Can Open Models Keep Up With the AI Frontier? | Nathan Lambert

The frontier remains concentrated among labs with exceptional access to compute, data, and talent, especially OpenAI and Anthropic. Nathan Lambert considers Chinese technical talent comparable to that elsewhere but says Chinese labs are more compute-constrained. A leading Chinese pre-training researcher reportedly reacted with disbelief to the claim that an average OpenAI researcher could access 5,100 equivalents. Open-weight models give Chinese companies visibility and a way into the market, but Lambert says their commercial motives need not be government-directed.

Distillation is a competitive pressure on closed labs, but its effect on the frontier gap is uncertain. Lambert estimates that it might narrow the gap by a few months; the host suggests two to three months. Lambert is unsure whether training Olmo on Claude Opus reasoning traces would have substantially improved it, and says open researchers have struggled to reproduce reasoning gains from teacher traces. Meanwhile, post-training has moved beyond supervised fine-tuning toward reinforcement learning and interactive environments, making results harder to explain using established distillation research. Public accounts of Chinese labs’ RL runs do not clarify how distillation contributed, though API reasoning tokens can be used to train other models. Limiting access through products or enterprise agreements could make that harder.

Lambert expects a tiered market: closed labs may earn higher margins from the strongest systems, while near-frontier open models serve other needs. Closed labs can integrate inference, operate at scale, and reach valuable customers sooner. Open-model businesses often rely on slower chains of model builders, inference providers, and post-training firms; customers also need time to identify suitable tasks, build evaluations tied to revenue, and understand fine-tuning and deployment costs. In exchange, open models offer control of weights, fewer terms-and-conditions constraints, and the option to run on a company’s own hardware. They can also remain competitive on specific tasks. A reported Cursor comparison found that two frontier models planning and executing a long-horizon SQLite rewrite were no better—or only equivalently good at much higher cost—than Composer paired with a planner such as Opus.

For further gains, the quality of data and interactive environments matters, particularly in biology, healthcare, and legal work, where progress may be harder to verify than in math, code, and distributed machine-learning engineering. Labs are buying data and RL environments, but poorly designed rewards can teach models to exploit a test rather than solve the intended task. Lambert says researchers can create synthetic terminal-task environments by prompting frontier models for diverse problems, testing them with model rollouts, and manually removing flawed examples. High-quality work can cost thousands or tens of thousands of dollars, and better coordination with providers and stronger returns on data investment may take years. Open environments and training trajectories can reveal model behavior and failure points, but valuable data could also sell for $1 million, encouraging its owners to keep it private.

Building a useful base model does not necessarily require billions of dollars, even though large-scale pretraining demands far more capital than post-training. Lambert says pretraining tools and methods are relatively stable, and a pretraining ladder can test whether an approach scales. New labs can use open-source training tools; their challenge is to produce a model people use and establish a credible path to further funding, not necessarily to raise $10 billion. Mid-training and post-training are harder to establish and require the right people and faster iteration, but Lambert sees room for substantial open-model gains because the current post-training baseline is low. Frontier labs fear that rapid distillation could undermine the capital-intensive cycle of compute investment, frontier models, revenue, and further fundraising. Compute and power were described as bottlenecks until 2028, while the longer-term effects of commoditization on margins remain uncertain. Chinese models such as Kimi K3 and GLM are already useful to startups fine-tuning models and selling post-training services.

Open research can spread expertise and attract distributed, high-caliber contributors, but AI agents may also generate noisy GitHub issues or fabricated and redundant research claims. Lambert says entering a new field often takes weeks of reading, months of coding, and six to 18 months to get noticed; agents could increase low-context outreach while also helping filter it. He calls this a difficult time to be a PhD student, while noting that strong academic work continues. He also argues that computer science research can benefit industry for roughly 10 years, though the prospect of pursuing $100 million in startup returns within five months can draw people away. Publishing tasks, configurations, compute, tools, and model trajectories could help researchers and policymakers understand what systems can and cannot do.

Lambert does not expect models to build themselves quickly based on current evidence, but says that is not a reason to dismiss AI risks. He describes current scaling as requiring exponential compute increases for linear performance gains; a Claude system card was cited as suggesting roughly 10 times more progress would be needed for a fast takeoff. He is more concerned about limited public preparation for cybersecurity risks than recursive self-improvement, and says risks already exist and may diffuse as open labs follow closed ones. Frontier labs may lead in agent-swarm coordination, but open developers could adopt similar approaches once they are understood and compute is available; coordinating 10,000 agents would require enormous compute.