without the noise, it felt like a refresh, like you cleaned your ears out to hear the nuances. She was very pleased with it”.

Once again, experimentation and passion have honed Visconti’s skills after having a lightbulb moment with iZotope’s mastering suite. “This feature – music rebalance – was just an eye opener – or an ear opener,” says Visconti. “The ability to go into a mixed track and make stems for bass, vocals, drums, guitars and keyboards. It’s something I dreamt about as a kid. I went nuts, I took everything apart like old Beatles recordings, and I was like, ‘What can I do with this thing’.”

Visconti sees AI as just another tool available to a music producer and having grown up in music studios, he’s seen them all. His father is the legendary music producer, Tony Visconti, who asked Morgan to sprinkle the technical stardust for the soundtrack of Moonage Daydream, a new film about the life and career of David Bowie.

“The director used a ton of found footage and one clip is of Bowie playing Rock And Roll With Me and the mix isn’t great. Like Festival Hall, this was recorded straight to mono. I was able to provide separate tracks of vocals, drums, keyboards and bass and Tony was able to remix the song for the film.”

Restoring decaying works of art

AI remains an experimental technology, even for the largest tech companies. Emil Wallner is part of Google’s Arts & Culture lab. “We experiment with the cutting edge of technology and try to find interesting areas to apply it to,” he says. “Sometimes it doesn’t work and we don’t publish things, sometimes we have interesting results and we share them”.

Recently the lab has experimented on the works of the world-renowned artist, Gustav Klimt. “We go to a lot of museums, to digitise collections so anyone can access them online. At the Belvedere Museum in Vienna, the topic to start working with these Faculty paintings came up.”

Klimt’s controversial Faculty Paintings, Jurisprudenze, Philosophie & Medizin, were destroyed by fire in 1945 and could only be viewed via a handful of old monochromatic photos. To give Google’s AI some colour training data, Dr Franz Smola, curator at the Belvedere Museum and Klimt expert, stepped in. “Franz looked at what all the critics said about these paintings,” says Wallner.

Smola scoured written records and gathered paintings by similar period artists to reference their palettes and paint types. His months of painstaking research took Wallner’s team another six months to convert into usable machine data. “We would just enable it to add a few pixels and a few motifs that we knew the colours of. We add pixels in the machine learning model, and from there, it could colourise the entire painting.”

Wallner is clear that this extraordinary project is not a restoration, but a re-colourisation due to the lack of detail in the monochrome photos. “One area would be to recreate the pointillism,” he says. “He used a lot of gold which you can’t really see, so you need to somehow augment that in a 3D environment to capture the impression of the metallic elements in those paintings”.

Everyday developers and data scientists are leveraging the power of machine learning from AWS to solve the world’s biggest challenges. From connecting patients and caregivers to supporting sustainability, machine learning models are at the core of this evolution.

This eBook explores how this transformative technology is pioneering impactful solutions.

Download now to read how innovators are tapping into the best practices, in-depth expertise, and powerful machine learning solutions to launch new initiatives that are improving lives and protecting our planet right now.

Provided by

Organizations across industries and geographies have long realized the importance of artificial intelligence (AI) and machine learning (ML)—with 9 out of 10 businesses having invested in the technologies as of 2020. However, with only 15% deploying AI capabilities, it’s clear that implementation is one of the main barriers to success.

This research from Harvard Business Review and AWS provides the keys to overcoming the challenges that businesses commonly face when attempting to implement AI and ML.

Download now to learn:

• The keys to successful AI implementation
• How to recognize when it's not working
• How to find the right partner
• and How to use data responsibly and ethically

Provided by  AWS

Businesses can unlock significant value across the organisation with the help of artificial intelligence (AI) and machine learning (ML) if they follow the proven path to machine learning success.

This eBook will guide you on your ML journey and help you achieve measurable results at every stage along the way.

Download now to learn about:

• A successful machine learning journey in 6 steps
• How to transform ML investments into competitive advantages
• Inspiring stories of industry leaders who’ve successfully implemented machine learning
• Insights from AWS machine learning experts

Provided by  AWS

By providing the broadest and most complete set of machine learning (ML) services, AWS is able to meet its customers wherever they’re at in their ML journeys and help them achieve specific objectives.

This eBook shares why hundreds of thousands of organisations use AWS ML to resolve challenges and create new opportunities within their organisations.

Download now to learn how to:

• Improve customer experiences
• Optimise business operations
• Accelerate innovation

Provided by  AWS

Microsoft could soon invest as much as $10 billion in OpenAI, the company behind generative artificial intelligence (AI) models such as ChatGPT and DALL·E.

If the deal goes through, it would represent a major step by the tech giant to consolidate its place in the AI market, as well as to improve its existing products and services. Microsoft already invested $1 billion (£816.9 billion) in OpenAI in 2019, so has a stake in the company’s progression.

People familiar with the matter told Semafor that the industry giant was in talks over the deal with OpenAI, and would take a 49% stake in the company if the deal went through. It was also alleged that Microsoft would take 75% of all eventual OpenAI profits until the investment had been recouped.

OpenAI’s large language model (LLM) and services that utilise it, such as ChatGPT, are likely to be central to Microsoft’s potential acquisition strategy. Operating on billions of parameters at a high degree of sophistication, OpenAI has already developed a number of models such as DaVinci and Codex for the input and output of text and code.

Integrating and expanding these models under the Microsoft brand could put the company in a highly competitive position for the next decade, especially in the face of competitors such as Google’s Deepmind.

LLMs are already being used by firms such as Secureworks to improve operations behind the scenes, and aggregate data for cyber security oversight.

“Our counter threat unit is reading the dark web all the time and we're seeing them take huge blocks of Russian language text, have it automatically translated and summarised so that the speed that we're able to observe and monitor the dark web is improving,” said Michael Aiello, CTO at Secureworks, to IT Pro.

Aiello described code reviews as another area where LLMs can shine, but noted no company is currently deploying this at commercial scale.

"The codex model that Microsoft works with is interesting because you can give it open-ended code and you can say, 'please find a vulnerability with this code.' You can see that on from both the black hat and a white hat side."

If Microsoft could find a way of monetising these capabilities, or likewise applying ChatGPT’s text-editing functionality to business applications such as Teams and Word, the investment may pay for itself in rapid time.

It has also been reported that Microsoft is looking into integrating ChatGPT with Bing search to improve its results and make web searches more of a natural process.

For back-office automation, it’s clear that LLMs can be used to rapidly improve efficiency and free employees from manual assessment of text such as code. Microsoft could benefit from this in itself, seeking to apply the models to its own back-end code in an exclusive deal that could see it edge out competitors through efficiency improvements alone.

However, it’s also highly likely that Microsoft is looking to monetise features such as this by acquiring a significant stake in OpenAI.

“It is clear that this investment move by Microsoft is a competitive positioning grab to gain a commanding position in what is being hyped as a transformative market,” Bola Rotibi, chief of enterprise research at CCS Insight, told IT Pro.

“Microsoft has placed its bet early on a leading start-up player that has an illustrious set of backers [such as] Elon Musk, Peter Thiel etc, and growing popularity and footprint - 1 million subscribers in less than one week when it released ChatGPT 3 on 30 November 2022. More importantly, the company has recognised the heavy cloud computing resources required by the AI models from OpenAI and its peers.

“A valuable benefit for Microsoft investing in OpenAI at this scale is the option of providing its services through Azure Cloud down the line. The range of uses for ChatGPT in particular is vast, and offering these through a consolidated, consumption-based package could be worth billions more in the near future for Microsoft.

“Much of that investment will likely be used to pay for the Azure resources so Microsoft is on a bit of a win-win basis: cloud services payment and 75% of profit until its investment is paid back - a typical styled outcome-based model.”

The move could also provide funding that OpenAI desperately needs. In December 2022, OpenAI CEO Sam Altman tweeted “we will have to monetise it somehow at some point; the compute costs are eye-watering”.

While Altman declined to specify the exact costs associated with meeting user requests, some have made estimates.

“ChatGPT acquired 1 million users within its first five days of operation. If an average user has made ten queries per day, I think it’s reasonable to estimate that ChatGPT serves around 10 million queries per day,“ tweeted Tom Goldstein, associate professor in the department of computer science at the University of Maryland.

“I estimate the cost of running ChatGPT is $100,000 (£81,773) per day, or $3 million (£2.4 million) per month.”

It is clear that OpenAI’s financial situation is far from solid, and that Microsoft’s cut of profits would only become beneficial through a major change to the company’s business model. Tola Sargeant, managing director of UK analyst firm TechMarketView, told IT Pro that the deal could be a “gamble” and one far from guaranteed to pay off.

“The proposed $10 billion (£8.1 billion) investment by Microsoft remains unconfirmed and seems speculative at best, with no comment from the company directly.

“It would be a significant investment into a company that barely generates any revenue and is still largely unproven outside its recent public test, which was constrained to text input only and not connected to real-time data sources. Monetising the solutions provided by OpenAI, especially ChatGPT, is not going to be an easy prospect, and it is likely profit (if any) may well be a long time in the making.”

If it does acquire OpenAI, Microsoft will also have to weather complaints against the ethics of the platform, as well as ensure safeguards are in place to prevent the model from generating hate speech as it has in the past. OpenAI’s image generation model DALL·E has also been the subject of much criticism, with some arguing that AI art represents an existential threat to the work of creatives the world over.

Although it would be a risky investment, Microsoft is well situated to harness the potential of OpenAI's powerful models. LLMs could radically change how customers interact with Microsoft products and put them a generation ahead of alternatives. This is a long-term opportunity for the tech giant, and one that could shape the tech sector for decades to come.

Biotechnology giant BioNTech SE has agreed to acquire InstaDepp, a UK-based technology startup specialising in the field of artificial intelligence (AI).

BioNTech aims to improve its research with the use of AI models, such as those developed by InstaDeep, to enhance the testing of novel drug samples and improve the efficiency of lab work.

InstaDeep has a track record in analysing samples for BioNTech, and in addition to the medical use cases for its models has demonstrated an ability to improve operations such as vaccine supply chains.

The deal, worth £562 million, is divided into a £362 million payment in cash and BioNTech shares, along with £200 million in future milestone payments to InstaDeep shareholders.

“Since the inception of BioNTech, we have focused on leveraging computational solutions to create personalised immunotherapies that can reach a wide patient population,” said Prof. Ugur Sahin, M.D., CEO and co-founder of BioNTech.

“The acquisition of InstaDeep allows us to incorporate the rapidly evolving AI capabilities of the digital world into our technologies, research, drug discovery, manufacturing, and deployment processes. Our aim is to make BioNTech a technology company where AI is seamlessly integrated into all aspects of our work.”

BioNTech has previously invested in InstaDeep, and in 2020 announced a multi-year research partnership with the firm. This was realised through the formation of AI innovation labs in London and Mainz to combine the expertise of the two companies to improve pharmaceutical design, analytics, and supply chains with AI and ML.

What is InstaDeep?

InstaDeep is headquartered in London, and was founded in 2014 in Tunisia. At the time of writing the company has approximately 240 employees and has twice been included in CB Insight’s ‘most promising artificial intelligence startups’ report.

It is the creator of DeepChain AI, a product that analyses protein sequences using an AI model trained on amino acids. It can be used to generate mutation probabilities of a sequence, or to predict stable variants of a given protein sequence, which is particularly useful for biotechnology research.

Previous work by the company includes collaborations with Nvidia using its Cambridge-1 supercomputer to accelerate biological research and perform complex tasks like genome sequencing.

InstaDeep is one of a select few organisations with access to the supercomputer, alongside founding partners AstraZeneca, GSK, Guy's and St Thomas' NHS Foundation Trust, King's College London, and Oxford Nanopore.

It's also a certified 'Service and Build' partner in Google Cloud’s Partner Advantage programme. The firm also partnered with Google Cloud in 2022 to train an AI model that could predict the traits of specific plants based on genomic sequence analysis.

It was hoped that this model could be used to assess the viability of crops necessary for world food supply, and to recognise genes that make cropper hardier, more nutritious and easier to grow. Together, Google Cloud and InstaDeep trained models of up to 20 billion parameters on available crop genomes

Art produced by artificial intelligence (AI) models has quickly gained popularity throughout the year. Driven in part by its ease of use, and coming off the back of years of generative AI development, AI art has become a defining subject in the tech conversations of 2022.

The exact enterprise use cases of ChatGPT and DALL·E are still unclear and, at present, the models are in a kind of public testing phase as people and businesses explore the potential and limitations of AI art as it stands.

How does generative AI work?

A mainstay throughout the year has been DALL·E 2, OpenAI’s model for generating images based on a user prompt. DALL·E works by translating prompts into machine readable keywords, which are weighted against a constantly-updating model that recognises images.

To achieve this, it uses a model called contrastive language-image pre-training (CLIP), which has been trained on many millions of images to determine the vector spaces and dominant features of images. This means rather than simply knowing what an image is, CLIP allows the AI model to identify how much of the image any given prompt relates to in order to work out each image’s individual features and descriptors.

By reverse-engineering this same model, DALL·E can generate images with a given percentage dedicated to visuals it thinks the prompt implies. For example, the prompt ‘dog on a chair’ might inform DALL·E that some of the image – say, a third – has to contain a chair, as its large-model training has suggested. Through training with huge datasets and, after a long development process, OpenAI set these models loose to the public through their website, where one can generate a set number of images for free every month.

In addition to its increasing popularity for generating images, OpenAI’s chat generative pre-trained transformer (GPT) is able to generate tracts of text based on a prompt. For example: In 2022, AI art is becoming increasingly popular. Some AI art is created by algorithms that randomly generate images, while other AI art is created by humans who use machine learning to create art that is more sophisticated and lifelike. Some people believe that AI art is not really art, but others believe that AI art has the potential to be even more creative and expressive than traditional art.

Woodward it isn’t, but as this was generated in seconds based on the prompt 'write a paragraph about AI art in 2022’, it’s an example of how the model can be used to quickly produce text that passes a basic bar for grammar and covers any given topic.

OpenAI has been through many iterations of its GPT natural language processing (NLP) algorithms and the current GPT-3 model has previously been withheld from public use over safety fears. The next iteration, GPT-4, is currently in development.

Happy AI-powered holidays

At IT Pro, this got us thinking: How can we test this technology in a manner that’s both thorough and festive? The answer, of course, was to combine DALL·E with ChatGPT to create a unique set of tech-themed Christmas cards.

Starting with artwork, it immediately became clear DALL·E was struggling with the term ‘Christmas card’, even when introducing the concept of Christmas into the visuals sparked by other elements of the prompts.

A Christmas card themed around virtual reality

It usually processed this by plastering a garbled (and frequently cut off) form of word art that read ‘Merry Christmas’ over the top of whatever it'd generated, as with the particularly vibrant virtual reality (VR) card.

The ‘cloud storage’ card was a little more in line with the art you’d expect to find on a Christmas card, but also basically nonsense. A very “Happy Cloaday” to you too, DALL·E.

Cloud storage Christmas card

The same cannot be said for the ‘CIO’ card, which was exciting but intense, festooned with garbled festive messages albeit far from ‘Christmassy’.

CIO Christmas card

To mix things up a little, we tried out a more generic prompt, ‘Hacker dressed as Jack Frost’, with the hope this could be a cheeky image to include on the front of an IT security team’s Christmas cards. We even mocked up a Grinch version.

“Cheeky” is not how I would describe the result, which is more than a little disturbing and reveals the limited extent to which DALL·E can generate believable human forms. Perhaps this is better as a Halloween card — it’s certainly not fit for any mantelpiece.

Hacker dressed as Jack Frost

Generative AI is rooted in the uncanny valley

Bringing in specific style descriptors appeared to rein in the madness somewhat, as seen in the ‘satirical cartoon’ prompt, which is the closest to anything you’d find on a real Christmas card. Yes, the face on the data centre technician is scrambled, and the text is gibberish, but the art style carried across is recognisable enough.

Satirical cartoon of Santa helping out in a data centre

The same is true of the watercolour painting example. In all likelihood, this comes back to CLIP, with each specific element introduced to the prompt changing the balance to which DALL·E can go wild with the other elements in the image. As what I’m assuming was a happy accident, the watercolour brush strokes omitted the faces of all but one of the ‘elves’ in this example, which ended up looking more like a stylistic choice than machine error.

Watercolour painting of software engineers dressed as elves, working on laptops in Santa’s workshop while a CEO dressed as Santa gives them a thumbs up

The charcoal engineer prompt did a more than passable job of recreating the physical medium, and could certainly pass as a photo of a sketch, rather than digital art. It’s only upon closer inspection that the illusion of the medium falls apart, with each stroke far too smooth to have been made with charcoal. The only real gripe with this image is that it ignored the Christmas theme — the engineer should have been climbing the North Pole, but we’ll just have to use our imaginations for that.

Charcoal sketch of a 5G technician climbing the North Pole

Remembering this balance and the statistical models that lie behind systems like ChatGPT and DALL·E is key to setting your expectations. It's not magic, and even if prompts are carefully phrased it doesn't make a bit of difference if there’s no training to support whatever use case you have in mind.

ChatGPT writes a haiku

Take this example of a ‘haiku’ prompt.

Prompt: A Christmas card written in haiku form, containing the word "Kubernetes".

A haiku is a Japanese form of poetry comprising seventeen syllables in three lines of five, seven, and five. This is not a haiku. Beyond that, it exposes the system's inherent weaknesses, as it falls back on cheap tricks when faced with a tricky prompt. Abandoning the metre and form of a haiku entirely, the model wasn’t even able to believably introduce the word “Kubernetes” into the output, instead choosing to dump the word front and centre in the out-of-place first line.

ChatGPT writes a corporate Christmas message

What if you’re not looking to send your Christmas card to one person, but a whole company of employees?

Prompt: Write a Christmas message to be shared via company email to a small tech firm with 20 employees, reflecting on the hard year the company has had but with optimism for the new year.

No one could call this heartfelt, and neither will anyone to raise a toast to the company’s CEO at the Christmas party. If it’s a bit vague and cliched, though, it’s a message that could believably be shared around via email at the end of the year, and not immediately recognisable as an AI-penned message.

For generative AI, results lie in the use cases

Considering the use case carefully applies to images too. Images the system can generate work best at a tiny scale. Take a step back from your screen, and some of the more passable images look ‘real’ enough. Look at them too closely, though, and you notice all the unsettling features that have become a trademark of AI art.

Take the ‘computer motherboard with Christmas lights’ prompt for example. If you’re reading this article on your phone, the illusion of the image may have held up longer for those viewing this on a monitor. Go back and zoom in on any of the fairy lights — or the circuitry of the motherboard itself — and you’ll see it’s really a mishmash of warped lines, hard edges and blurry stand-ins for detail.

A computer motherboard with Christmas lights

This also applies to the ‘charcoal sketch’, which looks roughly like charcoal from a distance, but up close appears plasticky, and instantly recognisable as a digital representation of the physical medium. Because AI image generation of this kind relies on ‘diffusion models’, in which random noise is edited pixel-by-pixel until it resembles the prompt given, DALL·E images tend to be wobbly on close inspection and fall apart under close scrutiny. This manifests most obviously in the programme's efforts at faces, which veer firmly into the ‘uncanny valley’.

Will generative AI replace creatives in the future?

We're not at all convinced AI has the necessary skill to supplant human artists. As we see, the models have a long way to go before they can begin to match the quality of human artists, and even the best outputs, a plethora of which can be found online, have strange flaws.

There are also strong ethical arguments against the widespread use of generative AI to create art, which include the damage it would cause to the value of creatives’ labour, and the legal grey area in which scraping work created by other people to train an AI model lies. For these reasons, AI art should not be seen as a revolution in the art world, nor a welcome tool for creatives, but a burgeoning area of development begging for close regulatory oversight. Arguments for legislation, such as the US’ proposed AI Bill of Rights, apply in every way to AI art, which challenges core human abilities to a degree that calls for greater debate.

