With artificial intelligence (AI) being all the rage right now, speculating about its future might be premature, especially while so many businesses are still getting their heads around its capabilities, and how generative AI might work for them.
But academics take a much broader view, building on a long corpus of research to anticipate AI’s potential as well as its challenges as the field continues to evolve.
While they agree it’s an exciting time, the craze around tools like ChatGPT is simply one facet of a much larger beast. There’s much to still figure out on the business side, such as monetization strategies, but there are also plenty of ethical and security challenges.
Three experts, speaking on a panel at Databricks’ recent Data + AI Summit 2023, charted a path for how AI might develop in the future – as well as everything enterprises must bear in mind as we collectively drift towards the future of AI.
Four reasons to be excited about the future of AI
1. Mass availability of LLMs
As part of his research into AI over the course of 15 years, Michael Carbin, associate professor at MIT’s Department of Electrical Engineering and Computer Science, has been exploring how to take larger models and make them more accessible for everyone.
“These models aren’t only the in the hands of the largest organizations, they can be in everyone’s hands,” he says during the discussion.
Carbin played a founding role in the AI startup MosaicML, which Databricks recently acquired for $1.3 billion. Companies like this aim to equip organizations with the tools to create their own generative AI systems in-house, using first-party data, rather than, say, scraping masses of information from the internet. Such services might see the barrier to entry lowered for enterprises hoping to take advantage of AI in their own specific industries.
2. Liquid networks
One of the most exciting innovations to emerge from MIT in recent years has been the idea of liquid networks, and it’s the current focus of Daniela Rus, professor of electronic engineering at MIT and director of its Computer Science and Artificial Intelligence Laboratory (CSAIL).
Achieving transformative business results with machine learning
How seven leading organizations are using machine learning to reveal new opportunities
Described as a new approach to machine learning, liquid networks take a compact approach as well as a continuous time approach for applications with time-series data. They’re provably causal, and have nice generalization properties, Rus says, when trained in one environment and applied in another, with huge distribution shifts.
“While the world is trying to make networks bigger and bigger, I want to make them smaller,” she says, before highlighting an example of liquid networks in action. “If you want to get a robot car to stay in lane and steer, it takes about 100,000 different networks to get good behaviour, and it only takes 19 of our liquid networks.
“If you have a solution that relies on a couple of nodes, or tens of nodes, you can then extract decision trees to explain how that system makes decisions. In doing so, you can begin to address safety-critical applications, where knowing how the model reaches decisions is very important.”
3. Smaller – not larger – models
On the theme of models becoming smaller, not larger, Carbin highlights small data as being essential to the AI models of tomorrow.
Right now, to build models that generate code, for example, you would “find as much data as you possibly can” and scrape datasets that are “large and pretty dirty in some way”, throw them into a large – but generic – model. But models with fewer but higher quality data points would be cheaper to build, and require fewer resources to compete with the likes of ChatGPT, claims Carbin.
“If you do a very good job of trying to curate your data, you can find these opportunities to process on orders of magnitude less data,” he says.
“ChatGPT’s absolutely amazing,” he adds, “but when you can focus on your particular domain and use case, I think there are really good opportunities for building these models, and building these models yourself, because scale is accessible to you.”
4. Multimodal AI
The AI models of tomorrow will be multimodal – meaning they can be trained on a vast variety of data at the same time including video, audio, speech, images, text, and numerical data sets, among other sources.
Natural language processing (NLP) might be considered a form of multimodal AI, for example, because it combines text with speech recognition. But future systems may incorporate all types of media, not just a handful, raising the sophistication of the technology exponentially. It’s fundamentally a new paradigm for AI that will bring the technology much closer to human levels of decision making – and Carbin says we’re within touching distance.
“I just want to see multimodal. We’re in this age where we suddenly have the capability to have multimodal interactions with computers. And that’s going to bring so much new capabilities to so many people who could previously only access computation via programming, and perhaps people were excluded, that’s just the future.”
Three reasons to worry about the future of AI
1. Unresolved security concerns
Amid all the excitement, AI practitioners don’t think enough about how services will be misused, says Dawn Song, professor of computer science at UC Berkeley, speaking on the panel alongside Carbin and Rus.
These models have plenty of security and privacy challenges. For example, as AI services will be “exposed” to the general user, there are concerns around responsible usage, she explains. They can easily be jailbroken, they can be fooled into giving wrong answers, and it’s easy to plant backdoors into them. Finally, models can be triggered under specific conditions to misbehave.
“One of the really big challenges for practitioners deploying these models is how do you know when the model is actually ready to be deployed," explains Song. "How [do you] even measure how the model essentially behaves? What are the metrics for measuring the trustworthiness and the different aspects of the model?”
2. AI’s massive carbon footprint
There are three categories of AI right now, according to Rus; solutions around pattern recognition, systems primarily designed to come to a decision, which is where reinforcement learning comes in, and finally there’s generative AI.
“In each of these three categories we have issues; we have issues around data, because they all require a lot of data. And that means the computation is huge,” says Rus. “That also means there’s a large environmental footprint.”
Deploying a very small model, by today’s standards, releases 626,000 pounds of carbon dioxide into the atmosphere, Rus adds, which is equivalent to the lifetime emissions of five cars.
“As we think about deploying AI in the future, it’s important to keep these numbers in mind, because we want tools that support you run it, but also [support] the planet.”
3. Big tech monopolizing AI
“We talk about these personal assistants that help individuals, but who actually controls these personal assistants?” posits Song. One significant danger as the fledgling industry evolves is that companies like Microsoft and Google dominate the field and shape development.
The alternative of this vision relies on an expansion in, for example, open source development, as well as in-house development of generative AI systems in which individual enterprises own the IP to the models they use. This is as opposed to simply leasing general-purpose services offered by the aforementioned firms.
