Anthropic could be the champion AWS and Google need

AI firm Anthropic is enjoying a period of immense public interest in the wake of large investments from AWS and Google, as the firm moves to compete with OpenAI.

AWS announced up to $4 billion investment in Anthropic in September, with the two firms penning a deal to make AWS Anthropic’s primary cloud provider and to speed up development of future models available for AWS customers. 

Now, Google has agreed to invest $1.5 billion in Anthropic in addition to the $500 million it has already sunk into the AI startup.

With this potential $6 billion in announced AWS and Google investments alone, Anthropic has quickly become a company of interest in the AI space and could become more of a direct competitor with the likes of OpenAI.

Anthropic’s founders, Daniela and Dario Amodei, formerly held the respective roles of VP safety and policy and VP research at OpenAI. The pair left the company shortly after working on its large language model (LLM) GPT-3, citing their strong beliefs over approaches to AI safety. 

In January 2023, reports stated that Microsoft was investing $10 billion in OpenAI as it put the firm at the center of its AI product strategy. In the months since, Microsoft has integrated OpenAI’s multimodal LLM GPT-4 throughout its product offerings such as Bing Chat and 365 Copilot. 

Microsoft has provided OpenAI with crucial infrastructure for training its models, and reportedly built a supercomputer using Nvidia AI hardware to help train GPT-4.  

This is not the sole reason that OpenAI continues to lead in the AI space. The firm has a strong track record of producing powerful AI models, and though models vary in their effectiveness for different use cases, GPT-4 is still considered the most powerful all-rounder.

A supercharged Anthropic

With Anthropic set to receive infrastructure support from AWS, the firm is now in a position to demonstrate its own unique capabilities when it comes to AI without limitations on its model inference or training.

AWS’ Trainium and Inferentia chips were custom-built for models with hundreds of billions of parameters, and offer 50% and 40% cost improvements over Amazon EC2 respectively. Anthropic will be able to use these chips to improve the performance and competitiveness of its models. 

With this in mind, Anthropic’s successful challenge against OpenAI will rely to a large extent on whether Amazon’s hardware can remain competitive with Nvidia’s. AWS’ choice to rely on its in-house hardware instead of Nvidia’s, as Microsoft and Google Cloud have done, could tee up a fundamental shift in the public cloud 'big three'. 

Having placed its bets with AWS, Anthropic could be dragged along for the ride on this eventual head-to-head with Nvidia, and could either become a misstep for AWS if it fails to deliver, or held up as proof of AWS’ AI legitimacy if future Anthropic models can match or exceed OpenAI’s. 

In the short-to-medium term, the firm can expect rapid customer growth and interest as a result of Claude being included on Amazon Bedrock, the firm’s generative AI platform. 

What is Claude and how is it different from ChatGPT?

Anthropic’s headline chatbot is named Claude, and even prior to its public release the firm was billing it as a safer alternative to ChatGPT.  

Claude offers several benefits over ChatGPT, such as faster processing times - even on its free tier - and the ability to provide up-to-date information. 

Another benefit of Claude over ChatGPT is its far larger context window: 100,000 tokens of information, roughly equivalent to submitting an entire novel as an input, versus ChatGPT’s 32,000 for paid subscribers.  

The firm follows the principle of Constitutional AI, in which generative AI foundation models are trained to follow principles set out by a written constitution at the highest level of their operating parameters. 

It has said that this allows it to overcome the limitations and risks of using reinforcement learning from human feedback (RLHF), in which humans assess two AI model outputs and identify the least harmful of the two. 

Anthropic believes RLHF takes too much time, cannot be easily scaled to keep up with the expansion of AI model output, and carries the potential to expose human workers to upsetting outputs.

Instead, its models are trained in two AI-driven training phases. In the first phase a model generates outputs based on its initial dataset, revises these based on constitutional principles, and is then trained on these revised outputs. 

The second phase sees the fine-tuned model trained via reinforcement learning led by a paired AI model, to further reduce harmful output.

Claude’s constitution was written by Anthropic employees, with the aim of reducing bias and improving the emphasis the model puts on concerns for human rights.  

Anthropic has also tested the use of a constitution written using the responses of a public consultation it ran involving around 1,000 members of the US public. 

“The process of training a language model to abide by qualitative public opinions involves a large number of subjective judgment calls,” Anthropic wrote. 

“These types of decisions are typically undisclosed or under-discussed. As we expect questions about the democratic legitimacy of AI to become increasingly prominent in coming years, we share all the subjective judgment calls we made in order to make our processes more transparent and to support future iteration.”