Undeniably, ChatGPT is more convincing than DALL·E. Perhaps this is because there's no written equivalent of the uncanny valley and that, unlike the more disturbing results of DALL·E, ChatGPT outputs come across, at worst, as the handiwork of someone who isn’t a very capable writer.

That isn’t to say it’s a silver bullet for those out who want to write poetry for their office, but don’t have the knack for it. ChatGPT is worse than you, I guarantee, so keep up the practice. But with its ability to generate prose that only needs to be passable – such as a quick email response – NLP models may become more integrated into software like word processors. It could be used to generate generic product descriptions or company emails.

Above all else, humans will still need to play a key role overseeing AI creations; none of the artwork or text here could yet stand up on its own. If used for public-facing use cases, organisations will have to assign teams to manage the output of the likes of ChatGPT to make sure the AI-generated content cuts the mustard.

Google has released Simple ML, a tool to help apply machine learning to data prediction and sorting tasks within Google Sheets.

Available as an add-on, the tool is intended to be stress-free to install and use, while powerful in its potential applications. Using a sheet with at least two columns, users can quickly train and refine a machine learning model to their own needs, opening up possibilities for automation across a company’s existing Google Workspace.

The tool is intended to enable enhanced activity in Sheets for both beginners and advanced users. It can be used for tasks like predicting how long maintenance could take based on past data, or spotting potential fraud when analysing unusual data in a client's finances.

By default, the tool is shipped with two functions: ‘predict missing values’ and ‘spot abnormal values’. For more complex use, it can even predict values within large data sets - Google gave the example of researchers using Simple ML to predict a person’s age based on their DNA in a molecular ageing study.

Simple ML was made by the same developers as TensorFlow, and is pre-trained to be as effective at processing data as possible. Despite the complicated nature of ML algorithms, and to combat the intense learning curve traditionally associated with training and deploying a machine learning (ML) model, the tool can be used on a fully no-code basis.

Nevertheless, advanced users that are already familiar with training machine learning models still seek to benefit from Simple ML, which allows for detailed evaluation of models to understand the quality of models that they have trained using the tool.

Models are broken down into a number of metrics, depending on the way the model was originally trained, and demonstrated through data and graphical representations.

Google recommends using Simple ML on sheets which contain columns with at least 20 values, though 100 is the recommended number for a model to really start training with accuracy.

Depending on the complexity of the task, far more may be required, and when a task cannot be completed, a detailed error message indicates what has prevented the model from running as intended.

A key benefit of the service, as cited by Google, is the fact that it runs entirely in a browser, so all sensitive data that could be included on a Sheet remains secure.

In addition, ML models are saved directly to Google Drive in their own folder, enabling them to be shared between a team in a business environment or downloaded for backup should they be crucial to new business processes.

In a blog post on the release, Google outlined the ease of use by listing the three steps necessary: opening the data; selecting the task that suits the use case, such as identifying anomalous data; running the model and seeing the results, which are weighted by probability.

On a technical level, Simple ML utilises decision forests - a group of randomly-generated decision trees governed by a machine learning algorithm to output predictive results.

Google announced the update at the annual Women in ML Symposium, an online event held by the firm to promote an inclusive community within the field. Simple ML is available right now in beta.

Automation may be a key tool in any company’s arsenal, but the promise of implementing end-to-end automation – or hyperautomation – with no human interjection, is an unachievable distraction for tech teams.

It seems only a few years ago that businesses began incorporating artificial intelligence (AI) and machine learning into their processes. This is a far cry from the widespread usage these technologies now enjoy. While they’re undeniably valuable, and a properly thought out strategy can be extremely beneficial, some companies are stretching too fast and too far in the pursuit of an unfocused form of hyperautomation.

“Digital leaders have been surprised by the complexity of automation”, Bev White, CEO of Nash Squared, tells IT Pro. Despite this, a recent Nash Squared report reveals digital leaders are, on average, planning to automate nearly one in five jobs (18% of the workforce) within the next five years. White adds: “With today’s unsustainable salary demands, talent shortages and increased operational costs, it seems the business case for digital labour is getting stronger by the day.”

For menial jobs like document processing, which increasingly contribute to worker unhappiness, increased automation has a clear role to play. If executed properly, customers won’t know whether a behind-the-scenes process has been done at all, let alone if it was done by a human or artificial intelligence (AI). Indeed companies continue to invest in automating the back office.

What’s really being promised is a seamless customer experience (CX), and this is absolutely something to which every business should strive. Businesses that buy into their own claims, though, would do well to take a step back and assess whether their automation goals really fit their needs. On the whole, the industry-cultivated idea of ‘end-to-end’ automated ecosystems is doomed to fail, because complete automation of services is’t only impractical, but also fundamentally unnecessary.

Chasing the best customer experience

The industry perspective on automation has been skewed by its development phase, argues Bernhard Schaffrik, principal analyst at Forrester, but now systems are beginning to mature, the endless use cases simply aren’t there.

“I think the challenge is that automation promises were so easily harvested, during the times where you were experimenting with an automation technology for the first time,” he tells IT Pro. “The appetite grew, understandably, but many missed out on sitting down together and defining an automation strategy, or an operating model. I'm talking to many senior IT decision makers, and they are just driven by removing humans from the equation because humans are too expensive.”

Schaffrik maintains that without a clear strategy, short-term over reliance on automation won’t help businesses, and will hold firms back from properly delivering on CX. He calls for automation to be implemented only where it’s urgently needed, but for this implementation to happen alongside careful consideration of where the “human touch” is preferable.

“Customer and employee experience are the key to transformation rather than an 'all-in' pursuit of end-to-end automation, and looking at tasks in this way will help businesses focus their efforts on using automation where it will give them the most gain,” states Dermot O'Connell, senior vice president of services, EMEA at Dell.

Even when it comes to back-end tasks that don’t affect the customer experience, automation can be more effective when used in tandem with human workers rather than in place of them. Just as humans can benefit from automated tools to speed up their work, automated systems can become more ‘intelligent’ through observation of human co-workers. Extremely complex or critical tasks can also be deferred to human workers, to ensure quality and to save time on the development and deployment of the automated processes.

“Previous generation thinking about automation was that it was either zero or one: either the work was done manually, or these robots would completely take over that work,” says Ronen Lamdan, CEO of intelligent automation firm Laiye. “The reality of it is, when you apply AI, you never get to the 100%. The winning combination is actually empowering these human workers with digital capabilities, with AI capabilities, to take over the mundane parts of their work, but still retain people to service customers, to handle edge cases, to do the more cognitive intelligent thinking.”

The hidden, human cost of automation

In September, NetSuite made a number of automation announcements at SuiteWorld, its annual conference. Powered by AI and machine learning, the new features introduce ways to automate tasks such as invoicing, planning work schedules, and tracking employee metrics. The promise of end-to-end automation makes sense within the context of the ERP giant’s full suite, and speaks to a goal of providing customers with barrierless integration between services. Nobody wants to pay for a suite only to discover the quality of life features are all siloed into narrow use-cases.

But the reality is there will always be a need, or at least an ability, to escalate decisions to management level. Even in NetSuite’s keynote presentation, it could only advertise reduced manual checks for invoicing, acknowledging that the machine learning system has to escalate anything it’s confused with, to avoid costly mistakes.

In some cases, automated systems can cause more trouble than they’re worth. Business leaders are already running into implementation challenges, particularly where human labour is still required to shore up automated systems.

A prime example of this is social media algorithms, sold to the public as self-sustaining, automated content systems that occasionally requiring minimal human oversight. One of the main benefits, often touted by firms, is their ability to curb illegal and harmful content pushed onto the users’ feeds. The likes of Instagram and TikTok now boast systems to prevent such content from ever being seen by others, or delete content that slips through the cracks.

But this process is far from automatic. Every algorithm that recognises harmful content has been ‘trained’ by humans, whose manual moderation of banned demands they be subjected to the most disturbing posts. The churn rate of such roles is sky-high – far above the industry average – and driven by the intense and upsetting nature of the roles.

There’s no easy route to successful automation

White also notes that complexity is currently one of the top three obstacles to implementing automation. “What may have at first seemed like a simple process to automate, can quite quickly turn out to be a spaghetti of logic branches – which will also require some level of manual oversight even when fully implemented.”

This complexity is particularly true for firms that have not engaged in the right amount of testing or strategising in advance of implementation. If a system has only ever been trained in a vacuum, hard-coded with functions rather than learning them on the fly from active employees, its use cases could be outdated by the time it’s properly installed. It could even be detrimental. Similarly, if a system has been designed with a ‘one size fits all’ approach to end-to-end automation, companies could find themselves rushing to rip it out down the line. Waterstones’ recent botched AI stock system affair is testament to just this.

This framing is important to understand the issue overhyping hyperautomation has created. It’s not a ‘one size fits all’ framework for increasing efficiency, nor a silver bullet for a sector facing a skills shortage. There’s no doubt hyperautomation in action can achieve powerful results, but only when taking factors like employee satisfaction into account.

On the same note, human labour saved by implementing automated systems can lead to quality improvements for workers and customers alike. Proper integration of automation alongside workers can be a win-win, improving satisfaction and productivity levels. But the lure of end-to-end automation should never overtake questions around its practicality, nor be used to hide human efforts behind the scenes.

Stack Overflow has made the decision to temporarily stop answers being posted on its platform that are created using the ChatGPT AI-powered natural language processor (NLP).

The moderators of the software engineering knowledge exchange platform said that it’s a temporary policy aiming to slow down the rise in answers created with the AI tool.

The decision was made since the average rate of getting a ChatGPT answer that is correct is 'quite low', said the moderators. Due to this, accepting ChatGPT answers to the site is “substantially harmful” not only to Stack Overflow but also to users looking for solutions to their problems.

In the post, moderators said that the main issue is that the answers have a high rate of being incorrect but appear to look authentic, as well as being very easy to create.

“There are also many people trying out ChatGPT to create answers, without the expertise or willingness to verify that the answer is correct prior to posting,” said the moderators.

“Because such answers are so easy to produce, a large number of people are posting a lot of answers. The volume of these answers (thousands) and the fact that the answers often require a detailed read by someone with at least some subject matter expertise in order to determine that the answer is actually bad has effectively swamped our volunteer-based quality curation infrastructure.”

This has led the moderators to ban the use of ChatGPT to create posts on the platform. If a moderator believes that a user has broken this policy, sanctions are set to be imposed to stop users from continuing this behaviour, even if normally the post would be acceptable. The moderators didn’t clarify what these sanctions would be.

The platform is still developing a final policy on the use of ChatGPT and similar AI tools which it said will need to be discussed with staff as well as users on Meta Stack Overflow.

ChatGPT is an AI-driven NLP created by OpenAI. It acts as an interface with which users can send prompts and it will return answers. It can simply respond to questions using natural language, as well as suggest additions or completions in code, and in some cases, spot security vulnerabilities in a codebase.

It uses a dialogue format where the tool can answer follow-up questions or reject inappropriate requests. However, OpenAI stated on its website that ChatGPT has several limitations, including that it sometimes writes “plausible-sounding but incorrect or nonsensical answers.”

Its creators said this is hard to fix, as during reinforcement learning there’s no source of truth and training it to be more cautious makes it decline questions that it’s able to answer correctly.

When IT Pro asked the AI tool what the problems are with ChatGPT, it produced this result:

Since the launch of ChatGPT on 30 November, tech enthusiasts have experimented with the API. People have found wide uses for it from the fun, such as text-based adventure games, to the professional, such as potentially innovative ways to educate students on various topics.

the advent of the technology has impressed individuals across the sector and some have claimed it's one of the biggest technological advancements of the year.

However, this isn’t the first AI tool from OpenAI to have problems. In November 2021, OpenAI brought to market its GPT-3 natural language processing API.

It opened up the technology to help developers make applications faster. However, it had been described in the past as “too dangerous to release to the general public” since there were fears over its ability to produce fake news that appeared to be real.

Explaining or predicting future trends and outcomes is not a straightforward task. Regardless of quality data science or determination, it's rare for predictions to be 100% accurate. In the modern world, this is where two practices come into play - namely, machine learning and statistics. 

Historically, humans have been a necessary part of the statistical process as it was we who were ultimately capable of examining and interpreting data. Through machine learning, however, computers can now answer questions and make predictions in the same way.

Yet despite the fact AI and analysis models are fast becoming ubiquitous, there are still common misconceptions about the benefits of using a machine learning approach as opposed to a statistical approach. Educational direction and career opportunities are just some of the differences here, with key differentiating factors between courses in machine learning and courses in statistics. 

Machine learning vs statistics: What is machine learning?

Machine learning uses computers to identify patterns in data, without the computers needing to follow explicit instructions. Instead, they are programmed with initial algorithms and models, and they learn to adapt these based on the data inputted to offer answers. However, this also means there’s less human control over the results.

Dr Michail Basios, co-founder of TurinTech, says: “Machine learning models are built for providing accurate predictions without explicit programming. Machine learning models can provide better predictions, but it’s more difficult to understand and explain them.”

Machine learning vs statistics: What are statistics?

Statistical models make their predictions from numerical data. This is often gained from questioning people and asking them to make a choice; for example, it’s how political and election polling works, or consumer shopping surveys. Human involvement is usually more involved with the interpretation of this data. “Although some statistical models can make predictions,” Basios explains, “the accuracy of these models is usually not the best as they cannot capture complex relationships between data.”

Machine learning vs statistics: What are the major differences?

Despite the fact statistics are entered into computers as part of the machine learning process, there is one overarching variation between both methods. This difference lies in the objective or desired outcome rather than in how the data is analyzed, according to Paul Clough, professor at the University of Sheffield Clough and head of AI and data science at business intelligence and data science firm Peak Indicators. 

Providing an example, Clough describes a comparison between the amount of chicken feed consumed by a hen and the number of eggs that chicken produces. Where a statistical analysis would look to depict the correlation between the amount of feed and the number of eggs, Clough says, a machine learning analysis would act differently. The data would instead be used to predict what number of eggs will be produced in the future.

Machine learning vs statistics: When would you use machine learning over statistics?

Clough’s egg analogy is useful in exploring this further. “The machine learning model doesn’t tell you anything about running a more efficient farm,” he suggests, “whereas the statistical model would be unwieldy if you owned tens of thousands of chickens.”

When looking at huge data sets, machine learning can be more optimal. Many concede statistics are unable to offer a deeper analysis on the relationships and correlations between the data when the levels of data are high.

They also cannot be relied on for causation, probability, and certainty because the danger is they might be misinterpreted or intentionally misused to back up a particular argument. As the famous saying goes: “There are lies, damn lies, and there are statistics.” Human involvement presents a weakness in statistics.

Basios, however, explains you can apply statistical modelling when you understand “specific interaction effects between variables” and “have prior knowledge about their relationships”. Machine learning, meanwhile, can be used when aiming for “high predictive accuracy”.

It tends to be ‘deployed’ more at the source of the data; as this is gathered and grows, algorithms automatically begin to provide intelligence. In manufacturing, for instance, it can predict when machines will need maintenance, reducing disruption. It can also analyse one option when compared to another – predicting which outcome would be better.

Machine learning vs statistics: When would you choose statistics over machine learning?

Conceptualizing problems is a good use case for statistics, especially problems that demand close consideration. According to Graham Upton, chief architect of intelligent industry at Capgemini, statistics are also great for the creation of metrics, tables, and key performance indicators (KPIs). Statistics is focused on humans and easy to explain, AI writer and researcher Eleanor Watson adds, as opposed to machine learning.

Machine learning has, in theory, greater power in more complex scenarios, Watson goes on, such as those that involve more variables than a human would be able to deal with. The most powerful tool is not always required, though. Watson adds that many organizations tend towards the overpowered “jackhammer” of AI for surgical problems in which traditional data science would be more cost-effective and more efficient. Data science has become an “unsung hero,” Watson says, as in reality it props up the world’s current economy to a much great degree than machine learning. 

Machine learning vs statistics: Is there a role for both machine learning and statistics?

Governments, businesses, and organisations are increasingly becoming reliant on machine learning as well as traditional statistics to help them make better and faster decisions. Each method, though, has its own individual drawbacks to consider.

“Statistical models have limited predictive accuracy since, sometimes, the underlying assumptions of the model are far too strict to represent reality,” Basios says. “Today’s businesses are adopting hybrid methods combining characteristics of statistical modelling and machine learning to understand in-depth how the underlying models work as well as to generate accurate predictions.”

Machine learning vs statistics: What does the future hold?

Decision intelligence (DI) is the space to understand in the future, highlights Dr Lorien Pratt, chief scientist at Quantellia. She explains it bridges both technology and the everyday decisions we all make hundreds or thousands of times to construct our world.

“There is a certain class of problem – ranging anywhere from day-to-day business strategy to the realities of climate change – that involves intangibles, complex systems dynamics, and human-machine interaction,” she says. “Machine learning, as a ‘supercharged’ form of AI, can cut through the noise and make a real difference.”

Pratt continues: “Rather than starting by thinking about the data we have, the key is to begin by asking what outcomes we want and what actions can be taken to achieve those outcomes, and then to fit the data into this action-to-outcome pathway.”

Amazon Web Services (AWS) has announced AWS Supply Chain, a new cloud application built to improve businesses’ supply chain management and lower their costs.

Central to the application is its system of connectors, which utilise machine learning (ML) models pre-trained by the cloud giant to draw data from across a company’s existing enterprise resource planning (ERP) software and supply chain management systems into a unified data lake.

AWS Supply Chain aggregates and contextualises this data, to provide real-time insight into inventory and quantity in the form of a visual map.

This information can be used by supply chain leaders, inventory managers, and other decision-makers within an organisation to identify bottlenecks and areas requiring efficiency improvements.

It can also help to identify more serious points of failure within a supply chain before they can have a wider, more significant business impact.

In addition to this, AWS Supply Chain will suggest actions to be taken based on problems it identifies. For example, the application could indicate a plan for moving products between two shipping hubs, taking into account the carbon impact, percentage for risk, distance between the hubs, and effect on overall spend.

With this level of actionable information, the application is intended to provide both short-term and long-term insight, enabling businesses to respond to the pressing issues within their supply chain.

With AWS Supply Chain Demand Planning, the application is also capable of analysing historical sales data to create forecasts, alongside contemporaneous information that adjusts models for accuracy.

Intending for the solution to cater to the needs of each organisation, the system learns from shifts in demand and common developer input so it can offer the information that is most relevant.

With the aim of improving collaboration, AWS Supply chain also comes with built-in chat and messaging capabilities to advance communication between teams.

The application’s compatibility with existing ERP and supply chain management systems used by businesses is a key selling point, as it negates the need for rip-and-replace or vendor lock-in. AWS states that the application can connect to current systems without the need for up-front licensing fees, or replatforming of any kind.

“Customers tell us that the undifferentiated heavy lifting required in connecting data between different supply chain solutions has inhibited their ability to quickly see and respond to potential supply chain disruptions,” said Diego Pantoja-Navajas, vice president, AWS Supply Chain.

“AWS Supply Chain aggregates this data and provides visual, interactive dashboards that provide the insights and recommendations customers need to take actions toward more resilient supply chains. And this is just the beginning - we will continue our investment in AWS Supply Chain to help our customers solve their toughest supply chain problems.”

Supply chain issues have been a hot topic for several years. In the immediate wake of the pandemic, the world experienced a damaging semiconductor shortage, the effects of which are likely to last until 2025. This prompted a scramble by nations to boost chip supplies, from Samsung’s proposed $191 billion Texan investment, to India’s chipmaking ambitions.

The advent of internet of things (IoT) technology has allowed for greater insight into stock as a whole, providing masses of data that can be leveraged by companies through systems such as AWS Supply Chain, or in private network environments like smart ports for maximum supply oversight.

