AI Race: Why Anthropic stands a good chance to pull ahead

The first distinguishing aspect of Anthropic is the composition and quality of its leadership. At its helm is Dario Amodei, who, along with three other co-founders, are former employees of OpenAI. They were instrumental in making the advances that made Chat GPT-2 possible, and they are also responsible for much of the foundational research that led to breakthroughs in LLMs.

Amodei, in particular, has a rich pedigree in the domain of Machine Learning and AI, and variegated academic exposure to supplement his technical acumen. He possesses an undergraduate degree in physics, a Masters in neuroscience, and a PhD in Physics. Further, he possesses a stellar track record of working with stalwarts of AI and Machine Learning.

He worked at Baidu with one of the pioneers of the field, Andrew NG. He then had a yearlong stint at Google Brain, which released the seminal paper “Attention is all you need” providing the base for models like ChatGPT. And, after that, he pioneered the research at OpenAI for five years. Overall, Amodei and his team from OpenAI have an enviable track record and have operated at the cutting edge of the field for a long time.

Constitutional AI

The next aspect that distinguishes Anthropic is its developmental process. It has an approach to AI model training that they refer to as Constitutional AI. Here, like in a Constitution, the model is given an explicit set of principles which it must adhere to, during training and deployment. This differs from the approach that has been adopted by ChatGPT for instance, where human feedback is used to build guardrails around the model, along with other methods like meta-prompting and reinforcement learning.

This deeply ingrained focus on principles by Anthropic for its models has given its AI chatbot Claude an edge over ChatGPT in various instances where tricky questions need to be answered without coming across as offensive. Arthur AI, a machine learning monitoring platform, found Claude 2, the second generation of Anthropic’s chatbot, to have the best self-awareness among all major players, including those by OpenAI and Meta.

Another interesting aspect of Anthropic’s model development is how it uses its constitutional AI framework to develop safe AI models much faster. To illustrate, imagine a case where the firm has spent millions of dollars and several months of training time to train a powerful AI model which aligns with the principles in their Constitution.

Now, once this model is available, it can be used to test every other new model that is being trained, to see if the output of the newer models is aligned to the stated principles of the constitution and reorient them if it isn’t. This method reduces training time and cost, produces a more balanced model, and give much greater peace of mind to the developer and client.

Anthropic’s Billion Dollar Bets

Thus, Anthropic boasts of a good team with a proven track record, a unique take on AI development, and a strong focus on alignment with principles. Besides these facets, Anthropic also stands out due to the size of its ambition.

Scaling, a popular term in the AI lexicon and something Amodei is a firm believer in, basically refers to the phenomenon that when more processing power and larger datasets are deployed for training a model, the performance of the final models improves as well.

Using this hypothesis, which is holding true for the time being, Anthropic has embraced an impressive target of where they want to take their models and the industry in the future. From the $100 million models that populate the industry’s cutting edge in 2023, Anthropic plans to transition to billion-dollar models by 2024.

They have also stated their intent to spend around $5 billion in the next two years to enter half a dozen industries. And, right now, they are also building Claude-Next, their frontier model which will be 10 times more capable than today’s most powerful AI, and will cost them over a billion dollars to build.

Thus, the industry in general and Anthropic in particular is poised for a massive bout of acceleration in the coming two to three years. Amodei, who has been a leading voice for AI regulation, has been unequivocal in stating that capabilities are bound to dramatically rise in the coming years, with AGI or Artificial General Intelligence also not being too far off. It will be interesting to track the trajectory of the breakthroughs that this new scale achieves in the coming year, and what role Anthropic plays in guiding it.

Srimant Mishra is a computer engineer and lawyer with a deep interest in the field of Artificial Intelligence. Views are personal, and do not represent the stand of this publication.