Open weights, closed secrets: Nathan Lambert on the frontier gap and Trillium
Key points
Key takeaways from an a16z interview with AI researcher and Trillium founder Nathan Lambert (October 2026):
Frontier progress is capital, not magic. Model progress comes down to compute, data and talent, and OpenAI, Anthropic and Google have had the most of each. Organisation matters as much as budget: OpenAI used its inference-scaling laws and reasoning-model push to drive the frontier, while Anthropic focused on coding for so long that it became Claude Code, and a bottom-up reset left Meta back on a "positive trajectory".
Chinese labs go open as a business strategy. Releasing open weights is how Moonshot AI and other Chinese builders get into the global conversation — a practical distribution choice, not a government directive or altruism. They are technically excellent and talent-equal but compute-starved: Lambert says an average OpenAI researcher can draw on around 5,000 H100 equivalents, a figure that leaves pre-training researchers in China stunned.
Distillation helps by "a few months", not years. The lag from open to frontier models has compressed to roughly two to three months, and training on a frontier model's outputs is effectively buying compute. But the open post-training literature shows how hard it is to map today's reasoning traces into a training signal, so Lambert is sceptical the established distillation story still holds. Open reasoning APIs leak "thinking tokens" that keep this copyable; moving best models into first-party products would make it far harder.
"That's called competition." Quoting Jensen Huang — "people distill my products every single day… that's called competition" — Lambert is unsympathetic to policy crackdowns on distillation: Chinese labs will keep doing it, the open-source and startup communities benefit, and turning terms-of-service violations into government enforcement would cause more harm than good.
Open and closed run on different time scales. Closed labs are vertically integrated — efficient inference and early entry into the highest-value markets, with coding revenue going from roughly zero toward $100B. The open-model economy takes far longer to build because information and revenue do not flow back to the model builder, and it involves more companies, such as Thinking Machines Lab or Prime Intellect specialising in fine-tuning. The payoff is ownership: weights you can run on your own hardware with no terms of service.
The frontier is now a data and environments game. In domains that are hard to verify — biology, healthcare, legal — progress comes from buying high-quality data and giving models worlds in which to act. Labs are pulling data work in-house partly to stop signal leakage, and the boom in reinforcement-learning environments is opening room for reward hacking as vendors cut corners on world-building quality.
Trillium will open up the research the labs keep internal. Named for the spring ephemeral flower that blooms before the forest canopy, Lambert's nonprofit will release open-source RL environments for high-leverage tasks such as terminal use, an open stack for supervised fine-tuning and mid-training, and published trajectories and configs so outsiders can see what "recursive self-improvement" actually looks like in practice.
Fast takeoff is not the near-term risk. Scaling laws still mean exponential compute for roughly linear gains, and human bottlenecks remain. Lambert is more worried about the thin public preparation around cyber security than about models building themselves, and argues more open research and more independent eyes keep society better informed.
Read more: a16z episode · Trillium · Nathan Lambert's Interconnects