On a company-by-company basis, careful supply chain management is an extremely important part of overall business strategy. Although there is a lot of software, and even artificial intelligence, equipped to tackle stock for businesses, improper implementation of such systems can end up costing more and irritating customers, as with Waterstones’ botched stock upgrade.

AWS Supply Chain is now available in preview in the regions US East (N. Virginia), US West (Oregon), and Europe (Frankfurt), and availability will expand to other regions in the near future.

With innovation changing the way of life and the way of doing business immeasurably in recent years, there are growing calls for regulation to oversee the pace of change in a controlled way. Artificial intelligence (AI), in particular, is an area that could most do well with structures to guide its development, although governments around the world have been slow to the punch, with the technology advancing exponentially in its capabilities.

In October, the White House’s Office of Science and Technology Policy announced a "blueprint" for an AI Bill of Rights. As the EU eyes its own AI legislation, and China moves towards further AI regulation, the American government has conceptualised a framework for AI practitioners to work towards. In its own words, it's a white paper "intended to support the development of policies and practices that protect civil rights and promote democratic values in the building, deployment, and governance of automated systems".

The AI Bill of Rights has no teeth – yet

Not all experts with skin in the game are convinced this is an ideal solution. The document lays out a wide range of principles – from data privacy and discrimination prevention to so-called “human alternatives” and a directive on “notice and explanation". But industry leaders like Aible founder and CEO, and former Harvard senior fellow, Arijit Sengupta, are concerned. He says some interested parties – particularly in the media – have jumped too quickly to talk about enforcement before considering the real-world impacts of any framework if translated into formal regulation.

“The only good thing about the Bill of Rights is that it has no teeth," he says. "If it had teeth, it would be even worse.

"Think of it this way, when new technologies come about, historically, what the US has always done is try to create a playing field where innovation could thrive. Our starting point was how do we make sure we are the leader in this technology? Interestingly enough, when you're looking at things like the AI Bill of Rights, that concept has gone out of the picture.”

Sengupta thinks the AI bill of rights, while possibly stifling innovation, is operating “from a place of fear”. While he welcomes the document's intentions – lest the space devolves into a dangerously discriminatory place to operate – the white paper’s authors need to take a deep look at the vagueness of the definitions of terms like "bias".

This level of scepticism isn't universally shared, though. Many in the space, like industry expert Sam Zegas, see the blueprint as a good first step and disagree with Sengupta's pessimism. Zegas is VP of operations at language AI firm, Deepgram, and was formerly a management consultant with the US Foreign Service.

“I agree with it being a set of guidelines," he explains, adding it would, of course, need more thought before it shifts towards any kind of formalisation, implementation and enforcement. "This isn't actually legislation, it's a position that the government is putting out there hoping people engage with, but it's not enforced in any way. I think that's the right approach because it's a very fast-moving space. I think there are a lot of different applications that may be different interpretations of these guidelines.”

Dr Chris Hazard, co-founder and CTO at Diveplane, and previously an employee at the US Department of Defence, agrees. For him, the framework can be made more valuable by thinking about second and third-order effects – like what incentives could spring up as a result of the document becoming more actionable, and how the US can balance business development to push AI’s corporate development.

“It's just a recommendation – a draft. I think that's exactly right, because it's so easy to get the legislation wrong," he warns. "If you get it wrong, you can stifle innovation, you can have people building the wrong things, and everyone spending a lot of money building things that don't actually help, but [do] meet the law.”

Other positives Zegas presents include the document's link with the ways in which the country has, in his view, previously looked to increase equity. “Rights-based frameworks have been incredibly influential in the last century in advancing the quality of the human experience in the United States and so I very much am in support of bringing a rights-based framework into the AI space.”

Can US AI regulation be a role model for the world?

Of course, none of this happens in a vacuum. Hazard says when the US thinks about the implementation of the AI Bill of Rights and how AI ethics can be further advanced, it needs to be thinking about how to lead globally.

“I think because of the market power of the United States, and the social and economic influences that [come with that], showing the United States is being thoughtful about this and is moving in the direction that many other companies or countries that lead in AI … shows that this is a global thing.”

Sengupta, on the other hand, isn’t convinced the US’ influence can be taken for granted. For him, the US has to be mindful if it wants to make its voice heard rather than relying on its historical dominance. “The point in time when the world naturally did what America did has passed," he says. "I think what's important for America to understand is that you can only lead by having a better example. You can’t just lead by example, right? And the issue right now is the approach they're taking is not a better example.”

Sengupta characterises the EU’s approach as “micromanaging industry” and China’s as focus on “societal cohesion". His worry about the bill of rights is that it will create so much bureaucracy that it'll overwhelm smaller companies, to the point they question whether it's viable to even attempt to compete in the industry.

On a global scale, Zegas says another risk is that an opt-out procedure related to bias with too much uptick could artificially sway the results being shown by any AI implementation. For his company, one that works in language, the variability of the data set, in his view, means that a complete balance of bias is impossible.

“The guidelines promote the idea that people should be able to opt-out of having their data included in algorithms and I think there's an interesting trade-off there which is that, for example, if a certain minority group were to were to opt out of including their data in a Deepgram algorithm that would actually bias our systems away from being able to serve them and their choice to do that would actually negatively impact the equitability of the product that we could put out.”

Meanwhile, Zegas says consumer choice is good and that the US is positioned, with the document, to exist in a “Goldilocks zone” between the highly regulated EU and the less controlled – but less IP conscious – China. He urges legislators to strike a balance that gives people enough freedom to innovate while freeing them from bureaucracy.

In the midst of the chaos, some are coming together to talk about equitability in AI. Hazard, whose company is part of the Data and Trust Alliance – which includes companies like IBM, Meta and UPS – says part of their work includes discussions around hiring practices, mergers and acquisitions, and promoting good data.

This work ensures companies are committed to not including unethically sourced data in machine learning and AI algorithms, while regulators across the world endure the long, hard slog of codifying such principles into law.

AI and machine learning offer great opportunities for organisations to improve workflows, enhance productivity, and strengthen monitoring. To ensure success, IT decision makers and developers need to choose the right hardware and software for their development infrastructure; it’s no longer just about the cloud or HPC.

This whitepaper shares how workstations are proving their worth in AI development, working together with cloud AI platforms to support longer, or less costly, tasks, and enabling the cloud to focus on more robust processes.

Download now to learn five ways workstations can be utilised effectively and efficiently for AI apps and how to get started.

Provided by

This automatically-generated transcript is taken from the IT Pro Podcast episode ‘The rise of automated contracts'. We apologise for any errors.

Rory Bathgate

Hi, I’m Rory Bathgate. And you’re listening to the IT Pro Podcast, where this week we’re discussing digital contracting. As automation has been ramped up, companies have sought to improve manual tasks such as document handling using machine learning and AI. But when it comes to automation in the legal space, strategy is key. Contracts can be very complex and require human input, especially in stages such as negotiation — achieving the perfect blend of machine and employee oversight involves striking a careful balance. Rich Somerfield, CTO at Summize, to discuss the opportunities presented by automated workflows, and how companies can implement systems such as contract automation to improve efficiency and insight. Rich, thanks so much for being on the show.

Rich Somerfield

Thanks for having me.

Rory

So just to start off with, what are some of the current problems in the tech sector that Summize is seeking to address?

Rich

So Summize is a product that aims to make legal easy, so trying to understand what's in contracts, trying to help legal teams work with the rest of the business to understand what's in those contracts, and just overall processing those contracts. So we see our product is really like changing the game, in terms of how businesses work with the contracts and understand them. But yeah, overall, it's really just trying to make that whole process a lot smoother and a lot easier.

Rory

So in trying to speed up that process and trying to make that process smoother. What kind of qualities have you identified, that businesses are looking for, in IT contractors at the moment? And then vice versa, what kind of qualities are contractors looking for when they're forming contracts with businesses?

Rich

So I think it's quite interesting. So I've worked with contracts, never have actually been a contractor myself in the IT field, but I've worked quite closely with a number of different contractors, and for a range of different reasons as well. So some have been very specific projects with a very specific requirements and milestones, and others have been much more open-ended, kind of almost like as an ability to kind of increase your development team quite rapidly. But without kind of the fixed nature of a long term employee, so really kind of accelerated. I've worked on so many different projects. And those ones have seen both successes and challenges. But then really, in terms of the contractors themselves, it's critical to try and obviously find the right people, get them up to speed as quick as possible, and really deliver value. And like I say, I've seen successes and failures in that. And I think typically, it really revolves around a clear understanding, when you go into the process, about what the milestones are and what you want to achieve, whether that be a short term or a fixed event, make sure everyone understands on both sides what's expected, or whether it's a like I say, a longer term addition to a team to kind of accelerate other expertise or capacity. But then in those scenarios, given they are sometimes longer, longer ended, or, you know, kind of over a longer timeframe with with a less clear end date, I think again, it's really important that both sides understand what's happening and what's what's going on. What the end point is really.

Rory

So on a contractor side, being transparent, very clearly communicating what what a business can expect when they're entering a contract with you.

Rich

That's right, yeah, like I say, I've seen some of that have been very much more open ended, which is, you know, you've got a team, it needs more capacity for for many reasons, or it needs some expertise. And in some cases, they have been multi-year, in the end, they might have started out as kind of a couple of months that gets extended, keeps on getting extended, and then you kind of look around and it's they might as well, you know, have been a permanent employee for the time they've been with you. But, it's different. And obviously, it gives both sides at that thing, get their own benefits and cons as well. You know, from a contractor's perspective, they get the opportunity to work in an environment, a little bit of stability, but without the typical constraints of being a permanent. And then obviously from the business' side, you know, if things go really well, then you can look to different options moving forwards or if things don't go well the project closes down or whatever it would be, then obviously, you can, it's a little bit easier to kind of just agree an endpoint and kind of move on separately.

Rory

So when businesses implement solutions, such as Summize, they're really looking at, they could see benefits in the short term, but they're potentially looking at really great long term ROI.

Rich

For sure, I mean, in the IT sector hiring is quite challenging. There's been a massive increase into the supply side of things. Since I started 20 odd years ago, the supply has increased, but the demand has increased and exceeded that supply as well. So everyone's looking for the right people, and looking for the right skill set. And I think that really, everyone's quite different, which is quite interesting, the diversity that we're seeing start to expand. Traditionally, it might have been someone who went into IT in high school, and stayed in it and beyond. I think what we've seen over the last 10 years or so, is people actually transitioning from other careers and sectors. And I think what we've found is that diversity is actually incredibly useful, incredibly beneficial people perhaps being in the services sector, for example, and bringing that to the IT team. I think, yeah, that that kind of diversity has been incredible, really, overall.

Rory

Yeah, there's been a lot of discussions around roles such as CIOs, for example, really benefiting from people with maybe an HR background or humanities background. Do you see that contracts are kind of following in that suit? And there are there are greater discussions around maybe more flexible roles and pairing up companies maybe that traditionally, you wouldn't have paired up before?

Rich

Yeah, I think that's a really interesting idea. I really like that that thought process really like you say, you're looking at those combinations of different experiences, lead to kind of unexpected results, and unexpected and beneficial results. So if you've got people from different backgrounds, different ethnicities, different kind of previous careers, those combinations actually really generate a much better end result is what we've experienced. And I've seen from the different roles I've done, as well, I always kind of look to like innovation, there's some people look at innovation has being like magic, but really, it's just connecting the dots between different things that have never been connected before. So if you've got someone that's perhaps worked in, like you say, HR, or the service industry or anything like that, that then comes into an IT role, then they've got experiences, they can connect in to new experiences, or new understanding of IT. And it's those connections that if you've been in the pure IT role on your career, you're unlikely to see. But the connecting the dots is actually quite simple, if you've experienced it. We've got a strong ethos at surmise round DEI, so a dedicated DEI person, and really trying to make sure that we're doing the right things. Because it is easy to kind of stick to what you know, and recruit from people that you know, but you're really trying to make sure that you're looking outside of that. And bringing those different, different experiences in is incredibly beneficial, whether that be you know like I say, all the different backgrounds, or contractors, what it would be like you think of typical contract has probably done quite a few different contracts, seen a few different things experience have experienced quite a lot. They've they've got a wealth of experiences they can bring into a team as well. So like keeping those combinations there is incredibly valuable.

Rory

So taking into account a range of data points for any hiring process, any contract process that really evaluates people are more than just the traditional metrics.

Rich

Absolutely right. Yeah.

Rory

Fantastic. So, on the sort of more technical side of the solution itself, Summize is looking at using AI powered automation, and improving workflows to automate contracts, automate hiring. How would you go about combining AI powered automation and human workers to achieve a balance. Presumably, there must be situations in the creation of some contracts where things have to be, for example, escalated to a to a manager for oversight.

Rich

Absolutely right. So, we typically try to think a little bit more about augmented intelligence than pure artificial intelligence. I think artificial intelligence, is obviously a very broad, a broad area with many different areas of applicability, but really thinking more about augmented, that ties in quite nicely with the legal, with it being very much an advisory service. So you've got people who have trained seeing different things, understands the current business strategy and the appetite for risk, all that kind of stuff. And then really kind of tying that together with the direction that you're assessing contracts that coming in, or that you're drafting new contracts, we really see that as kind of a, very much a human powered activity, really, at the moment at least. But then bringing that in with where are the areas where computers can help make that process more efficient, more effective, perhaps more predictable in the areas that make sense? So coming into legal I, prior to joining Summize, I was had not been in the legal sector at all. And from a programmers perspective, you kind of look at the look at contracts, and you think "well obviously they're going to be consistent language, you know, everything's going to be like a, almost like a coding language that turns into a contract". And the reality is extremely, a long distance away from that. Contracts are very much bespoke. Everyone's kind of got slightly different favourite clauses and different words and approaches to tackling them. So I think what's interesting is how can you blend that variety with something that potentially is more predictable? There's been some initiatives in the industry for things like one NDA and one DPA, which is trying to kind of narrow down the ranges in contracts that perhaps don't need that ultimate flexibility in the drafting side of things. But, but in other cases, you know, you've got your most important supplier, and negotiating the contracts with them, you're highly unlikely just to kind of click a couple of buttons and let the machine generate the contract for you, you're going to want to be much more involved, understand, on both sides of things, and also build that personal relationship as well with the third party that as you go through that negotiation process, you both understand what the best result for you is. I've looked into negotiation practices in the past, there's some quite interesting podcasts out there. And they often talk about aim for the win-win, because almost certainly, whether you negotiate contract with that, that party again, or like a friend of a friend, you know, there's always kind of relationships. And if you get a reputation for kind of, even for win-lose negotiations, it's always going to come back and cause you challenges in the future. So yeah, the win-win's where you want to get to, typically you need relationships to be able to do that. And like I say then the computer software itself, just trying to support you and make that process a bit easier. And anything that is repeated, really take on the load of those ones. But keep the advisory nature for sure.

Rory

Right, so it's it's drawing on a kind of repertoire, that it builds up and understanding of, like you were saying, in-house clauses that are used more often while leaving leeway for, you know, leaving the door open to somebody to come in and create the more bespoke aspects of a contract.

Rich

Exactly right. I mean, one of the one of the capabilities we've built into Summize is to, like understand payment terms, for example, it's good example, payment terms. Now, obviously, if you're a supplier, you might want those payment terms to be going in a particular direction, you know, kind of payment terms want to be as short as possible. If you're a purchaser, you might want the payment terms to be slightly larger, obviously. So you've kind of got some kind of thing you're looking to, to optimise or to average out overall. If a contract comes in, and it's massively different than your average payment terms, that's something you probably want to know about. So, Summize software can do that, we can automatically tell you what the average payment terms are, how this relates to that one. And then obviously from from the human perspective, you're going to understand this particular contract with this particular third party, how important is it for us to get that one really tight on the payment terms? Or are we a bit happier to be kind of slightly more flexible? And that's really where that advisory part comes in. But as long as you have the data points to kind of guide you, then we see that as a real win.

Rory

So are you looking for the software to kind of learn over time? So you're building up examples for it? Are you looking to go for a more traditional route, of sort of setting out rules for the AI?

Rich

I think it's both. As with all these things, everything's a bit of a hybrid, isn't it? So you start with the rules to get a good baseline to get the data. But then obviously, over time, the learned behaviour element is incredibly helpful. We kind of typically think around things over three phases, really, when we talk about the augmentation. So, you start with the manual. So here's all the tools as a human, you're totally in control of everything. And then from that point onwards, then the next stage beyond that is like a suggested based approach. So "we've seen you do an X, we think Y would be the next best thing to do", or we think this is interesting to you, so really trying to help cut down the time it takes you to see the data point to understand what's going on. As you started to look at those suggestions, if the, if your use of those suggestions gets to a certain threshold, effectively what you're telling the system is that the suggestions are pretty good value. And then that's when you look to the automate. So everything we really do follows those three phases, really. So we'll start with the start with the repetitive manual thing, shift through suggested and then turn to automated. And we see that's a really nice, nice approach for tackling AI overall and integrating that into people's experiences, that you gain the trust, and then once you've — you know, really when when you look into any role via the if you've got an assistant or someone that's helping you do something or team of people, you'll typically follow the same process anyway, there'll be an education phase. Then you'll kind of monitor and then the trust builds up and you say, well, great, we're happy with all that goes and and then until something goes wrong, then we'll just keep going.

Rory

Right? Putting the AI on probation.

Rich

That's not it's not a bad way to think about it. Yeah.

Rory

So on that point of rules learned over time: some businesses, when they implement automation workflows, worry about leaving losing oversight of the process, over time, or pretty quickly. How do you manage this?

Rich

It's a really good point. And I think it's, I don't think there's a free lunch, when you're talking about this sort of stuff. So whether you're looking to automate something, or whether you're looking to do it yourself, there's always going to be some involvement in the process. So even in a fully automated system, you want to understand how well it's doing is it being effective. You know, you might have a fully automated manufacturing plant, but if everything that comes out, the other side has got defects in it, it's totally pointless, you've just wasted even more time. So I think keeping that regular quality control, staff check on how things are going, is it actually working is critical. So you just kind of change your involvement in the process. And that might, you know, may be the case that you can offload that monitoring capabilities. But it doesn't go you know, you don't go from having work to do to having zero work to do. It's always important to stay in the loop, even though obviously, the time involved in that, and perhaps the skills needed to do that oversight changes might free you up for more time. But yeah, if it's doing something on your behalf, I think I would always recommend that you keep an eye in to make sure it's doing it correctly.

Rory

Right. And then on the other side of that, I know that Summize, offers increased data insight into in contracts with a data system. What kind of data are clients typically looking to draw out of contracts in particular, and and how are they using those to improve their decision making going forward?

Rich

We say that businesses run on contracts. And I think when you start to think about it, you might might not think about it often. But when you start to think about it, it's absolutely true, everything a business does, is really signed into a contract. Whether it be for your employees, suppliers, or obligations, or whatever it be, it's in a contract, almost always in a contract somewhere, if not always in the contract. And then what we often see, is that people don't really have a clear view on what contracts they've been signed up for. Whether the contract sat in someone's inbox, which is quite often the case, that it gets signed, stays in the inbox, nobody ever knows about it apart from that individual, or whether partially centralised often in a slightly ad hoc way. So I think one of the key bits is like, can you get all the contracts into one place? If the business runs on contracts? Which obviously does, you kind of need to know what's in them, the first step of knowing what's in them is actually getting them gathered them all together, and have a process which keeps them all together. So it's one of the things that our product offers is the ability to kind of centralise your contracts. Then the next part there is to actually understand what's in them, and how that differs or aligns to your business risk appetite. So there's an element of that one. And then the other bit is simple things where they like, end dates on a contract, or maybe you've got to lease on your offices; when do you need to renew the lease? When when's the timeframe that you perhaps need to break the lease if you want to move? The cost implications of some of these things missing that date, is incredible. So one of the elements that we've got in our product is we automatically understand dates and their relation within the contract, whether it be a termination or renegotiation that whatever date, and then surfaced that into a central calendar, so even if it's even simple things like that one, giving visibility around kind of like the calendar for your contracts is incredible, but like I say it starts really around collecting them all together, understanding them, and then different parts of the business are interested in different pieces of information. Whether you're like, the C level might be interested more in the kind of holistic risk you have, and where strategic is whereas down to individual account execs are probably more interested in the specifics around renewals and delivery targets.

