Can Open Models Keep Up With OpenAI and Anthropic? | Nathan Lambert

gist

AI researcher Nathan Lambert tells a16z partner Lisha Li that OpenAI and Anthropic lead on compute, capital and organization, not talent. Chinese open-model labs match US talent but are compute-starved, and distillation explains less than people assume. He expects open and closed models to persist as tiers, and launches Trillium, a nonprofit that will publish RL environments and document how AI automates AI research.

ideas

• Compute is the moat. Lambert frames progress as compute, data and talent; Chinese labs are equal on talent, but an average OpenAI researcher reportedly has thousands of H100 equivalents on tap.
• Distillation is overrated. Post-training has moved from SFT on strong-model completions to RL-heavy environment training, where copying outputs from frontier models is much harder to cash in.
• Data and environments drive the frontier. Gains outside math and code come from expensive, iterative data purchases and RL environments, a booming industry with quality and reward-hacking problems.
• Open wins on a longer clock. Closed labs monetize premium markets like coding first; open weights spread later into privacy-sensitive and cost-optimized deployments where firms own fine-tuned models.
• Demystify recursive self-improvement. Scaling laws give linear gains for exponential compute, so Lambert doubts fast takeoff; Trillium plans to publish how its own work gets automated, with check-ins every six months.

quotes