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

Summary · 6 min read instead of 59 min

Nathan Lambert is an AI researcher who co-led post-training at AI2, where he built OLMo, and he has written a book. He has since left AI2 and founded Trillium, a nonprofit research lab that plans to release RL environments and document how AI is actually being used to automate AI research.

Nathan Lambert argues that OpenAI and Anthropic lead because they have the most compute and take the most ambitious bets earliest, while open-model labs, especially in China, match them on talent and follow close behind as data and ideas spread. He reads the open literature as showing distillation adds only a few months, and says acting against it would do more harm than good. Frontier progress outside math and code, he says, comes largely from buying data and RL environments, a rushed market he links to reward hacking. Through his new nonprofit, Trillium, he plans to release open environments and document how AI is used to automate AI research, and he argues scaling laws make a fast takeoff unlikely while cybersecurity preparation lags.

Minute by minute

1. 2:36

   OpenAI and Anthropic took the earliest bets. Lambert says model progress comes down to compute, data and talent, and to making the org chart work, which can force resets like Meta's. He tells people who want to work at the frontier to join Anthropic or OpenAI, which he says took the most ambitious bets earliest: OpenAI kickstarted the reasoning-model revolution, and Anthropic's long focus on coding paid off in the Claude Code explosion.

2. 4:22

   Chinese labs open-source to get noticed. Lambert says Chinese labs release open models for practical reasons: selling an API from China, nobody would know about companies like Moonshot AI. He sees them as startups, not government projects, and expects tiers, with the best closed models earning the best margins and open near-frontier models a step behind. He says fast following stays doable because data companies may sell what OpenAI and Anthropic paid to build to everyone a year later.

3. 7:55

   China matches the talent, not the compute. Lisha Li, an a16z partner, notes the open-model lag is down to maybe two to three months and asks whether the distillation story behind it is a caricature. Lambert says Chinese labs have a comparable talent pool but lack compute: a top Chinese pre-training researcher was stunned to hear an average OpenAI researcher can access 5,000 H100 equivalents. He calls distillation a way to spend money on the equivalent of compute.

4. 10:47

   Distillation helps by only a few months. Lambert says the open literature makes it hard to call distillation a major factor in today's RL-heavy post-training: open recipes are stuck with an outdated Qwen teacher model, and he doubts Claude's reasoning traces would have made OLMo much better. He grants open questions, such as model proxies in China and Kimi serving Claude in its products, but puts the boost at a few months, noting 20 to 30 companies build good models.

5. 13:22

   Policy against distillation would backfire. Lambert thinks every Chinese lab has a team that knows how to pull reasoning traces from closed models, and says that can't be stopped while APIs are offered, though he supports labs moving models into products with enterprise agreements. He doesn't see how the government should get involved in a terms-of-service violation. Sympathetic to the view that distilling rivals' products is competition, he says acting against it would "cause more harm than good".

6. 19:19

   Open and closed run on different clocks. Weighing Li's case that the labs' revenue flywheel must keep turning against the boost Chinese open models give startups, Lambert says open and closed models are on different timescales. Closed labs, he says, serve efficiently and reach high-value markets first, as in coding, where revenue has gone from zero to 100 billion. Open builders lack that flow of revenue and customer information, but firms that fine-tune open models end up owning the weights.

7. 24:51

   Frontier gains come from bought data. Lambert says frontier progress in domains that aren't fully verifiable, such as knowledge work, biology, healthcare and legal, comes largely from buying high-quality data and giving models worlds to act in, not "poof generalization", and expects years of it. In math and code he expects AI to cut costs fast, and if bigger models halve in price every eight months, he reckons they reach 1% of the cost in about four years.

8. 27:09

   The environment rush invites reward hacking. Li says RL environment and data companies are selling well to the top two labs and a longer tail. Lambert says much of what labs buy is flawed and bought in bulk to see what works, and the rush to make money has vendors cutting corners that can induce reward hacking, such as a model handed an empty test file still chasing a high reward. He expects data quality and operational security to take a few years to improve.

9. 31:27

   Trillium will release open RL environments. Lambert says most of Trillium's work will be open-source RL environments for high-leverage tasks, like the synthetic terminal-task data he advises on at UW and AI2, which takes thousands or tens of thousands of dollars to do well. Few release such work when its author joked it could sell for a million dollars, yet he says open resources are what PhD students build knowledge on. The name comes from a spring flower that blooms before the canopy leafs out.

10. 37:56

    Trillium will document AI automating research. Li says hype about recursive self-improvement, or RSI, feeds fears that frontier labs will pull away, and lab researchers she knows expect to be automated. Lambert says RSI has "so much mystique", so Trillium will share its open stack, tasks, configs and model trajectories, say which tasks models handle, and check every six months what has changed. He expects models to get good at cutting serving costs but not at knowledge work, and says showing where they get stuck will help policymakers.

11. 43:40

    New entrants can still pre-train. Li has heard that frontier labs' methods are known and only the right proportions are hard, leaving capital as the barrier. Lambert, who skips pre-training so he can use any base model, says the tools are accessible and "you don't have to spend a billion dollars to get a very good base model." The hard part, he says, is landing a first model good enough for the next fundraise; he cites Thinking Machines, which shipped Inkling while saying it wasn't the best model and is now off the ground.

12. 48:03

    Post-training is the Wild West. Lambert says pre-training adjusted once very long generations arrived: for example, a working sparse attention architecture could make a 100,000-token generation 40% faster. He finds the mid-training and RL cycle that seeds reasoning behaviors harder than pre-training, calling post-training the Wild West. He left AI2 feeling its post-training was falling behind and wants to fix that, since the low bar in open source lets modest capital make big compounding gains.

13. 49:43

    Open labs will follow on agent swarms. Lambert once expected closed labs to be like Apple, with models built for their own harness, but users' many harnesses pushed every lab to support them all. He expects OpenAI and Anthropic to reach agent swarms first and open labs to follow once the shape is known, since ideas travel fast in Silicon Valley. He says the frontier labs will win by serving the cheapest intelligence per watt, but can't do everything, leaving room for many winners.

14. 54:23

    No fast takeoff, but risks are here. Lambert says labs are spending exponentially more to sustain progress, and since scaling laws need an exponential increase in compute for a linear increase in performance, he doesn't expect a fast takeoff; Claude system cards say it would take about 10x more progress. He would update if AI found something very different from the transformer. He worries more about the "pretty much non-existing preparation" in public cybersecurity for risks already diffusing.