Rory

So on the product more specifically again, Summize offers integration with Word, with Teams, with Slack. Is that something that was based on customer demands? Or was that led by a specific concept of the system?

Rich

When the company started, it was really trying to help legal teams be more as effective and efficient with contracts, and legal needs, and I think what we what was pretty clear to us straight away is that a law department so in the existing business law department is often considered the cost centre. They don't generally generate money, but they really like obviously. but even though they're providing value to the business, they're not really like a revenue generator. So what what was clear is that any kind of budget constraints that they might have, for that team themselves, imposing a totally brand new system on the rest of the business, and getting the rest of businesses switched over can actually be quite challenging. And you look at businesses as well, they typically have services that they've integrated with, whether it be like CRMs, or whether it be your IoT side of things, or your document management, you've almost always implemented some kind of system. And you're unlikely to want to rip all those out and replace it with an all in one solution from your legal team, for example. So you're really kind of always trying to look to see how can you integrate that into the existing environment. And that's really something that's been critical to the way we've thought about the product, so integrated into office 365, or Microsoft 365, as it's now called, or whether it be into the the G Suite side of things, Slack, Teams, you know, these are the things that people are using today and continue to want to use, and obviously Microsoft Word as well, as is hugely used in the legal sector. So integrate what people are already doing, I think there's a couple of benefits there, the ease of adoption, obviously goes massively, massively improves this limited need for training. So that becomes a lot easier. And then obviously, working with the rest of the business as well, there's less of that pushback, because it's kind of where they already are. So one of the elements that we've got is the ability to interact with your legal team, we call it legal front door, whether it be for creating new contracts, or reviewing contracts. That capability is available Teams, Slack, or through the web application. So it's really easy to adopt. And then also, because they've adopted that piece of technology and something like Teams or Slack, for example, that chat experience is where we can start to preempt questions. So typically, and this always raises a bit of a laugh with any lawyers I talk to, but they'll typically get an email on a Friday at 5.30 with a Word document attached from somebody in the business that just says "FYI". What does that mean? When do you need? What do you need me to do? Do you need me to review it is an important client, you know, all that kind of stuff? The first questions you’ve got to ask are that information. Well, our software means we can preemptively ask that question before you even see it as a human. And I think when we come back to what we talked about earlier about augmented intelligence, that simple step of pre collecting that information is a huge value add. So, when the legal when the person in the legal team receives the request, they know who it's for how important it is, and with it being 5.30, can they go for the beer? Or do they need to kind of stay around and just and just do the work? Because that is what it is. So I think like I said, integration into existing tools. And practice is absolutely critical in for being a success. And it's really important to us as well, that our customers are successful, obviously, a long term partnership, but really kind of starts as quickly as possible. How can we get you to the point where you now see this project as a success? We'll talk about kind of expanding beyond that in the future. But yes, always around that first point where you're actually seeing the value. And I think the other bit to touch on as well is obviously we've all been through COVID and experience, experiences around that. And Teams and Slack, the virtual office. Before that it was used in some places, but probably a little bit kind of sporadically. More often people were used to kind of the in person meetings, I think what we saw very quickly was those virtual offices being adopted. And you only have to look at the Microsoft Teams, the usage went from double digit numbers in the millions, to many, many hundreds in a very, very short space of time. And I think even though we're you know, people are starting to move back into the office, bit more of a hybrid style approach thing. It's really, really centred around those virtual offices, still, people are still using Teams, because almost certainly somebody on your team is going to be working from from home on particular day. And I think really, that's where business happens now, in those environments, Teams and Slack. So it's critically important that the we're there in that mix, and we can be part of that story as well.

Rory

Right. So yeah, you'd say that if you're not coming into implementing automation with a specific strategy in mind for how employees can access that and whether it's recognisable or not, then you're almost at risk of immediate tool sprawl. I know that that's something that employees run into increasingly, with a range of different digital tools that they're being asked to use, it makes more sense to have them in in a single place.

Rich

Absolutely. I think I mean it's a great question really, because it talks a little bit to what are you trying to get out? You're buying some software, or product or whatever it is, what is the what's the win scenario for that piece of software? I think it's always an important thing is trying to identify, looking at, you know, six months or whatever, looking back, what is it that we get to this point of being successful? So, some products are more tool-based, like you say, and they are excellent at what they do. But they're probably confined and constrained in the kinds of impact they're going to have into the wider business. So the legal team, I'm sure, could make use of many of these tools, but they're never really going to go beyond the legal team in terms of effectiveness, really. So we think a lot more around kind of like business optimization, how can we help the business work better with contracts. Well, we're talking about the business, clearly, that's all departments, and every area. And in those cases, like you say the one off tools are probably not going to achieve what you want. And it's much more around the entire business process, making that successful in as easy a way as possible. And I think a lot of that really, many experiences I've had, and people I spoke to had as well. It's really that implementation phase where you have those business transformations, obviously, but it's the implementation phase where the business transformation really stalls. And so that's one of the real areas of focus for us, that we want to identify, like I say, the first milestone where you started to get successful, and then how it helped you through those. But that first point, if you can get that first point of being successful, the rest really follow a lot more naturally. Yeah, absolutely.

Rory

Something on once again, the more technical level, but that I noticed that Summize is advertised as offering the ability to scan documents for legal risks. How is that achieved? Is that using the same kind of repertoire of clauses that the AI has become used to? Is it identifying problem areas in that way?

Rich

I think exactly right, exactly what you've said. So we've got the ability to analyse contracts for clauses, whether they be PDFs or or Word documents, whether it's a scan, effectively a photo of a contract, and doesn't really matter to us. So we kind of ingest those, when we ingest them, we start to analyse them for the different clauses. The often what you'll see in a contract, as well as even though there might be a primary paragraph related to a clause, there's probably other sections all the way through the documents that have additional additional information or constraints around that clause. And I think typically, when you're looking at contracts, you might find the first bits or finding the subsequent bits is a lot harder from a human perspective. So one of the things we've got is to try to tease out those bits. So we'll tell you the highlights, the one we think is the most important, but then also, it's really easy to navigate through the other bits as well, that you may or may not have found. So I think that's kind of the ability, to kind of analyse it. And obviously, when we're starting to look at other entities, like dates, places, times, etc. What are they in relation to? Is it in relation to, like, termination or payments or whatever. So we can understand that information as well. Coupled with our ability to run the analytics across your other contracts, whether they're similar contracts or different contracts, you can start to kind of build that cross-contract picture now. How does this contract compare to others of a similar type? Is it better or worse? Or would it be? And what party is it with? Is that party something we genuinely have different terms with? I think all that kind of stuff from an analysis perspective, which which learns and adapts from usage. And then the other bit that actually ties in quite nicely to this as well is the concept of a playbook. So in case you didn't know, a playbook is typically you have a Word document, or maybe an Excel file of "given this particular scenario, this is what we recommend you do. If they disagree to X, Y, Z and it's come through three negotiations, then please escalate to somebody else on the team". So that kind of internal guidance, on how to deal with certain things. Again, typically, it's in a Word document that people print out, they're probably not always got the most up to date version, they probably always don't refer to it. So one of the critical things rules is for example, in Microsoft Word, which is where most most legal work typically tends to happen, we surface that playbook actually in Word itself. And we can be context aware as well, what you're looking at, so we can tell you about what the favourable positions are for or who to escalate to when you're in the contract, when you're reviewing it, and the time saved from that is incredible. Because that play book is also centralised, it means that any modifications to your positions over time are automatically reflected on everyone's system that's using using Summize. So there's real benefits to be had there as well.

Rory

Fantastic. So that that rating system as well, just to circle back to that, that that would be in house per customer. So what constitutes maybe a risky contract or a good contract for one customer, for another, it's a different set of a different metric that it's measuring it by?

Rich

Exactly. Yeah, exactly. Right. And I think it's not just different per customer, it's almost different per party per customer as well. So with inside one customer, they may have different parties, like I said that you know, that their most important supplier versus maybe some of the not most important ones, then you're going to be looking those and analysing those in a slightly different way as well. So that's all alternately possible.

Rory

Right, fantastic. And then just broadly, on the kind of outlook for automation in the industry: do you think that going forward, the old idea of kind of carving out the back office and just plugging in automation is maybe not reflective of what's actually going to happen? And that what will happen will be more of this integration between automation providing context, but work is still very much being in control?

Rich

I mean, for sure, let's say we talked about augmented intelligence, and I think it's much better framing, artificial intelligence feels a little bit scary, a little bit unknown. I think when you talk about it as augmented, it becomes a lot clearer about what you're trying to achieve. So look for the repetitive stuff, get the computers to do that. And then the advisory elements really keep that with the humans. But what benefitted or accelerated through data, to really try to combine those bits together, and when you look at some of the advancements we've seen recently in like text to image, some of these, the image generation stuff, it's phenomenal what AI can do. But it's, I don't know, if you've used any of these image generator things it's very, very much an iterative process. You put some keywords in, get some interesting images out, which probably aren't what you're after, three or four, ten, fifteen tries later and it's, actually something from that section is quite interesting. And that's a really good example for where the way we see things is you still got the humans involved in the loop. They're kind of managing and coordinating work that's done by a computer, you perhaps don't always understand how the computers got to a particular point. But if it's good data, then it's good data, whether it came from a computer or whether it came from a human.

Rory

Thanks so much for being on the show. As always, You can find links to all of the topics we've spoken about today in the show notes and even more on our website at itpro.co.uk. You can also follow us on social media, as well as subscribe to our daily newsletter. Don't forget to subscribe to the IT Pro Podcast wherever you find podcasts. And if you're enjoying the show, leave us a rating and a review. We'll be back next week with more insight from the world of IT but until then, goodbye.

Automation has been widely adopted across the sector, and firms have repeatedly applied it to laborious tasks such as document handling. Using machine learning and AI, a great number of documents can be processed, and decisions made without the need for painstaking human input.

But automation in the legal space requires careful thought, as contracts can be very complex and often require a human touch — here, a hybrid approach is needed.

This week, we spoke to Rich Somerfield, CTO at Summize, to explore the opportunity of automated workflows, and the demands of the legal sector when it comes to implementing tech of this kind.

Highlights

“I think what we've seen over the last 10 years or so, is people actually transitioning from other careers and sectors. And I think what we've found is that diversity is actually incredibly useful, incredibly beneficial people perhaps being in the services sector, for example, and bringing that to the IT team. I think, yeah, that that kind of diversity has been incredible, really, overall.”

“So typically, and this always raises a bit of a laugh with any lawyers I talk to, but they'll typically get an email on a Friday at 5.30 with a Word document attached from somebody in the business that just says "FYI". What does that mean? When do you need? What do you need me to do?”

“It's really that implementation phase where you have those business transformations, obviously, but it's the implementation phase where the business transformation really stalls. And so that's one of the real areas of focus for us, that we want to identify, like I say, the first milestone where you started to get successful, and then how it helped you through those.”

Read the full transcript here.

Footnotes

• Gartner: RPA software market will continue to grow as business demand increases
• What is Robotic Process Automation?
• What is machine learning and why is it important?
• How AI can help – and hinder – the supply chain crisis
• AI is now powerful enough to automate the back office
• How can CIOs help to close the tech skills gap?

Subscribe

• Subscribe to The IT Pro Podcast on Apple Podcasts
• Subscribe to The IT Pro Podcast on Google Podcasts
• Subscribe to The IT Pro Podcast on Spotify
• Subscribe to the IT Pro newsletter
• Subscribe to IT Pro 20/20

This automatically-generated transcript is taken from the IT Pro Podcast episode ‘How can we make procurement smart?'. We apologise for any errors.

Rory

Hi, I’m Rory Bathgate. And you’re listening to the IT Pro Podcast, where this week we’re discussing procurement.

Whether a firm is looking to ramp up its manufacturing, expand its IT solutions or improve quality checks across its supply chain, the initial stage of accessing good partners and contractors to make these changes happen is key. Procurement is an ongoing, complex process, and can be a drain on any company’s resources, even with increasing levels of automation.

This week we’re speaking to Sudhir Bhojwani and Lalitha Rajagopalan, co-founders of supplier engagement platform ORO Labs, to talk about how the implementation of automation has to be balanced with a clear strategy in order to cater to specialised services.

Lalitha Rajagopalan

Thanks, Rory. Glad to be here.

Sudhir Bhojwani

Thanks, Rory. Nice to meet you. Nice to be talking to you this morning.

Rory

Lovely to be talking to you. So just to start off with, broadly, what kind of issues in the market in the landscape? Did you identify that that led you to creating or, and smart workflows?

Sudhir

So the core issue that we identified is the amount of manual effort and pain that a business user fields in a company when they have to procure something, procuring something very simple like, for example, Amazon business has made great strides in the procurement world. And people, more and more companies are connecting with Amazon business and other suppliers. Amazon business is just one example. Of course, other such companies make simple buying easy. When it comes to procurement people are not just buying stuff, which are commoditized on Amazon business, but they're also buying something more complex like services, for example, equipment, for example, software, technology purchases. So these purchases are larger purchases, they are pretty complex purchases, and purchasing through the system. It's a lot of manual effort, actually today, in every organisation. We come from a city Alibag, where we spent a lot of time a lot of big companies, we found even the best adopted procurement companies who have a procurement system in place, a lot of their purchases is still highly manual. And experience is, let's say, not Amazon-like and it can take weeks and months to work with suppliers. And if you take on supplier perspective, the pain is even more acute, if I'm a small supplier trying to work with a large company. And it takes me two months to get to the process. Because I just happen to be a small supplier and the customer is treating me like a big supplier. So that conflict that exists drove us to thinking hey, can we automate this workflow in a more interesting way where the friction between buyer and supplier can be reduced considerably. And that's what drove us to building this. And because current solutions were not built to solve the core user experience and removing the friction problem, to the degree we think we can and it's also driven by the complexity of the semantics, people assume people are buying standardised goods and services. The current procurement systems are designed for standardised goods and services. But people forgot that people buy a lot of non-standardised goods and non-standardised services and technology purchases have gone up, compliance requirement has gone up, everything has just actually gone up. And how do you still keep things ticking along and make the workflow automated? That's kind of why maybe Lalitha can chime in as well.

Lalitha

I think one thing I'll add here is that if you look at the current procurement technologies, they were all designed for a very simplistic spend approval, like yes/no. And most of it is basically, you know, should you be spending this money or not? And most of the time answer is no, unless it's been, you know, pre-procured by procurement kind of a thing. So even procurement has a big pain here, because right now when it comes to anything that is bespoke, or something that is specific to a business need, like I'm in marketing, I want to work with an agency for a specific project or something like that. Procurement really doesn't have any tool beyond throwing people at the problem they can take a ticket and then they have to figure out how to help the user navigate it. So feel the procurement technology is very stuck in a simplistic world. And procurement is not getting the help that they need, to help the business user make do it faster, do it better and keep the company protected. And I think that friction is in all user points and context is missing in how decision making happens today, if you just make it about spend approvals, and we felt that it was time for a richer and more cracked experience which will make it faster.

Rory

So not just, kind of, smoothing those those decisions, but also deepening those decisions and putting them into like you're saying into a wider industry context, and allowing businesses maybe to view a wider range of suppliers with with all of the necessary context? I think at this point, it could be good for those who don't know if you both could briefly outline what exactly Smart Workflows does in comparison to the standard model.

Sudhir

If you think about a procurement workflow today, most people will relate directly to this, we'll use the word approvals. People think of it as approvals. But the truth is, a lot of things happen in a procurement which is which goes above and beyond approvals. So as an example, if I want to buy the marketing service that Lalita mentioned, first thing I would like to know is, is the vendor already a vendor that we have done business with in the past? Second thing we'd like to know is, is an MSA in place? Are the contract rates in place place? What ave they done for us in the past? Do we need to get IP agreements in place? Is this an engagement that requires GDPR compliance? Is this an engagement that will require some more security consideration? So that's lots of considerations that are way beyond approvals, budget approvals. And, I also like to use the word, it really is a workflow, things go through step different steps in the process. Different people actually do certain things beyond approvals, like for example, negotiating a contract, I would not think of this as an approval, you negotiate a contract, and potentially afterwards, the contract needs to be approved, actually. So a lot of stuff actually happens. So when you think in that direction, so workflow is actually quite more complex, and it's not just approvals. That's how I would think of the problem more than anything else. And the reason we do things we decided here, we call it smart procurement workflows, it's plural workflows, because we don't believe procurement is one workflow. Because if you think about it, the way people procure marketing is not the same way, the way people buy software. It's definitely not the same way people buy raw materials, very different processes, you can imagine the workflow is different, the requirements are different, everything is different people involved are also different. And the reason we thought 'smart' is also equally important, is for the workflow to be effective, it needs to be smart as an example. It should not be the same whether I'm working with a brand new vendor, or when I'm working with a vendor, which I've already done business with, but my certificate is in already. Let's say I'm doing some construction-related work. And the health and safety certificate is already in place. Should we again, do health and safety? The answer could be yes or no depends. If the safety was done 30 days back, and I'm just starting a new, maybe you don't need to do it. On the other hand, if the safety certification was done two years back, maybe you need to do it again. On the other hand, if it was never done, you definitely need to do it again. So, the smartness really also means it needs to understand these concepts natively, and there are many such concepts, health and safety is just one. Sustainability is another concept, security is another, privacy is another, there are lots and lots of interesting concepts that companies have to deal with. And if you don't do smartly, you will end up doing same thing again, and again, for no good reason.

Lalitha

If you look at organisations today, you know, compared to say, 10-15 years ago, 10-15 years ago, it was very command and control kind of a model, which is very hierarchical. And you could have, you know, the process kind of laid out top-down, right. And I think the big difference in how people work today, modern companies are much more democratic employees have a lot more agency to do what they need to do to get the job done. And when people talk about oh, "speed of business" and everything, there's something fundamentally at core that's happening, where, you know, companies give a lot of room to people to operate the way intelligent people would want to operate based on the context, they're not really dictating the procedural aspects as much. But that also means that anything that's a centralised function, whether it's procurement, whether it's IT, whether it is legal, etc, they are stuck in a world where they have to have a customer service mindset with the employees and organisation to help them you know, to do what they need. And we see saw in IT, right? It went from, 'you can't even upgrade your browser' to you know, 'bring your own device', kind of a thing. And so it's really important to understand that then the central functions there is, it's very hard for them to always keep up with the context, and they are shrinking and shrinking their NPS score, and the operational SLAs are highly scrutinised. And so, our goal is that smart procurement workflows have to have context embedded in it, and it should adapt organically to the context. You don't have to rely on a good person in every department reading every request correctly, and doing the right thing. Smart nurses can auto context triage the workflow as it moves through the system, can the system itself basically make it easier for legal to review or IT to get the task done, or whatever it is? And it also needs to have understanding that oh, legal likes work in a different system, or IT uses something else, to be able to integrate a process across different tools. So 'smart' implies all of these things.

Rory

That's interesting what you're talking about with the kind of cross silo experience there. Something I know that a lot of firms implementing automation right now are contending with? Is this idea of things having to still be flagged to a to a higher level, how do you balance a kind of end-to-end, cross-silo solution with that need for maybe certain huge procurements or maybe ones that are flagged by the system as a one off? How do you balance that with that seamless end-to-end experience?

