Anthropic Co-founder: Building Claude Code, Lessons From GPT-3 & LLM System Design
Tom Brown, Anthropic co-founder, explains how scaling laws, Claude 3.5 Sonnet’s surprise breakout, and Claude Code’s internal origins shaped Anthropic’s trajectory.
- Claude Code started as an internal tool hacked together by engineer Boris for Anthropic’s own engineers, not as a planned product.
- Claude 3.5 Sonnet’s success surprised even the Anthropic team; so did 3.7 Sonnet’s agentic coding unlock.
- Anthropic deliberately avoids benchmark-teaching teams; Brown suspects that is the primary reason Claude outperforms benchmarks in real developer use.
- Global AI compute spend is on a roughly 3x-per-year trajectory; Brown says the buildout will exceed Apollo and Manhattan projects combined by next year.
- Power — especially US grid capacity and permitting — is Anthropic’s biggest infrastructure bottleneck, not GPUs.
- Anthropic uniquely runs GPUs, TPUs, and Trainium simultaneously, trading engineering complexity for chip flexibility and total capacity.
- Brown got into OpenAI by cold-messaging Greg Brockman, positioning himself as an engineer who knew both ML and distributed systems — a rare combo at the time.
- Scaling laws held over 12 orders of magnitude, which Brown says is what convinced him to pivot entirely into scaling research at OpenAI.
2025-08-19 · Watch on YouTube
