Anthropic CEO says cancer, mental illness 'will be relics of the past' due to AI

Anthropic concluded their IPO roadshow yesterday, and their roadshow video contains some new details regarding their business operations. Their focus on the "Approach to compute" segment, featuring co-founder Tom Brown, is particularly noteworthy. Brown explains how Anthropic trains and serves Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs.

This multi-chip strategy was an unconventional choice back in the early days, as many believed that running AI models on different chips wouldn't be worth the effort. However, Anthropic has now honed this approach to such an extent that it takes just three days, with minimal human intervention, to bring Claude up and running on a new chip platform. For instance, they successfully utilized AMD GPUs for the first time this year, needing only three days to get Claude operational, with most of the work done by Mythos.

The second key factor in Anthropic's compute advantage lies in their discovery rate of compute multipliers. They concentrate on total effective compute, which is the product of physical capacity and the compute multipliers they've discovered. For example, a 4x compute multiplier allows them to train a model with the same intelligence using only one-fourth of the physical units.

As they discover more multipliers, they stack on top of each other, creating a cumulative advantage. Anthropic believes this can accelerate further as Claude approaches recursive self-improvement, potentially allowing Claude to discover new multipliers on its own.