Lalitha

So it's really interesting, you bring that up, right? Like, one of our customers is, you know, a fast growing biotech firm. They have a policy, which says that, hey, anytime you spend more than $100,000, you need to have competitive bids. Okay? Now, oftentimes what happens is, when a scientist is trying to work with something, you may have a very unique thing that you want to get. And there's a bit pretty much a single supplier, who basically is the right supplier there, okay. Now, what you do not want to do is to automatically keep red flagging this, and it becomes a red herring, it's a false alarm kind of thing. So how do you basically make sure that you're helping the business user, and you're helping the procurement, and you're helping the executor who has to approve the one off? Think about it that way, right. So what a smart workflow will do, is it will actually not wait till the user fails, saying that, "Oh, you don't have an alternative," it will tell him right up front, it'll give him the in-guided policy guidance to say, "Hey, you're trying to buy something that's going to cost more than 100k. The company policy says that, you know, it should have a competitive bid, do you have a competitive bid? Or an alternate supply you'd like to consider? Or is it a unique situation? Like it's a unique supplier or something like that?" So the user is guided to provide exception, right there, they don't fail it and then go back. And then the second thing it does is that, from a procurement process perspective, it allows them to say, "Yes, I know this to be true." So they can kind of comment or annotate that. And it also allows, nobody wants it to go to like, you know, the CEO without the you know, CEO minus one kind of being known notify. So you can compose a workflow to say, hey, heads up to someone else, but then summarise it for their review to say, hey, the reason why this is so is because of this. So it comes with the notes there. So you're not trying to then think back and say "hey, why are we doing this?" or something. So how do people feel comfortable making decisions only if they understand the context? So, can you gather context along the way? So smart workflow doesn't mean that oh, intake is complete? And after that is just, you know, yes/no, yes/no. It basically collects context along the way. So that's one of the ways in which we make sure that, because at the end of the day, it's human beings who are deciding and human beings are doing it asynchronously. How do you make sure we carry context along?

Rory

That's very interesting. It seems like instead of keeping those processes under the hood, as it were, with automation, maybe that just that extra bit of context, that extra insight into why an AI is making the decisions it's making can kind of in turn, make all the difference when it comes time for people to make decisions.

Lalitha

What you said is exactly right, the visibility. Because today, users feel frustrated with procurement, because they are like kids stuck in the backseat of the car asking the parents "are we there yet? Are we there yet?" It's like a complete black box, and one big part of it is making the visibility of what's going on and why it's going on.

Sudhir

All right. I think it's such a, I mean, if you this is people think of this procurement as a ticketing problem. So hey, I just need to create a ticket and somebody will work on it. The truth is procurement is actually quite a complex workflow. That's why we really wanted to focus on the workflow. And we had, we took a workflow from from one of our customers who is live with us globally, now. They're doing, they're executing workflows in 60 plus countries with us. And they had the current workflow implemented on a ticketing system. And what they realise is, even though everybody was creating a ticket for someone to execute, it was not consistently even executed. Different people who got different tickets, they're executing it differently. Because it was not thought of it as a workflow problem, it was thought of as a ticketing prop. problem. We believe what procurement is much more of a workflow problem than a simple ticketing problem. Of course, the ticketing involved in the middle of the process sometimes, right, of course, but it's much more of a, "Hey, what is the step one? What is the step two? And how do all these things come together?" It is a complex data problem. How does the data flow from point A to point B? How does it and how do the decisions get made? How did the risk event how is the risk evaluated along the way, and that lots and lots of interesting stuff that you have to do to make the workflow be executed in a consistent and accurate way every single time? Regardless of who's involved?

Rory

That's really interesting. On that wider point of visibility that came up, I know that something that your system advertises is the idea of qualified suppliers, showing suppliers that overtime, accrue a kind of rating or qualification. I was wondering if you could talk a bit about how precisely that system works, and also how you balance a kind of internal qualification system with maintaining competition in the market, and maintaining a range of suppliers for any given customer.

Sudhir

So I'll start broadly and then I'll let Lalitha also chime in. So it's a good question. Firstly, let me start by saying, in procurement, the most complex object is supplier object. Like in CRM world, you would say customer is the most complex object. In procurement world supplier is the most complicated. If you look at the supplier in ERP, it has more fields than we can imagine. And every customer we talked to, including small customers, by the way, and I'm not talking just about really multinationals, even small customers, they have problems such as duplicate supplier records. My data is not clean. It is like my mailbox, I clean it, it gets dirty before I know it. So that problem is absolutely true. And then the question then becomes is, why is this even important? If you think about it, I'm trying to do business with a supplier. And I don't even have a good view of a supplier object. How can the workflow be smart? So our core premise is, we need to have a good understanding of supplier objects. So when you think about auto, what one of the important things in smart also is, we are not just a horizontal workflow engine that can solve any problem and every problem. We are geared to solving procurement, because we have a deep understanding of the core supplier object. So we understand how does the supplier object map into ERP, whether it's Oracle, whether it's SAP, whether it's NetSuite. We know why duplicate vendor records exist. And people think, "Oh, they exist because of a mistake". No, sometimes duplicate vendor records is the right answer. Let's say like, one company acquires another company, they both were doing business with the same exact supplier. It's not reality to just merge the two vendor records off requisition immediately, it's not reality, because there could be some transactions out there invoices to be paid, etc, etc. And there could be other terms too, right? So various reasons why I cannot merge the two records, but I want to still merge my ERPs. Right? So duplicate vendor records will exist. But one thing we saw actively is we give a view of the supplier, which is beyond duplicate supplier records, but now think what supplier itself, it also happens to be a company, which may have its own legal structure. They like, we are a very small company, always a small company, we are a startup. But we have two legal entities, one in here, one in India. And if you look at some of the other companies, they will have hundreds and thousands of legal entities. When you try to do all of this stuff. User is possibly extremely confused, "I have no idea what this jargon means. I don't understand what ERP means to begin with. I don't know what legal entities mean to begin with, secondly." But these two worlds have to come together, unfortunately. And the workflow has to do it effectively. So to answer your question, a lot of that revolves, that's the first piece of the puzzle, this is just the basics. The second thing after that is if you ask — the running joke in procurement, if you want to know what your company is doing with the supplier, ask the salesperson. And the supplier's like, he will tell you what your company is doing with them. And that running joke continues in the world. So our system is kept, and the running joke happens for a reason. There is no system, there is nobody who owns a supplier object today actually effectively gives a view of a supplier, that is really useful to the business use. Think about it. If I'm a business user, somebody in marketing, I want to find out what are the, what agencies we use to create some sort of production videos last year, or a year before last. Nobody can answer that question today. So our answer, we want to build a system that can answer those kinds of questions in a much more effective way. Lalitha maybe you can add.

Lalitha

I think the only thing that I will add here, is that the end vision for us, right end vision for us is we want to essentially be equivalent of CRM on the supplier side. Okay. And the reason why workflows and this kind of come together is because we realise it's not a data problem, it's actually an engagement problem. So it's really important to have a self-curating view so that when you think about a supplier, the system should logically be able to provide context of "hey, what are all the activity engagement that's going on with the supplier?" Again, going back to what I said about how companies operate today, today, even having view of what supplies my company does business with is very well hidden. So the core of our experience is being able to have that view. And also to enable to your question peer to peer, you know, discussion. And peer to peer ability to understand supply contracts, like think about is the water cooler effect, right? Like, I'm in marketing, I'm a new employee, and I want to start a project, I probably want to talk to other people in my team to say, "Hey, have you guys work with these guys will be like to work, how would it go?" And then you can very nuanced feedback, you get feedback, like, "yeah, this team is great, make sure they assign this project manager, that's the A team. And when you do the contract, make sure they do this, they'll say no, but we've done it in the past, just tell them, you know, this contract, we've done it." This is normal, tribal knowledge that is completely lost. So it's not just, you know, company name, company address, company logo kind of thing, you need this richness of context to enable better decision making. So at the end of the day, right, if great CRM for suppliers is going to help customers, companies essentially work better with their suppliers and make better decisions with the supplier, it should result in two things, I want to be the customer of choice with the supplier because I can work with them better, which means my company employees in my company can work with them better. Which means that every time a supplier shows up, it shouldn't be like, you know, you're starting from scratch, again, name of the company, everything. I should know you I should give you a red carpet welcome, I should know exactly how to work with you. And that tribal knowledge is kind of enabled. So that's kind of how we envision it and it was a long, a little bit of a road to go to kind of build it out that we believe that technology today. That's possible.

Rory

That's fantastic. So do you think that this, you described it as tribal knowledge, do you think that AI offers a kind of new opportunity to retain this tribal knowledge in organisation far longer than normally, you would be able to retain it manually? Because you know, people would leave them and with them, these relationships would go. So do you think that AI is a kind of a it can it can store that memory and retain that through through the years?

Sudhir

To be honest with you, I would say the two parts of the puzzle answer. First is not even AI. It's just the basic structure of the system. And the system is geared to store the data in a some, in I would say smart way. AI is different in what you learn from it, right? So first is can the system, I think you also asked the question, what granularity can you sort the data? Is the system designed to, does it understand the complexity that, hey, the supplier object exists in ERP with duplicate records, or the supplier object also exists in my contract management system. But that looks very different. And there's no connection between the two systems. Then the supplier is an existing company. So that whole jargon, it understands first and foremost, and that's not an AI, that's a more of a structural problem. Do you have the intelligence to do this? Now second, is what you can derive out of it. And I believe that's where the AI can be really, really helpful. A very simple example. We don't think of it as an AI. But nobody ever goes to Google on page number two, every result comes on page number one, is it AI? Or is it not AI? It's definitely doing a lot of learning, I tell you that much, the algorithm does a lot of learning. It's a highly, very curated high, it does a lot of learning, based on which it knows what what results to put on the rock top rank. So I think those are the ways we are thinking about AI problems. If people are saying, "Hey, can you make sure that if I type in, you know, almost what I'm typing and why I'm typing and based on that the recommendations that come out, are very accurate?" So this way, I can make decisions faster, smarter, better. And that's how at least the AI will come into play. And then of course, we can take it to the next level. How can natural language be used in a more interesting way? Because there are some really interesting power in procurement. Like for example, when you look at procurement of the cost of a commodity, what are you trying to buy with the supplier? Everybody's trying to map to this concept called 'commodity tree'. Some of some commodity trees are extremely complex, there's something called UNSPSC, which has like 27,000 nodes in it. Now picking a node from 27,000, good luck. I mean, and they're horribly — I mean, not horribly, it's not designed for human being let's call it, it's designed for systems. So the human being can't necessarily understand that. Can we use natural language for user, to describe what they're trying to do, and we figure out the rest? Simple problem, of course, compared to the supplier problem, actually. How do you make sure the suppliers are aligned correctly, etc. So yeah, that's how we're thinking about AI. So there's a lot of AI work that is needed to make this happen. But also the base, base application layer has to be strong to handle this, otherwise, nobody is gonna help.

Rory

So with all of that in mind, bringing in the improved structure, bringing in AI, where it helps. Would you say that this is more of a long-term investment for businesses, that the savings or the returns that we'll see are, are going to be more in the long term? Or are you also finding that customers are finding that, in the short term, they're seeing things dramatically improve by adopting automation within their workflow?

Lalitha

So I definitely think they have an immediate need, and an immediate ROI to this. So one of the ways in which it's immediate is that we are an audit ready system, which means that companies can feel confident about having an accurate picture of are we compliant in our engagement with the supplier X, supplier Y, okay? And what was done, what was not done? What's missing? So that's the question that you don't have to do analytics and reporting to do is like, right there. The second way in which I think the benefit is pretty immediate, is that often times when Sudhir said that it's all bespoke custom complex things that people are trying to, you know, work with the supplier for, you know, you are throwing people at the problem in a regular ticketing kind of world, only the procurement can keep up in a fast growing companies. Like I'm adding, you know, a dedicated procurement person for, you know, this department, that department. And by composing smart workflows, which kind of bring the context along, you can do more with less people. So I can, my immediate ROI is just operational cost management, right? I can scale more and better with a workflow tool, which is more sophisticated like this. The third value is actually, if you think about the supply chain disruptions and everything, it's really important to know, not just for procurement to know, but for the organisation to know, who are all the suppliers that we already kind of have an engagement with, what are they capable of doing so that I can basically start with searching or existing suppliers, and front and centre that is what is there in our experience. So I definitely think the ROI is immediate. However, in the long term, the long term, it's even more. I think it's an exponential value for you to say that I can have a consistent global process. That I can start my rollout and say UK, but then I can do North America, Asia, whatever. But from a process governance perspective, it can be consistent process, but it is done in a local context way, right? Like in UK, your quality check can be done by the quality team in Manila, maybe you outsource it to somebody. But so this is a checklist that needs to kind of work through etc. Right? So you can do that. And the second thing is also like, you know, I think having, it's a big competitive advantage for customers who say, I know how to work well with suppliers. And to build that internal competency, it's a constant curation and data problem. And that is what rolling this out in a global way would allow them to do.

Sudhir

But I think I like two use cases that highlights the problem. So one use case is, we work with a customer, we work with their marketing team, in particular in this use case. And marketing team says "Hey, my current procurement process will take me two months to work with a new vendor on this kind of a topic on this kind of engagement." What is the value if it comes down to two days instead of two months, to the business? It's not just the savings, definitely the savings. And I'll talk about that in a second. But the in marketing, that means they can do the lead gen faster, which is amazing. So the outcome procurement has been dreaming about how can I do that? How can I have a direct impact on top line, I believe speed is an important factor to deliver value on top line, especially when you talk about the business driven spend not. So that's one. Second thing I would say is, just talking to a CRO of a company. They say we our CFO does a monthly meeting, where we are all trying to say look at given the market conditions, we're trying to see how do we drive the savings? What contracts are possible that we can put a potentially put a pause on so we can save some money though savings is still an important part. And it's gonna get trickier next year, I think, given what we're seeing in the market, right? And they say this meeting is extremely complex the amount of preparation that goes into this meeting, and we still don't have visibility into what we can put a pause on in terms of discretionary spending. So our system, when we implement the workflow correctly, it also generates the right amount of data to make literally real time decisions. Now you can make real time spend management decisions, which actually are very tough to do in current systems.

Rory

That's very interesting, I think that point of of insight through real time data, and also the automation of compliance, are these both pain points that you're seeing common throughout throughout the industry throughout your customers?

Sudhir

Yeah, yeah, absolutely. I think if you look at the spend management problem, if getting more serious, more serious, and people really, really need to have a system. Look, I think spend, you can't manage the spend until there is a full adoption by the business user. So you need a system that first and foremost has the business users, so that adoption is high. Once the adoption is high, and then you need the smart system that gives the data in a shape that you can do real time decision making on that, so that you can control the spine, so generate the right data. So it should not be that it should not just be act as a repository, it should not act as the repository just. So we that so we see that as a pretty big driver. On the other hand, if you look at some regulated industries, like say, financial, financial institutions, like banks, for example, for them, this is not an optional thing. I mean, the kind of fines that you will see without onboarding and offboarding a supplier correctly, not following a consistent process to onboard — not just onboard, offboard too, by the way. Consistently we have seen fines of 60, 70 80 million dollars being applied by OCC. And this is a global problem. So I think on both ends, you'll see really lots and lots of examples.

Lalitha

I mean, to underline the smart procurement workflows, right, Rory, I think that simplistic spend approval is an ok system, if all you want to do is to track bookkeeping, record keeping, cut the PO, pay the invoice somebody approved it. But if you want a modern, I think modern enterprise software has to go beyond that it has to be a system that enables better decision making. And in the CRM side of the world, you kind of see that right? The modern ref tag, the CRM tag, all of them allow for a lot of, you know, a curation to be able to make better revenue management decisions. And I think that there, you don't hear revenue systems CRM system talking about order to cash and stuff like that, right? Procurement technology today still spend approval, you know, source to pay, procure to pay, it's kind of still stuck in that world. And I feel that in order to have a user experience where users actually want to use the system very early, it was only possible if it actually allows them to work better, work faster and make better decisions. And if it is just going to be like, you know, I need to, you know, file taxes kind of thing they will avoid as long as possible.

Rory

So it's it's not just about, it's not just about the fear of replacing workers, which I know is always tied to automation, but it's also about giving workers like we were saying earlier, that context, that ease of mind that they can go in, and the system will assist them in doing their job properly.

Lalitha

Yeah, like if I'm a facilities manager, I'm in marketing, if my job is important to me, and my job depends on me working with a supplier. So I'm, I'm inclined to do whatever I need to do to get the job done. And if a system is default, designed to say "no", for anything that is not standard, I kind of want to avoid it. That makes sense. So I can always ask for forgiveness later I think.

Rory

I think expanding on that, too, do you see this as a model that companies will be aiming for, or should should be aiming for in the next few years? Implementing not just more optimised workflows, but also AI and machine learning in this way to, as a kind of a backbone and the back office, but also to provide this this extra context?

Sudhir

Yeah, absolutely. I mean, I think the supply chain disruption has made supply chain as well as procurement, a whole lot more important and opened people's eyes that, hey, this actually impacts us. This is not just a back-office function, where a sub-optimal system will suffice. Or just a back-end process on an Excel is also okay. It's okay for us to not even know which supplier does what for us. And that is hidden in an Excel sheet. And we can't figure out from a spend management perspective which projects or which contracts we can put on pause and save money. All these decisions are important, and if not done in real time, I think people are starting to see this. So if you look at the general investment in SaaS, obviously CRM always leads the way. Everybody wants to make money first, I think before they worry about saving money, generally speaking, so the CRM always sees a lot of investment. And HR investment is natural, you have to do an HR investment, you cannot not have hire people and not know how to pay them, or not know how to track their benefits. So HR investment is clear. So the procurement has been the almost a little bit, in the back in the queue. But now the supply chain disruptions, and the dependencies on global economy are now more and more things are coming local as well. So this whole shift in the global economy is, people are realising, oh, this is an important part of the puzzle. This is not just Excel kind of a system that I can live with, I really need to have a smarter system. So I believe this, this area is gonna see a tremendous investment in next five years. That's kind of our bet anyways.

Lalitha

Yeah, so basically, you were using the word backbone, right. And I think that, you know, I see ourselves as a new age platform, which is actually not under or behind, but more on the top right. And I think that a good analogy to think about is actually smart home automation, right? Like, everybody has old homes with old wiring, etc. But if you think about modern home automation stuff, it is doing two things. One is my smart light bulb senses, its context of that, "oh, it is getting dark so I need to switch on". Or my you know, "resident is away, so I need to pretend that there are human beings here", right? You know, and it basically I can talk to it, it's very talk to me baby kind of thing, I can say, "Hey, turn on the music in the living room, it's gonna rain, therefore, I should turn off the sprinklers" kind of thing. So and I think that in enterprise software, especially the SaaS explosion, with every department having a very specific tool for doing some job, there's a new layer of technology, which is more like a fabric, which is the word for these users, which is going to unify the entire context. But it's not just you know, routing, it's not like an RPA kind of thing, like take it from system A to system B kind of thing, just a routing logic kind of problem. It's a context gathering, and context service problem. So that you can essentially provide a layer of seamless start to finish automation, which is kind of what, you know weird enough, in a consumer world, you're trying to do the same thing, like, you know, can the can an operating system run my house kind of a day? Sudhir, can you talk about just one thing, I don't think we touched upon it in a specific way, which is one of the things we're trying to do is to work with external data repositories. Like for example, you know, if a company is a UK-based company, then we know how to work with Companies House to kind of validate that entity and things like that, right? So can you just bring in this data context?

