Anthropic Co-founder: Building Claude Code, Lessons From GPT-3 & LLM System Design

Brown describes learning to take initiative at early startups, where the team had to find its own direction. After Grouper struggled against Tinder’s better solution to the same problem, he spent six months studying machine learning, linear algebra, and statistics before joining OpenAI to build a StarCraft environment.

At OpenAI, Brown shifted toward scaling after seeing research showing that more compute, with the right training recipe, reliably improved model capability. He says the safety and scaling teams that founded Anthropic took the stakes seriously, though the seven co-founders initially had little certainty about what product to build.

Brown focused early at Anthropic on training infrastructure and securing compute. The team hesitated to build serving infrastructure before ChatGPT launched, a delay he describes as a lesson. Claude 3.5 Sonnet’s coding performance brought stronger demand, and Claude Code began as an internal tool built by engineer Boris for developers at Anthropic.

Brown says Claude Code’s team treated Claude itself as a user, giving it tools and context to work effectively. He sees a broader opportunity in helping models handle business tasks that require context and coaching, beyond pair programming. Anthropic also aims to offer a strong API platform for developers building those tools.

Brown describes power as a major bottleneck for data center growth in the US and supports making it easier to build data centers. Anthropic uses GPUs, TPUs, and Trainium to increase capacity and match hardware to workloads, while accepting the added engineering effort. His advice to younger people is to take more risks and work on things they would be proud to succeed at.