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

TL;DR

Tom Brown (Anthropic cofounder) recounts his journey from MIT graduate and early employee at multiple YC startups, through engineering at OpenAI (GPT-3 scaling from TPUs to GPUs) and Google Brain, to cofounding Anthropic.

Key insights

• The scaling laws paper (Dario Amodei’s) showing a straight line over 12 orders of magnitude was the primary motivation to pivot all work toward scaling – a counterintuitive insight at a time when many
• Anthropic deliberately does not have a team dedicated to improving public benchmark scores because that incentivises “teaching to the test” and creates a train‑test mismatch; they rely on internal ben
• Claude Code began as an internal tool hacked together by an Anthropic engineer (Boris) to help other engineers, and its external success was a surprise – the company had fully bet on the API as the on