Sudhir

Yeah, I think while customers find it very interesting that we actually integrate with their ERP, and our, one of our goals is to give a good visibility of an existing suppliers. That's a really tough problem, actually. Because as I said, duplicate vendor records exist. So how do we take the existing data and turn it into something that is fully searchable? And I always felt this way, is if I searched for IBM as a company in a procurement system, then I searched for the same IBM on Google, the knowledge panel of Google on the right hand side gives a pretty interesting piece of information. And that information seems to be more interesting than the information I have in the company. When I'm spending $4 million at IBM. How can that be? Public information, public sources have gone a few notches, let's say a lot of really, really interesting data sources. And one thing we are doing is, we are enriching data from various data sources. Obviously, our use case is procurement, we're not trying to build ZoomInfo for our goal is not to be ZoomInfo. But our goal is to really get the right data, so the right decisions can be made faster, and that lots of interesting data sources we're using right now, to get that enrichment done so we can deliver higher value in a smart workflow. So our goal is smart. Smart is important is not just candy, candy word, in the slogan. I believe, without smart workflow, people will continue to do the same thing. Same mistakes will happen, and then we'll wonder why it takes two months, what should take two days, and we believe without proper data, and without the proper, proper context, it will never happen.

Rory

Well, thank you both so much for being on the show.

Sudhir

No, thank you, Rory. It was pleasure, pleasure talking to you, and hope to connect with you again sometime in the future.

Lalitha

Thank you Rory.

Rory

As always, You can find links to all of the topics we've spoken about today in the show notes and even more on our website at itpro.co.uk.

You can also follow us on social media, as well as subscribe to our daily newsletter.

Don't forget to subscribe to the IT Pro Podcast wherever you find podcasts. And if you're enjoying the show, leave us a rating and a review.

We'll be back next week with more insight from the world of IT but until then, goodbye.

Procurement has long been treated as a ticketing problem, with check boxes to fill and suppliers to onboard according to regulation. But with increasingly sophisticated automation systems, and a growing need for exact partner agreements amidst global supply chain issues, companies could do well to rethink their linear approach to procurement.

This week we talked to Sudhir Bhojwani and Lalitha Rajagopalan, co-founders of supplier engagement platform ORO Labs, about its ‘smart’ procurement strategy, and the place that automation has within the procurement workflow.

Highlights

“If the safety certification was done two years back, maybe you need to do it again. On the other hand, if it was never done, you definitely need to do it again. So, the smartness really also means it needs to understand these concepts natively, and there are many such concepts, health and safety is just one. Sustainability is another concept, security is another, privacy is another, there are lots and lots of interesting concepts that companies have to deal with. And if you don't do smartly, you will end up doing same thing again, and again, for no good reason.”

“Think about it that way, right. So what a smart workflow will do, is it will actually not wait till the user fails, saying that, "Oh, you don't have an alternative," it will tell him right up front, it'll give him the in-guided policy guidance to say, "Hey, you're trying to buy something that's going to cost more than 100k. The company policy says that, you know, it should have a competitive bid, do you have a competitive bid? Or an alternate supply you'd like to consider? Or is it a unique situation? Like it's a unique supplier or something like that?" So the user is guided to provide exception, right there, they don't fail it and then go back.”

“Users feel frustrated with procurement, because they are like kids stuck in the backseat of the car asking the parents "are we there yet? Are we there yet?" It's like a complete black box, and one big part of it is making the visibility of what's going on and why it's going on.

Read the full transcript here.

Footnotes

• ORO Labs
• What are digital workflow solutions?
• What is customer relationship management (CRM)?
• Gartner: RPA software market will continue to grow as business demand increases
• What is Robotic Process Automation?
• What is machine learning and why is it important?
• What is the supply chain?
• How AI can help – and hinder – the supply chain crisis
• AI is now powerful enough to automate the back office
• How smart homes are the next frontier for independent IT providers

Subscribe

• Subscribe to The IT Pro Podcast on Apple Podcasts
• Subscribe to The IT Pro Podcast on Google Podcasts
• Subscribe to The IT Pro Podcast on Spotify
• Subscribe to the IT Pro newsletter
• Subscribe to IT Pro 20/20

Intelligent automation firm Laiye has forecasted the end of subscription models in the IT industry within the next five years, citing the inadequate extent to which such models can serve growing automation needs.

Laiye's chief strategy and ecosystem officer Braham Shnider said 'rip and replace’ requirements currently imposed in the software as a service (SaaS) market, in which companies have to remove current systems in order to update to new ones, are an ineffective model for automation.

This is because of the way that Laiye argued end-to-end automation will have to work, as a combination of automated 'digital workers' - programmed bots performing specific manual tasks - with humans working alongside them.

In such a system, digital workers could be loaned for specific tasks like a rental car, rather than paid for as part of a wider systems package in a costly, yearly subscription.

Instead of subscription models, the firm is proposing a more ‘open’ automation ecosystem, replacing digital transformation and vendor lock-in with ‘digital transition’, to address the skill shortage gaps experienced by each company.

In the coming years, the company believes digital workers could be paid for by firms through a consumption model, with a set amount given to vendors per process carried out, akin to virtual freelancers.

“For example, we can provide an invoice-processing digital worker, and we can charge by the number of invoices that digital worker processed,” said Yichuan Hu, co-founder and CTO at Laiye, to IT Pro.

“It's no longer 'you need to pay ten thousand for this digital worker, for a year'. You just take, say, 10 cents for every invoice this digital worker processes. So that will actually be more attractive and more elastic for customers.

“There are three stages, and at the third stage — that is ten years from now — our product will become ‘work as a service’.”

Launching the Work Execution System

Laiye’s claims were made at an event held in London marking the launch of its new Work Execution System (WES), an automation system which provides partners with digital workers powered by machine learning (ML) and artificial intelligence (AI).

WES digital workers learn rules, taught by human employees in no-code interactions, which they can then contextually apply to tasks of varying complexity, from menial form processing to more sophisticated handling of forms.

Laiye advertises that WES can be integrated with systems such as UiPath, Slack, Teams, and SalesForce, and Shnider claimed that it “doesn’t care” if its customers choose to go with Laiye robotic process automation (RPA) or a competitor’s.

At the same event, the company highlighted the importance of harnessing the metadata generated by automated systems, in order for organisations to adequately address the ‘work execution gap’.

Trillions of pieces of data, from integers to complex information such as documents and videos, are caught up in automated systems, capable of producing vast amounts of metadata for big data analytics.

With an aim of getting a better view of this data, the firm announced that upon release, users will be given the choice of a free version or paid enterprise version.

In the former, users will be able to leverage the features of this system at no cost, in exchange for sharing their metadata for public use, giving companies insight into non-identifiable industry and use-case benchmarks and feeding this data to better train ML algorithms.

For example, a company could compare their invoicing automation stats to the industry benchmarks, use this information to determine whether more changes need to be made to their system, and in the process teach their system gaps that should be eliminated in future.

Paid enterprise users will generate only private metadata, for internal benchmarking purposes — IT Pro was given the example of the figures from different departments being compared against one another.

Laiye claims that companies that have adopted its automated systems thus far have seen a 60% improvement in productivity, and is willing to offer a full refund if it cannot integrate its WES by a company's deadline.

Another benefit of accurate metadata is explored in an upcoming white paper by Laiye, which proposes a carbon calculator for its customers.

Through the WES, this will determine the sustainability of companies in comparison to the overall market, and offer partners discounts in the form of credit against future purchases in exchange for meeting sustainability goals, which Gartner highlighted as the key strategic technology trend for 2023.

The beta version of Laiye’s WES will be launched on 30 November, with general release slated for 20 January 2023, and any company that buys a current Laiye product before 16 December will be given free access to WEP in 2023.

AI has the power to transform industries and set businesses up with lasting competitive advantage. However, the full impact of AI can only be achieved when AI can be trusted.

This latest eBook identifies three key goals of MLOps and trustworthy AI: trust in data, trust in models, and trust in processes, and analyses building blocks to help achieve those goals such as data collection and access, model building and deploying, and monitoring and managing models.

Download now to learn how MLOps, trustworthy AI, and the data fabric are related, with real-world examples from the healthcare, manufacturing, and banking/finance industries.

Provided by

Even though the UK has a temperate climate, managing water remains a significant challenge. This year became among the hottest and driest on record, yet the UK has also suffered major flooding. These seemingly diametric aspects of water management must be reconciled to guarantee water security in the future.

Published in 2020, the National Water Strategy sets out the challenges the UK faces in meeting its water needs, and how infrastructure will need to evolve. Specifically, new technologies will be required to conserve fresh water and solve these problems, particularly as the effects of climate change deepen.

Proposals to create a national water grid have been previously considered, based on the electrical power grid. Under this system, abundant water supplies could be transported between regions of the UK, based on need. This network’s energy requirements, however, and knock-on environmental effects could pose challenges.

The University of Manchester did consider alternative ways to transfer water between regions in specific scenarios, but the cost and environmental impacts make this technology largely unfeasible. Even desalination – the process of removing mineral salts in water – are part of the strategy but, again, the ecological overheads with emissions push this technology beyond practical use.

There can be potential, however, in new technologies, such as the Lockheed Martin’s perforene graphene filter, which could make desalination cost-effective. Systems, too, producing fresh water from wastewater will need to expand. Perhaps water companies can even look to NASA, which has been developing wastewater reclamation systems for decades.

Using IoT to make water smarter

Understanding water usage is a crucial component of any water strategy. Here, the Internet of Things (IoT) could hold the key to rapid improvements in leak identification, and in tracking the movement and consumption of water. In addition, smart wireless networks, which are coming into practical use as 5G rolls out, provide a communications environment upon which water tech can be based. Indeed, Ericsson predicts that five billion IoT devices will be in existence by 2027.

The water industry has many centralised facilities which can be easily adapted with sensor technologies such as supervisory control and data acquisition systems (SCADA) to improve their efficiency. However, the distributed nature of the water network provides a more significant challenge.

As the wireless environment expands with IoT maturing, though, the wider water network could be integrated into existing management systems to deliver end-to-end insights, in much the same way the smart metres are transforming how energy is consumed. If water is awarded the same status, many similar technologies can be applied to its consumption and management.

The impact of leaks on the water system's overall performance can't be understated, and it’s one of the issues that plagues Britain’s creaking water ecosystem. There are examples of technology being used at home and abroad that can come to the rescue. Systems like Fuji Tecom’s Quatro Core LC-5000, for example, can locate leaks using noise detection. Water consumption is tracked in Saudi Arabia by IoTSens without the data loggers breaking into the water mains to install monitors, while CityTaps is one of the first companies to offer a prepaid water metre.

“Acoustic logging of water pipes is one area where I see a lot of potential to revolutionise our approach to leakage,” Jack Lawton, data science principal at Aiimi, tells IT Pro. "By listening to the sound of leaks along pipes at various points, it’s possible to triangulate the exact location of a leak – in some cases within one metre accuracy. We have seen this effectively deployed with our customer, Anglian Water, and there is a growing interest elsewhere in the sector.”

For many water companies, digitising their network process management is essential. Here, creating a digital twin is gaining traction across the industry, as water companies see the advantages digital twins bring to other businesses. Mirroring their networks and support systems in a digital space could open the power that IoT sensor technologies and machine learning could bring to all water companies.

A handful of technologies will be needed, in fact, to make the UK’s water system as efficient as it needs to be, says Joseph Tabita, energy and commodities lead, SVP at digital transformation consultancy Publicis Sapient. “An improved digital ecosystem is essential using the full gamut of available technology – from IoT to robots to machine learning – will help us manage a dated and complex water system.”

Asset management and conservation

Water is a precious asset that’s often taken for granted. However, a multifaceted approach is needed to ensure water supplies are maintained, but also that water is used efficiently.

“Actions to manage water demand, such as reducing usage and leakage, have a vital role to play in meeting future needs,” concludes the current National Water Strategy. “These actions can be delivered quickly, are scalable and are often less costly than increasing water supply through infrastructure or moving water from areas of surplus to areas of need.”

A more integrated approach to water management that uses technology at its foundation is already in operation. Led by Northumbrian Water, the Stream project aims to use open data sharing between all the major water companies to help them manage water on a national scale. The project has oversight from the Open Data Institute (ODI), which will deliver user personas, technical design architecture and mapping of the broader water data ecosystem in the UK.

“We know the importance of a clean and efficient water supply and strongly believe in encouraging innovation to both solve problems and improve services to the public,” says managing director at the ODI, Louise Burke.

For Dom Couldwell, head of field engineering EMEA at DataStax, collecting and analysing data is only the beginning. “Managing the UKs water management systems requires processing massive amounts of data covering water quality, distribution, capture and demand,” Couldwell says. “This data is captured from a wide variety of sources, in different formats and frequencies and then needs to be analysed in real time to react to changes. It must support multiple data models from time series to unstructured events. We must organise that data and make it useful to cross functional teams with differing skill sets, use cases and allow them to take action on that data in a timely way.”

There are myriad technologies that could be used to resolve many of the water management issues companies currently find challenging. Connecting several technologies – especially advanced sensors – to create a digital ecosystem that can be used to manage water from extraction to recycling is how every water company should approach their digital transformation to ensure resilient water security.

Automated transcription software isn’t exactly new as an idea or a finished product. Most of the notable players in the market, like Otter, Trint and Temi came into prominence over the last ten years, with Rev launching as far back as 2010. All services that fall into this bracket are powered by artificial intelligence (AI), particularly machine learning and natural language processing (NLP), and promise the ability to do away with laborious and time-consuming manual transcription, creating fewer bottlenecks in business.

The need for accurate automated transcription in both the business-to-business (B2B) and business-to-consumer (B2C) worlds – as well as occupations like journalism – has been ever present, but this need isn’t being served as well as it can be. Far from any particular provider bearing the brunt of the guilt, the ecosystem as a whole is underserving customers for various reasons – from inadequacies in the technology to a lack of accessibility.

Picking apart the faults in AI transcription

Much of transcription software’s marketing is directed at businesses, with features that focus on picking out key moments from recordings and enabling collaboration. There’s another area, however, where these tools have chosen to focus their energy: accessibility.

Prior to these tools being available, transcription was often prohibitively expensive. Since Rev burst onto the scene in 2010, primarily focused on being a marketplace for freelancers to provide less expensive transcription than was available, the market has seen a large influx of venture capitalism cash. With newcomers like Airgram and Verbit garnering huge cash injections in initial funding rounds – $10 million and $23 million respectively in 2022 – it’s important the ecosystem begins to listen to the needs of their customers. Initial promises of lower costs, and greater access, seems like a reason to jump for joy – but there’s a catch.

Transcription software is notorious for generating errors, no matter how finely-tuned the AI is. One very high-profile example is on YouTube where the company has come under fire for removing a community captioning feature that allowed creators to lessen the burden of fixing the often clunky transcriptions provided by parent company Google.

Social media abounds with examples of poor transcriptions, all powered by various forms of AI, from the likes of major tech names like Facebook and TikTok. Many of these tools argue the solution is usage; the more diverse the user base, the more accurate the transcriptions are – and vice versa. While there are significant differences between companies, some offer multi-language support while others are far more geared towards English. What’s clear – even after just a cursory look at this segment of the market – is that it isn’t as open and shut as marketing would have you believe.

AI transcription’s accessibility problem

Improper captions are one reason why Svetlana Kouznetsova, a New York-based B2B accessibility strategy consultant who previously worked in web development, believes AI transcriptions are only fit for purpose when context is understood by all parties. She says people should be wary of deploying such tools in business environments and presuming all will be fine following implementation.

“I do not use auto-captions for work,” Kouznetsova, who has impaired hearing, says. “I use auto-captions mostly for informal conversations because auto-captions are often inaccurate, and it's more than just words. There are many small things to consider that machines cannot do, also it's important that captions are readable and understandable. Even if captions are accurate but not formatted or designed well they are hard to read.”

Kouznetsova says that as tools have rushed to herald their accessibility features, they have seemingly forgotten that context is key when trying to address accessibility in these spaces. She cites Otter, as an example, announcing an integration with Zoom in 2020 to provide live captioning. For her, it’s not just about which words can be spit onto the page, and in which order, but also about how transcriptions feed (or don’t feed) into readability.

“Many people don’t realise that those who can hear can fall back onto their good hearing, if text doesn't make sense to them. However, for us deaf people it's important that text is of good quality, otherwise it causes cognitive dissonance. I'm not against auto-transcribing tools. They may be useful for some situations. But they aren’t the best accessibility solution. It's like hiring a writer to write a book. The writer may do a great job writing a book. But if a book isn’t well designed, it's hard to read.”

Boosting the AI transcription ecosystem

Accurate transcription software might have played a hand reducing the gap between what the IT team thinks customers want and what they’re actually asking for, quips Michelle Symonds, who worked for the CITI Group in various roles, including IT project manager, for more than a decade, before founding her own UK-based SEO company.

“I think one really big issue then, and now, is that stakeholders and end-users make assumptions,” she explains. “IT professionals make assumptions, too, and they're not always the same assumptions. If you could have some record of that it might help to solve that sort of problem. It’s quite simply that IT people and business people, for want of a better word, don't speak the same language; that the IT people are thinking in a different way.”

Although there are problems, the quality of the technology has been improving over time, despite the fact that, for example, much of the software is anglo-centric. Using the technology also requires a significant amount of human intervention to sharpen up transcriptions generated from conversations. There might be hope on the horizon, though, in the form of more advanced technology branded Whisper, developed by OpenAI, which claims to recognise and translate audio at near-human levels. Should the likes of Whisper, or other ventures, live up to their potential, it could raise standards across the entire ecosystem.

Just as pilots run through a pre-flight checklist before they ever leave the ground, taking steps to increase your understanding of and confidence in your data in the cloud is key to better data-driven decisions that send your business initiatives soaring.

Our eBook “Setting Your Sights on Data Intelligence” is your pre-flight checklist for navigating your data landscape in the cloud and landing safely with horizon-expanding insights. You will learn about:

• Ensuring you can find, understand, trust, and access your data
• Understanding data relationships and flows
• Harnessing AI and machine learning to support data-driven decisions

Provided by

BT Group has confirmed a new machine learning (ML) platform has been deployed within the business that promises to shorten the development of artificial intelligence (AI) products from six months to six days.

The new platform is called AI Accelerator and is responsible for continually monitoring the health and performance of the AI models built by BT’s digital team.

AI Accelerator will specifically hasten both the administrative and technical process involved in ensuring an AI product or use case is fit for deployment, BT said.

The introduction of standardised tools and data handling make up a number of the platform’s key capabilities that will accelerate the deployment of safer AI models across the business.

Months, it said, will also be freed for the company’s data scientists and analysts thanks to streamlined software updates and monitoring.

From a performance perspective, the new platform will indicate any ‘drifts’ from baseline norms while the model is analysing data. This could involve showing alerts when analysis performance is becoming weaker or less accurate, for example.

There will also be parameters baked into the tool’s monitoring functionality that are designed to check if the AI models are in keeping with the company’s responsible technology principles.

AI is often scrutinised for the biases it can hold and the unethical ways in which it’s deployed but BT is aiming to make its models as cyber-secure and ethical as possible, with data privacy also a key priority.

The AI models developed by BT’s digital team underpin all the AI-driven technology that’s implemented across its products, pulling in data from a 29-petabyte estate BT calls its ‘digital brain’.

Potential use cases for the models soon to be developed in rapid time include products to help personalise services that BT already provides to customers.

Another is the development of propensity to churn models, it said. These models are designed to offer estimations on the likelihood that any given customer will abandon a company’s service within a given time frame.

“A core driving principle for BT Group is to find a way to safely accelerate the time to value, pound per petabyte, of data,” said Adrian Joseph, managing director of data and AI at BT Group.

“AI Accelerator gives us a path to more rapid value with clear oversight of AI use case parameters and performance.” Digital’s goal is to underpin over £500m of internal value from data and AI in aggregate, through improved customer experience and enhanced efficiency within each of the Group’s business units, over the next five years, and help power the transformation required to support BT Group’s long-term success.”

The new platform was built by BT’s digital team together with Datatonic, a London-based data and AI consultancy that’s also a Google Cloud partner.

BT and Google Cloud already enjoy a close relationship. Earlier this year, BT selected the cloud giant as its main partner driving to drive a company-wide digital transformation project.

