Oct 14, 2026 · 60 min
Open models seek leverage beyond frontier-scale spending
Can Open Models Keep Up With the AI Frontier? | Nathan Lambert
The future of AI competition may depend less on matching frontier labs’ budgets than on opening the tools, environments, and specialized markets around them.
- 1Frontier labs compound advantages in compute, capital, data, talent, infrastructure, and inference spending.
- 2Open models can compete through distillation, faster post-training, interactive environments, and specialized domain markets.
- 3Trillium aims to make frontier progress more transparent through open research tools, evaluations, and alignment infrastructure.
Don't miss
Lambert explains why Trillium will focus on open post-training infrastructure rather than attempt to reproduce the full cost of frontier pretraining.
The brief
Nathan Lambert and Lisha Li frame open versus closed AI as a contest between frontier labs’ capital and infrastructure advantages and open models’ reach, talent, and adaptability.
Lambert argues that compute, data, and talent drive progress, but distillation and international open-model efforts complicate attempts to preserve a lasting frontier-lab lead.
The conversation turns to reinforcement-learning environments, synthetic data, and reward hacking: flawed task specifications can make purchased training data look more useful than it is.
Lambert presents Trillium as a nonprofit effort to build open environments, evaluations, and tools for post-training, alignment, and research into recursive self-improvement.
The closing argument is less about one model winning than many specialized models competing across knowledge work, while society struggles to prepare for rapidly advancing capabilities.