“Thanks to the work the team has done accelerating our migration into Google Cloud, we have a healthy testbed for AI Accelerator and are seeing it catalyse AI acceleration across the Group,” said Dr Zoe Webster, artificial intelligence director at BT Group’s data and AI team.

“As we progress, we’ll be able to track the health and the performance of our AI use cases in real-time, ensuring consistent, safe, ethical delivery of value from our phenomenal data resources.”

The past few years have seen record wildfires in California, Europe and Australia, with the annual conflagrations worsened by hot, dry conditions caused by climate change. That makes fighting such fires all the more costly – financially and in terms of firefighters’ lives.

Even the UK faces more frequent and more intense wildfires, according to research by the University of Reading. The best solution is to reduce greenhouse gas emissions to rein in climbing temperatures as much as possible, but technologies such as sensors, artificial intelligence (AI) and drones are also lending a helping hand to predict, track and fight fires, helping to reduce the damage and keep firefighters out of harm’s way.

There’s a host of ways firefighters already use such technology. UK company FireAngel uses AI to predict incidents for at-risk people, such as dementia sufferers, while US government agency NIST developed AI to predict flashovers. This is the phenomenon when items in a room ignite all at once, only limited by the amount of available oxygen.

San Francisco startup Qwake Technologies, meanwhile, has developed a heads-up display (HUD) using augmented reality (AR) in helmets to help firefighters see their way through a fire, while 3M’s Scott Sight embeds thermal imaging into helmets. Then there’s our old friend Dubai, who have flaunted – but are yet to use – jet packs for fighting fires in tall towers. They’ve also developed jet skis called “dolphins” for accessing coastal conflagrations.

The best way to fight fires, though, is to predict where they’re going to happen – and spot them as soon as they light up. Machine learning systems to do just that are under development in the UK, and could help firefighters around the world.

Tracking where a fire plans to go

While the UK hasn’t been hit by wildfires at the scale of those seen in Australia, California or in Mediterranean countries, we are home to leading wildfire research. This is particularly true for the Leverhulme Centre for Wildfires, Environment and Society, where Dr Rossella Arcucci, a data scientist from Imperial College, leads efforts to bring machine learning to the wildfire fight.

That’s possible because the Leverhulme Centre for Wildfires has data ready to work with. Some AI systems are data-led, requiring training on massive sets of data. But others apply general models to existing data sets to help make sense of them. The latter is what Arcucci and her team of researchers are developing. “The models we develop are to be applied to lots of different applications, because we start from the point that we have data,” she explains. “The meaning behind the data comes from collaboration with experts in the field. As soon as you have good data and some physical meaning… you can work on the models.”

One project sparked by these ideas is led by Dr Sibo Cheng, using data assimilation and machine learning algorithms to improve and accelerate wildfire forecasting. His aim is to build a system using a recurrent neural network that lets anyone fighting a fire enter a few key bits of information and receive a prediction on how the fire will behave and where it will go next.

While Cheng’s wildfire work is data-led, there isn’t actually enough data on wildfires, as they remain rare enough incidents. To fuel his machine learning project, Cheng uses a physics-based simulator that simulates how fires grow and move. That model is already widely used to study fires, but running it is time consuming, meaning it’s not helpful for fighting active wildfires. “It will take several days or even a week to run a high-fidelity simulation,” he says. “So we’re investigating how to use machine learning to speed up the prediction of wildfires by learning from these physics-based simulations.”

The system is fed with compressed satellite photos of the fire, as well as key inputs such as time and location, as well as wind direction, local vegetation, slope and elevation. The model then looks at the area that’s already been scorched to understand how it’s propagated so far. A few seconds later, it will make a prediction based on that data, which can be handed to on-the-ground experts to help aid their decision making.

The aim, says Cheng, is to deliver a prediction about any wildfire in almost real-time, to help firefighters make better decisions and ensure evacuations happen effectively. “We’ll be able to predict with some certainty and uncertainty where a fire is most likely to propagate, so we’ll have a smart way to distribute the resources for firefighters,” he explains.

Arcucci adds that such “nowcasting” is important as it helps save lives of firefighters and local residents as well as reduce how forests are impacted by wildfire. Simulating different scenarios gives decision makers all the tools they need to take the best course of action.

Of course, no simulated system is perfectly accurate. So Cheng is pairing the simulator-based model with another technology that will act as a real-time predictive autocorrect for the system. “To make a real-time correction, we’re using a technology called data assimilation, which means that when you make predictions, you observe what is actually happening, and take a sort of smart average between what you simulate and what you observe,” he says. “In that way, you can better adjust your predictive model for the current time, and future predictions that can be more accurate. It’s an iterative process.”

So far, the system hasn’t been used on live fires, but tested on wildfires after the fact, with photos fed into the model from early on during an incident and the predictions compared with the eventual real-life outcome. “It’s not yet perfect,” he says, “but overall we have some quite good results, it’s quite encouraging.”

Arcucci notes, however, that what a mathematician deems imperfect is different from what’s useful in the field. If the predictions are out by a few metres here and there, that’s mathematically imperfect and will be tweaked in the algorithm – but it’s still incredibly useful data to firefighters on the ground. “For us mathematicians, we say ‘oh, it’s not perfect’ – but for people working in the field, it may be perfect,” she says.

Next, the team hopes to set up a user interface to make it easier to input the data for anyone, so firefighting teams don’t need Cheng to make use of the system. Plus, there’s work to apply the idea more widely by using social media. While the machine learning model works in near real-time, it requires satellite photos – and those take hours to download and process. To help remove that half-day delay, another Leverhulme researcher is looking at pulling in social media photos to spot where fires are happening, supplying images as well as geolocation. That requires smart filtering – someone posting a hot new outfit and declaring “I’m on fire today,” shouldn’t spark a call to emergency services – but if such a system can be made to work it could become an additional fire alert system globally.

Fire sensors by drone

What if we didn’t wait for a human to spot – and then tweet about – a fire? That’s the idea behind a system designed by Izzet Kale, head of the department of engineering at the University of Westminster. “Once you’ve spotted [a fire], it’s too late – you need to identify it very quickly and very early,” he says.

His solution is to drop hundreds of sensors across a fire-prone forest, pull in the data they collect in real time, and analyse it to spot if a fire has started by watching for rising temperatures or specific gases. To avoid unnecessary firefighter callouts, the system will automatically deploy a drone to fly over any potential fires to snap photos to confirm the sensor’s data indicates a new fire and isn’t just caused by an unseasonable sunny spot.

The research came out of a wider project into wildfires funded by the EU several years ago, that looked at detecting and better understanding fires; necessary as Mediterranean countries suffer worsening effects from climate change sparking dangerous fires in dry landscapes. “We’ve seen this every summer in Spain, in Portugal and Greece, with lives lost,” Kale says.

The sensor system’s aim is to spot fires as soon as they start, as well as to give information to firefighters on where and how to battle it. Plus, says Kale, the information gathered can shape evacuation efforts, so people aren’t accidentally sent into zones that are about to become dangerous. “For example, I think it was the Spanish fire last summer, lots of people were not killed because of the fire but because of smoke poisoning, because they drove into an area trying to escape fire where there was poisonous smoke,” he says. “Had they had an advanced sensor network, they would have known what the wind direction was, what the wind speed was, what the gas concentrations were, and which direction the smoke was going to go or was going in real-time. So you could actually have crowd control in this.”

The low-power wireless ground sensor nodes, as they were called, are set-and-forget. They have long battery life and harvest energy from their surroundings, Kale explains, waking up now and then to take readings of temperature, oxygen and carbon dioxide levels, humidity and other parameters, sending the data back to a server. The sensors also always collect their location, in case they’re moved by an animal.

Each sensor has a range of about 200 metres, though that depends heavily on local landscape; a large flat area will have better coverage. Installing them by hand was naturally onerous, as the aim is to cover as much of a forest as possible to spot fires not seen initially by humans. To help, the team miniaturised the sensors and equipment as much as possible, tucking them inside tiny balls – and then shot them out of drones. “The drones will drop them at any location on a map, and the sensor will active and go live,” says Kale.

The data collected by the sensors was augmented by intelligent systems: it was compared to a profile for the area built up over time, to help spot false alarms from sudden temperature changes. Basically, the sensor collects enough data to be able to spot when something goes wrong. “We can put some intelligence into it and distinguish between something that is a false alarm and something that is real,” Kale says.

Plus, there’s the drones. Once a threshold is met by sensor data, a drone is automatically sent out with the coordinates to fly over and collect photos as well as take chemical analysis of the local terrain, to better estimate how the local plants will burn. That data would then be passed to local emergency services. “Here is the biomass burning, here is the map of where the fire is now, and where it will be in five minutes,” he says.

Won’t the sensors burn up in case of a fire? Yes, some do – but you’ll know there’s a fire there, at least. “We make these out of high-temperature materials,” Kale tells us. “As a fire went, some of them died, and some of them continued to give information. When the fire comes to it and it stops working, you know that the fire is on it.”

At this point, firefighters could be deployed to the scene, but wildfires are dangerous and difficult to fight, so researchers came up with a system to fight wildfires by dropping biodegradable pellets from a cargo plane to smother it. “It was precision bombing with coordinates,” he says of that aspect of the project.

Did it work? “It wasn’t flawless, but it was almost foolproof in the sense that our sensors on the ground augmented information,” says Kale. “Once it decided it was a fire, we had a visual very quickly, and that was 100% accurate – you can’t really go wrong.”

So far, trials have happened in Greece, Spain and Israel, and Kale is hoping to commercialise the system to roll it out more quickly – especially as fires worsen each year. “The world has to move in the direction of knowing exactly what they’re fighting, what they’re dealing with,” he said. “It will save lives.”

Oracle Netsuite has announced NetSuite AP Automation, a banking solution designed to make it easier for businesses to process bills and pay vendors.

The new solution, launched at the company's annual SuiteWorld conference, provides firms with a full cloud enterprise resource planning (ERP) system that automatically captures and processes vendor bills.

Using AI (artificial intelligence) and ML (machine learning) object detection, the system is able to capture bills in a faster and more accurate manner than manual entry, and from sources including emails. Captured bills are converted into digital text.

NetSuite AP Automation also automatically creates bill records, with accurate filled fields for vendor details pulled in the capture process, and matches bills to ensure payment details are accurate. Bill matching is fully integrated with SuiteApprovals for a seamless oversight, review and approval process, with bills routed to relevant employees.

Through the use of optical character recognition, the NetSuite AP Automation is also able to accurately capture cheque payments, an important feature for the 81% of US businesses that still use this as a B2B payment method.

Many companies are already finding that automation is a powerful tool for increasing both efficiency and employee productivity, with back office automation freeing up workers to tackle more complex duties. In a landscape increasingly led by digital transformation, Oracle NetSuite is attempting to target those companies that lack the resources to build whole systems from scratch, and need easily deployable, pre-packaged solutions.

On transactions, Oracle NetSuite is continuing with its HSBC partnership, working with the bank to facilitate automated ACH, cheque or virtual credit card payments through NetSuite AP Automation. Customers can use their own current account to fund payments or the HSBC virtual credit card, which provides security for payments by generating single-use card numbers. The SuiteBanking Center will act as a dashboard for HSBC online checking and credit card accounts, and allows users control over transactions and payment statuses.

Greater integration between HSBC and the automated system also provides benefits when it comes to payment reconciliation, with NetSuite AP Automation’s “intelligent rules engine” used to flag irregularities for review, and automatically creating bank fee entries. Users are encouraged to set their own parameters during setup, for example allowing a degree of discrepancy between payment totals.

HSBC hopes that the efficiency of the new system would in and of itself drive businesses to adopt its virtual credit card system. Other benefits include the increased security of its single-use numbers, and the decreased time it takes for the bank to verify payments through the method in comparison to mediums such as physical cheque.

In contrast to simple filtering scripts, which can automatically action corrections to common mistakes such as typos, Automation AP's ML elements learn what to flag and what to change over time. For example, if a typo is flagged as appearing in the 'company name' field of a bill, a user with appropriate authority will be prompted to select which company the bill actually references. Future instances of this typo will result in the system automatically sorting through the problem on its own.

Similarly, the system will learn from a company’s finances in order to prompt alerts for any anomalous payments, as well as to anticipate regular payments without the need for developers to manually assign each one.

As part of future updates, Oracle NetSuite is working on a framework to allow companies better control over the rules that AP Automation learns. For example, under the current system it is possible that the system might take on a one-off payment as a ‘rule’ for future payments, and while this can be manually rejected by the firm, Oracle NetSuite has indicated that future releases will give more transparency over AP Automation’s decision-making process.

“Accounts payable plays an important role in helping organisations manage cash flow, control costs and maintain strong relationships with vendors, but all too often the process is slow, tedious, and error-prone,” said Evan Goldberg, founder and EVP, Oracle NetSuite.

“By simplifying and automating the entire bill payment process – from data capture to payment and reconciliation – NetSuite AP Automation eliminates these challenges. This helps businesses streamline and improve a key operation, empowering AP teams to operate more efficiently and ultimately reduce costs.”

NetSuite AP Automation is available now in the US, and the first 1,000 companies to purchase it will receive a 50% subscription discount, free implementation and bill scanning at no charge.

Hewlett Packard Enterprise (HPE) has been chosen by the King Abdullah University of Science and Technology (KAUST) to build its next-generation supercomputer to deliver supercomputing and artificial intelligence capabilities for research in fields like food, water, energy, and the environment.

The supercomputer, Shaheen III, is set to be 20 times faster than KAUST’s existing system. It will also be the most powerful supercomputer in the Middle East to address critical areas that have a societal and environmental impact, said HPE. The company hopes it will help KAUST process vast amounts of data at immense speed and scale and help users make new discoveries.

Shaheen III will be built using the HPE Cray EX supercomputer which it hopes will allow the university to apply significant computational power towards modelling and simulating scientific problems faster with higher accuracy. It’s designed with 25 HPE Cray EX supercomputer cabinets. The system will be made up of HPE Slingshot networking cables to deliver Ethernet fabric and closed-loop liquid cooling technology to remove heat. Each cabinet will also have 4608 CPU compute nodes, with two AMD EPYC processors, making up 884,736 cores in the entire system.

Through the design of the Shaheen III, HPE aims to help KAUST’s AI-at-scale mission by integrating the HPE Machine Learning Development Environment, an optimised software stack for model training and development. By combining the machine learning software platform with key supercomputing technologies, it hopes that KAUST’s users will be able to easily develop and train AI models faster, and bigger, for increased accuracy.

“The new HPE Cray EX system will allow us to conduct research on a larger scale, resulting in significant scientific, economic and social advances,” said Tony F. Chan, president of KAUST. “In line with Vision 2030, we strive to meet the ever-increasing demands of our active and solutions-driven faculty, and also those of external partners, for faster and more efficient computing resources. KAUST supercomputing resources are used by more than half of our faculty, students, postdoctoral students and researchers, and researchers from more than 20 external organisations in the kingdom.”

The supercomputer is expected to be fully operational in 2023 and will process data sets in areas like clean combustion, Red Sea ecosystems, climate modelling and the Arabian tectonic plate while delivering analyses, models and simulations at a superior level of resolution. It will also be used for the design of new materials for solar photovoltaics, personalised preventative healthcare, the discovery of new medicines, and increased hydrocarbon recovery.

The Shaheen III, equipped with over 2,800 Nvidia Grace Hopper superchips, is set to be used to develop novel and scalable techniques in core AI fields like deep learning, reinforcement learning, visual computing, and natural language processing. It’s also hoped to deploy AI in science and engineering applications, like computational chemistry, biology and material science, which previously were unfeasible given the large-scale nature of data sets produced and used at KAUST. The supercomputer is expected to deliver 100 petaflop/s of performance and outstanding AI modelling capabilities.

The Shaheen III is building on the experience of its predecessors, the Shaheen I which was launched in 2009 and the Shaheen II, a Cray-based supercomputer 25 times faster than its predecessor at 5.54 petaflop/s.

“Shaheen II has delivered 6.8 billion computing core hours to more than 1,467 users, resulting in data used in about 1,030 publications to date,” said KAUST Research Computing Core Labs director Dr. Jysoo Lee. “Shaheen III will ensure that the KAUST community will maintain its edge as a world-class university in computational sciences.”

Informatica offers a suite of cloud integration and cloud data management solutions, including Integration Platform as a Service (iPaaS), for connecting disparate applications and data deployed in diverse environments. Informatica Intelligent Cloud Services (IICS) is a comprehensive iPaaS consisting of data integration, application integration, and data management services for multi-cloud, hybrid environments.

To better understand the benefits and costs associated with an investment in Informatica iPaaS, Nucleus conducted an in-depth return on investment (ROI) assessment of several customers using Informatica technology. These customers realised an average ROI of 321 percent over a three-year period, with an average payback period of four months.

Download this guide to learn more.

Provided by

In 2021, organisations continued to embrace modernisation initiatives to improve agility and reduce administrative costs. Over the past year, companies have continued to deploy low-code environments to simplify integration processes, reduce costs, enable business users, and redeploy developers to more advanced areas and projects.

As we progress in 2022, increased demand for increased functionality in areas such as artificial intelligence (AI), machine learning, and automation to simplify integration tasks and reduce the cost of supporting continuously changing software environments.

Download this paper to read how, by complementing large-scale modernisation initiatives with efficient integrations and workflow automation, the Integration Platform-as-a-Service (iPaaS) market now has a record number of new users.

Provided by

Artificial intelligence (AI) threat detection and response provider Vectra has selected the Nuvias Group as its only strategic distributor across the UK.

Vectra specialises in using AI to detect cyber crime and help organisations thwart attacks before they become full breaches. The company’s platform enables unified threat visibility for public or hybrid cloud, software as a service (SaaS), identity and corporate networks, extending to environments such as AWS, Azure AD, and Microsoft 365.

The new distribution agreement builds on Vectra’s existing partnership with Nuvias company Cloud Distribution as the firm moves to harness the role of the channel in the UK.

The collaboration will leverage Nuvias’ range of partner enablement services, which utilise the latest business intelligence tools, to address the digital transformation needs of MSSPs, GSIs, and SIs, and their customers, Vectra said.

Vectra UK&I Regional Director Garry Veale said demand for intelligent, threat-led security is “at an all-time high” as the likes of ransomware and wiper malware continue to target enterprises.

“To meet demand, we selected the Nuvias Group, because we value its proactive approach to distribution, knowledge of the market, and technology expertise,” he explained.

“The Nuvias Group’s integrated offering will give more organisations the opportunity to use Vectra as part of a holistic solution and reduce their mean time to detect attacks. This is a logical next step to expand on the great work from Cloud Distribution”

Vectra said it will leverage Nuvias’ 1600-strong UK partner network, as well as its integrated solutions and ancillary services, to help its expansion from the SMB market into medium and large enterprises.

“With its focus on innovation and significant investments into R&D, Vectra fits perfectly into the Nuvias portfolio of cyber security and intelligent networks, complementing the existing range of solutions,” said Lee Driscoll, MD of Nuvias UK&I.

“Vectra’s NDR and cross cloud capabilities give commercial partners, SIs, and MSSPs a great value proposition to bring to customers, offering them a simple way to adopt advanced security measures and complete their SOC (Security Operations Centre) visibility triad.”

Meta has announced that PyTorch, its open source framework for machine learning, will become part of the Linux Foundation and reorganise under a new PyTorch Foundation.

The move, which comes after six years of Meta (formerly Facebook) control over PyTorch, comes as part of a move to emphasise the neutrality of PyTorch, which has become a key tool for use in the development of artificial intelligence (AI) and the acceleration of deep learning.

Since its launch in 2016, over 18,000 organisations have utilised Pytorch for machine learning (ML) in production environments, as well as to lead research throughout academia. Notable examples of PyTorch in real-world use include Tesla’s autopilot system, as well as the popular Python library Hugging Face Transformers, which provides users with pre-trained models for speech, text, and image recognition.

“The creation of the PyTorch Foundation ensures that decisions will be made in a transparent and open manner by a diverse group of board members for many years to come,” Meta stated in a blog post.

The Linux Foundation is a worldwide non-profit consortium for the tech sector, aiming to foster the development and curation of open source software. On its website, the foundation states that its goal is to “democratize code and scale adoption, for all projects equally,” and “strive to create new technology categories by identifying new trends and accelerating the growth of nascent technologies by removing barriers to adoption.”

Dr Ibrahim Haddad, vice president of strategic programs at the Linux Foundation, will serve as the PyTorch Foundation’s executive director, and the PyTorch Foundation’s governing board is expected to include a range of representatives from the tech sector. Founding members AMD, Amazon Web Services (AWS), Google Cloud, Meta, Microsoft Azure and NVIDIA were named in the announcement.

In May, the Linux Foundation announced its ‘World of Open Source’ research series, with the intention of furthering understanding of the scale of open source on the international, public sector stage, as well as surveying the overall state of open source within Europe. The results of the survey are expected to be announced at the Open Source Summit Europe, currently being held in Dublin.

Meta announced a partnership with AWS last year with the aim of expanding its AI offerings, and improving the ML models used to train Pytorch. Together, the companies aim to provide firms with the training and deployment of deep learning models at scale, such as large language models and automatic handwriting recognition.

“Growth around AI/ML and Deep Learning has been nothing short of extraordinary—and the community embrace of PyTorch has led to it becoming one of the five-fastest growing open source software projects in the world,” said Jim Zemlin, executive director for the Linux Foundation.

“Bringing PyTorch to the Linux Foundation where its global community will continue to thrive is a true honor. We are grateful to the team at Meta—where PyTorch was incubated and grown into a massive ecosystem—for trusting the Linux Foundation with this crucial effort.”

Cyber security-based artificial intelligence (AI) specialist Darktrace has teamed up with attack resistance management provider HackerOne to combine AI with Attack Surface Management (ASM).

The collaboration will combine Darktrace’s PREVENT/Attack Surface Management technology with the continuous security assessment capabilities of the HackerOne platform.

Building on HackerOne’s OpenASM initiative, the partnership is aiming to help organisations secure their digital estate through leading technology and a community of ethical hackers.

HackerOne said it recognised the need for an ASM partner that could enhance the asset discovery efforts of its hacker community. As a result, it selected Darktrace’s PREVENT/ASM offering, which uses AI to perform reconnaissance on a target attack surface by simply knowing the name of an organisation and identifying threats to that target.

In an announcement, the firm said the combination of AI and security expertise will deliver continuous insight to eliminate blind spots across a company’s digital landscape. Darktrace and HackerOne will also collaborate to train on ASM best practices.

"We are partnering with Darktrace because of their leadership in security, global coverage, and the flexibility of their asset discovery product to protect organisations of every type and size,” explained Ashish Warty, SVP of Engineering at HackerOne.

“The innovation and quality of their AI technology combined with the HackerOne Attack Resistance Management platform will improve the effectiveness of vulnerability discovery and provide the foundation for jointly developed security solutions that combine the best of human and machine intelligence."

AI and machine learning are fast becoming key facets of successful cyber security tools, enabling crucial functionality such as intelligent data analysis, automation, and identification of emerging threats.

When combined with ASM, organisations can more intelligently protect their infrastructure through continuous discovery, classification, and monitoring from an attacker’s perspective.

Back in June, HackerOne launched OpenASM, which helps organisations combine external scan data from ASM products with the firm’s own proactive security testing capabilities to assess attack surface risks.

Back in February, Darktrace also completed a €47.5 million (£39.6 million) acquisition of attack surface management startup Cybersprint, which offers penetration testing services, asset mapping and continuous, real-time insights.

With this new partnership, Darktrace said customers will now be able to quickly gain visibility of their external assets, while hacker expertise will enable targeted testing and data enrichment to address most critical risks.

"We have always embraced the hacker community, but with this partnership, we are making that more visible to the outside world,” explained Pieter Jansen, SVP of Cyber Innovation at Darktrace.

"By providing ASM data to HackerOne programs, participating hackers become more efficient at finding vulnerabilities and spend less time on reconnaissance.”

Visual inspection software provider Covision Quality has joined NVIDIA Metropolis, a partner program focused on enabling vision AI solutions for increased operational efficiency and safety across industries.

A spin-off of Covision Lab, Covision Quality’s visual inspection software builds on unsupervised machine learning. With just an hour of training, on average, this tool can minimize pseudo-scrap rates, a KPI in manufacturing, by up to 90%.

Additionally, workstations deployed at customer sites leverage NVIDIA RTX A5000 GPU-accelerated computing, enabling the software to process images, inspect components, and communicate decisions to a programmable logic controller (PLC) in real-time.

As an NVIDIA Metropolis member, Covision Quality has the added advantage of gaining early access to NVIDIA platform updates to further enhance and augment AI application development.

“Joining NVIDIA Metropolis marks yet another milestone in our company’s young history and in our relationship with NVIDIA, which started with our company joining the NVIDIA Inception program last year,” said Franz Tschimben, CEO of Covision Quality.

“It is a testament to the great work the team is doing in providing a scalable visual inspection software product to our customers, drastically reducing ‘time to deployment’ of visual inspection systems and ‘pseudo scrap rates’.”

“We expect that NVIDIA Metropolis, which sits at the heart of many developments that are happening in the industry today, will give us a boost in our go-to-market efforts and support us in connecting to customers and system integrators,” added Tschimben.

Advances in artificial intelligence (AI) and machine learning may increase the chances of reducing carbon emissions through carbon capture or geo engineering projects.

Ranveer Chandra, managing director of Research for Industry and chief technology officer (CTO) of Agri-food at Microsoft, has spent his career researching climate and agritech solutions. Microsoft, in this regard, is pursuing its own commitment to carbon reduction. Chandra says AI can help "full solution" geoengineering projects become more affordable, targeted and transparent.

With AI, researchers can better estimate locations for, as well as impacts of, ocean or solar geoengineering to decide on the most appropriate, effective approach. AI may also replace expensive simulations or augment process-based models – especially with new and promising technologies such as the Python-based Causal ML.

Tackling an existential crisis

Chandra, however, warns that huge challenges remain: "The first is the cost of these solutions, to be implemented at scale,” he tells IT Pro. “Second, there's a limited understanding of the full impact – primary and secondary – of any of these geo engineered solutions.”

Carbon capture and storage (CCS) is typically seen as more promising, but it's "still a very expensive way to remove carbon", he adds. "With AI, we are able to do seismic modelling at scale, at a speed-up of more than 1,500 times existing approaches that use partial differential equations and simulations."

It means better modelling of the flow of carbon and planning of CCS operations; Microsoft is working with a range of partners, including Nvidia, on developing AI and machine learning approaches that will power CCS projects.

Ongoing and planned projects could sequester a combined amount of approximately 40 megatonnes of CO2 per year. But to keep temperature rises at 1.5ºC versus pre-industrial levels, a hundred times the storage capacity is required, he warns. "This is an existential issue; therefore we have to help invent techniques to mitigate climate change," offers Chandra.

Professor Ted Shepherd, Grantham chair of climate science at the University of Reading, notes physics-based simulation models have been typically more reliable in terms of working out cause and effect than pure data-based approaches.

"Data science, as it is usually understood, is good at finding efficient solutions in situations with lots of data to explore, but climate intervention strategies are, by their very nature, out of sample -- that is, not represented in existing data," Shepherd says. "Data science methods tend to fall over when applied to out-of-sample problems."

Innovations in causal AI – which identifies underlying causes of a behaviour or event that predictive modelling fails to – can help lower the risk of unwanted outcomes by getting a much better picture of cause and effect relationships, Shepherd says.

Microsoft research projects driving CCS advances

Northern Lights: Working with Norway's government, Equinor, Shell and Total to standardise and scale CCS, for a North Sea storage reservoir from 2024.

KarbonVision: Using a computer vision approach to mapping geological faults from seismic data, reducing processing time it takes to detect potential leakage pathways of CO2.

Q-FNOs for 3D flow: Developing industry-relevant, scalable 3D simulations for CO2 flows and storage, involving typically complex, high-compute coupled PDEs.

Redwood: Working towards clusterless supercomputing on Azure, by building a more easily managed distributed programming framework on top of existing Azure HPC services.

Hyperwavve: Using a cloud-native fault-tolerant framework for hyperscale 3D seismic imaging, with Docker, Kubernetes and Dask parallel containerised seismic workloads at scale on Azure.

Innovating our way out of a climate crisis

Beyond CCS, more ambitious concepts include stratospheric aerosol injection (SAI) and marine cloud brightening (MCB). The idea is to reflect sunlight back into space to reduce global warming – either by spraying reflective particles into the stratosphere or 'seeding' clouds with salt crystallised out of the oceans, respectively.

Dr Vitali Avagyan, data scientist at TurinTech, notes that AI could help predict failures of CCS plants in real-time, as well as helping compare different, complex decarbonisation strategies.

"Rapid growth of environmental data from sensors, weather and climate models makes it difficult to interpret at speed," Avagyan says. "AI can help measure collective impacts of CCS on entire energy systems."

Shepherd notes that SAI seems "very feasible", but no one knows exactly how it would play out. What if, for example, an SAI implementation in one country caused, say, problems in another place? A failure of the South Asian monsoon would be a disaster, for example – suggesting governance challenges.

Dr Timothy Farewell, head of science at Dye & Durham, highlights the need for solid assessing, filtering and cleaning alongside a strong understanding of interactions and processes involved.

"Some blind AI or machine learning models will look to extrapolate beyond the range of training data to more extreme conditions – leading to serious issues with accuracy," Farewell confirms.

Jim Haywood, professor of atmospheric science at the University of Exeter, also tells IT Pro that greater knowledge of physical science is still needed to manage the risks and opportunities of SAI and MCB especially.

Shepherd agrees, adding: "We don't really have that much confidence in the regional aspects of climate change. You have lots of factors; you need a very structured way of doing it, and that's where the causal AI comes in."

Also, applying even well-established laws of physics to an atmospheric simulation means breaking up an earth-wide system into grid boxes – yet the equations needed should be continuous in space, rather than discrete. A typical grid box for a climate model might be 50 square-km in the horizontal and 1km in the vertical, Shepherd says. "It's pretty coarse. You're not representing a lot of things," he confirms. "There are uncertainties. Increasing the spatial resolution makes the model much, much more computationally expensive."

Ideally, an "ensemble" of runs, of lots of processes at very high spatial resolution, is needed to work out all the possible realisations as well, says Shepherd. "And different scientists will argue differently, for different trade-offs, as well as there being some processes that are not fundamentally understood – such as mixed-phase clouds of ice and liquid."

Discover why IBM ranked second highest in the Data Fabric Use Case according to Gartner analysts.

“A data fabric enables faster access to trusted data across distributed landscapes by utilising active metadata, semantics and machine learning (ML) capabilities. It is an emerging design; data fabric isn’t a common use case in the market yet. We picked it as a forward-facing use case — not every vendor has the full set of capabilities to deliver a data fabric design.”

Download this free Gartner® report to see how IBM is addressing this emerging design and has ranked second highest in the Data Fabric use case.

Provided by

.

Google Cloud and Vodafone have today announced the launch of a new artificial intelligence (AI) and machine learning (ML) platform dubbed the “AI Booster.” The fully scalable platform promises to spit improvements to customer experience, research speed and network performance.

Both companies are committed to expanding ML and AI use cases on an industrial scale, and this was the driving force behind the 18-month-long project.

Google and Vodafone say the collaboration will fuel the next generation of AI development, with Vodafone saying it can handle thousands of ML models per day. In the announcement, Vodafone credited AI Booster with an 80% reduction in time between proof of concept and production — from 5 months to 4 weeks.

AI Booster is built upon Google’s Vertex AI, which allows customers to design and deploy ML models at speed. Fully cloud-native, Vodafone promises that the platform will utilise all the benefits of its integration with Vodafone’s Neuron platform, with customers able to leverage powerful AI Booster environments simply by filling in an online form with their preferences.

Pre-trained tooling will complement custom options, with completely automated ML lifecycle compliance activities and embedded security giving customers the flexibility to deploy use cases as simple or complex as they need them to be, without fear of compromising their data.\

Google Cloud partner Datatonic has also provided reusable MLOps Turbo Templates, which allow developers to rapidly train models without running into technical debt and reducing deployment or adoption risks.

In 2019, Vodafone announced the integration of Google Cloud controls into its custom Vodafone Neuron platform, pointing to exciting opportunities the collaboration provided such as improved network optimisation, and even more detailed analytics to improve systems such as those handling commercial offers for customers.

Although ML models are still only gradually being utilised by organisations, Google is heavily invested in expanding the reach and user-friendliness of such models throughout its cloud, with recognition of the boosts to efficiency they can provide across all sectors.

“As a technology platform, we’re incredibly proud of building a cutting-edge MLOps platform based on best-in-class Google Cloud architecture with in-built automation, scalability and security,” stated Ashish Vijayvargia, Analytics Product Lead at Vodafone.

“The result is we’re delivering more value from data science while embedding reliability engineering principles throughout.”

OpenAI has announced GPT-4, its latest model that will contribute to the rapidly-developing field of AI in business. Organisations are already making use of increasingly sophisticated AI tools throughout their day-to-day operations, and OpenAI’s GPT-3 and GPT-3.5 have become particularly popular in the last six months, powering the headline-grabbing ChatGPT.

The new model is multimodal, meaning users can use both text and images as inputs and has been heralded as the most creative and powerful yet made by OpenAI. The firm has also stressed that GPT-4 is safer than previous versions, adhering more rigidly to the boundaries set out by its developers and users without sacrificing its capacity for creative output.

GPT-4 vs GPT-3

Across all metrics, GPT-4 is a marked improvement over the models that came before it. Putting aside the fact that it can handle images, long something that has evaded OpenAI’s previous GPT iterations, it is also capable of more nuanced, reliable, and challenging output than GPT-3 or GPT-3.5.

In simulated exams designed for humans, GPT-4 greatly surpassed the results of GPT-3.5. It placed in the 90th percentile for the Uniform Bar versus GPT-3.5’s 10th percentile results and 88th percentile for the LSAT versus GPT-3.5’s 40th percentile. GPT-4 also scored highly in the Massive Multitask Language Understanding (MMLU) benchmark, which puts models up against 14,000 multiple choice questions based on a range of topics not covered in training sets.

OpenAI also claims that GPT-4 outperformed competing large language models (LLMs) like Google’s PaLM and Meta’s LLaMA in a range of machine learning (ML) benchmarks. In the MMLU GPT-4 scored 86% versus GPT-3.5’s 70% and PaLM’s 71%.

Hallam back-end developer Oliver Fokerd notes that while GPT-3 enabled users to input natural language, it still took a bit of skill to craft your prompt in a way that would give good results. “GPT-4 will be much better at inferring users’ intentions,” he adds.

GPT-3 was once thought too dangerous to be released to the public, after early versions showed a propensity for generating misinformation. The model was also accused of exhibiting racism against specific religions and genders. More recent versions use reinforcement learning from human feedback (RLHF), which uses human helpers called labellers to assist the AI in its learning and flag unwanted responses.

To combat these same fears, OpenAI has committed to more rigorous RLHF for GPT-4 to minimise harmful outputs.

“Our mitigations have significantly improved many of GPT-4’s safety properties compared to GPT-3.5,” OpenAI states in its blog post on GPT-4.

“We’ve decreased the model’s tendency to respond to requests for disallowed content by 82% compared to GPT-3.5, and GPT-4 responds to sensitive requests (e.g., medical advice and self-harm) in accordance with our policies 29% more often.”

With regards to the structure and size of GPT-4 versus GPT-3 , OpenAI is keeping its cards close to its chest. The firm has not stated how many parameters GPT-4 has in comparison to GPT-3’s 175 billion, only that the model is “larger” than its predecessor. It has not stated the size of its training data, nor where all of it was sourced aside from "a large dataset of text from the Internet".

GPT-4 has required such a large overhaul that OpenAI rebuilt its entire deep learning stack for the model. It also made use of a supercomputer built in partnership with Microsoft Azure, for which GPT-3.5 was used as a dry run to ensure stability issues were resolved.

What are the business benefits of GPT-4?

A standout feature of GPT-4 is its multimodal nature, which allows businesses to feed it a wider range of data than ever before. Firms can use the model to summarise PDFs, aggregate and contextualise chart data, analyse and critique contracts, or quickly identify visual irregularities in physical infrastructure.

OpenAI’s president and co-founder Greg Brockman demonstrated the degree to which GPT-4 can extract and process information from images by sketching a mockup of a website into a sketchpad, which was then translated into the working HTML for the website by GPT-4.

The model’s image processing capabilities also make it an ideal workhorse for computer vision systems. OpenAI has stated that in addition to recognising objects and scenes within images, the model is capable of providing suggestions based on context within these images.

The app Be My Eyes, which pairs people who are blind or have low vision with volunteers who help them navigate everyday tasks via video call, has tested a virtual volunteer with GPT-4. It was able to provide recipe ideas based on an image of a fridge’s contents, but could also assist with complex tasks such as navigating a train system.

The firm is bringing the same tech to screen readers, enabling disabled users to access websites and computer systems through advanced descriptions of the key elements on a page, summaries of search results, or talking users through content management systems that were not built with navigability in mind.

Its prowess with languages other than English also opens up GPT-4 to businesses around the world, which can adopt OpenAI’s latest model safe in the knowledge that it is performing in their native tongue at a higher level than GPT-3.5 could in English. It also has potential for international trade and contract negotiations, with the model capable of translating, summarising, and analysing written text and screenshots in one fluid process.

This could be especially helpful for AI use in cyber security, with firms like Secureworks already having found success in using LLMs to aggregate and translate threat actor posts on the dark web.

Another new feature of GPT-4 that businesses could find particularly useful is its greater adherence to the boundaries set out by its training and by users. OpenAI highlights its improved ‘steerability’, which keeps GPT-4 in line with instructions while reducing "hallucinations", a common issue in generative AI where models begin to confidently lie.

Brockman demonstrated the potential of this steerability by instructing GPT-4 to act as ‘TaxGPT’, a chatbot that can read and apply the tax code while explaining its steps in detail. It was then able to provide advice on the standard deductions available for a theoretical couple.

Businesses can use the same steerability functions to improve the user experience of chatbots that will eventually run on GPT-4 through ChatGPT API, or make internal systems more helpful and company-specific. For example, a chatbot that explained company processes and regulations could be tailored by job description and experience so that employees always receive information at an accessible level.

This could also prove useful for training purposes, which are already being demonstrated by Khan Academy in its use of GPT-4 as a source of knowledge and context for students seeking to learn maths, science, and humanities. The same principles apply to the GPT-4’s vast array of knowledge, meaning business leaders could use the model to upskill workers on subjects like programming or other languages to beat talent shortages rather than let the model replace workers altogether.

When can my business use GPT-4?

If your business is subscribed to OpenAI’s paid tier ChatGPT Plus, you now have access to GPT-4 as a chat model. Firms that wish to integrate GPT-4 into their stack via ChatGPT API will have to wait a little longer, however, with OpenAI currently accepting applications for its GPT-4 waitlist.

Firms that submit model evaluations deemed high quality through OpenAI’s ‘Evals’ service may also receive access to GPT-4 early. So if GPT-3 or GPT-3.5 are already part of your stack, this could be a route for early adoption of the powerful new model.

On the other hand, if your business is using Microsoft’s Bing AI then you may have used GPT-4 without even knowing it. The Redmond giant has admitted that its search engine AI has been powered by the new model all along, giving some indication of the performance that can be expected from GPT-4 on a widespread scale through chatbot boundaries.