Tan argues that American open-weight labs like Meta and Allen Institute are unilaterally disarming by respecting ToS clauses that Chinese labs like DeepSeek and Qwen openly ignore. He frames this as a national competitiveness issue: every month US labs comply, the default open-weight model on Hugging Face becomes more Chinese-dominated.
The editorial notes that OpenAI, Anthropic, and Google all prohibit using outputs to train competing models, but these clauses have never been tested in court against foreign labs and become unenforceable once distilled open-weight models are released. DeepSeek-V3 and Qwen 2.5 show obvious behavioral fingerprints of GPT-4 training data — including telltale 'As an AI language model developed by OpenAI' slips — yet face no meaningful consequences.
Training a frontier model from scratch requires north of $100M in compute, while distilling a competitive student model costs one to two orders of magnitude less. DeepSeek's own V3 paper openly credits synthetic data pipelines, demonstrating that distillation isn't a marginal optimization but the primary path to competitive open-weight models.
At a Y Combinator event this week, Garry Tan told a room of founders what most of them already suspected: American open-weight AI labs are losing the distillation race because they're the only ones still pretending the terms of service matter. His pitch, per TechCrunch's coverage, is that labs like Meta, Allen Institute, and the newer open-weight players should openly train student models on outputs from GPT-5, Claude, and Gemini — the same technique DeepSeek and Qwen have been using for two years to close the gap on frontier capability at a fraction of the compute cost.
The legal fig leaf is thin and everyone knows it. OpenAI's usage policies prohibit using output "to develop models that compete with OpenAI." Anthropic's are similar. Google's Gemini terms carry the same clause. None of these terms have ever been tested in court against a foreign lab, and none of them apply once a distilled open-weight model is in the wild. DeepSeek-V3 and Qwen 2.5 both show clear behavioral fingerprints of GPT-4-family training data in their outputs — the occasional "As an AI language model developed by OpenAI" slip, the identical refusal phrasings, the same math-problem solution templates. The labs have been asked about it. They shrug.
Tan's argument is that the asymmetry is now a national-competitiveness problem, not a licensing problem. Every month American open-weight labs spend respecting a clause that only binds Americans is a month the default open model on Hugging Face gets more Chinese. As of this week, seven of the top ten trending models on HF by download volume ship from Chinese labs or Chinese-lab derivatives.
Distillation isn't a shortcut — it's the shortcut. Training a frontier-class model from scratch costs somewhere north of $100M in compute alone; distilling a competitive student model from a frontier teacher costs one to two orders of magnitude less. DeepSeek's own paper on V3 openly credits synthetic data pipelines seeded by "strong reasoning models," without naming which ones. Qwen's technical reports are slightly more coy, but the ablations show the same story: student models trained on frontier-quality synthetic traces converge faster and land at higher benchmark scores than models trained on organic web data alone.
The reason American labs haven't openly copied this playbook isn't technical — it's that they're downstream of the frontier labs' lawyers, VCs, and enterprise sales orgs, all of whom would rather not test the ToS. Meta trains Llama on its own scraped corpus. AI2 uses Dolma. Neither ships a distilled-from-Claude model, even though the resulting weights would almost certainly be stronger per parameter. This is the moat Tan is calling out: it's a moat only Americans are honoring.
The geopolitical framing is where his argument gets uncomfortable. If you buy the premise that open-weight models are strategic infrastructure — the way TCP/IP or Linux became strategic — then the country whose labs define the default weights defines the default assumptions baked into every downstream fine-tune. Tokenizer choices, RLHF value alignment, refusal patterns, factual defaults on contested topics. DeepSeek-R1 refuses to discuss Tiananmen. Qwen has specific behaviors around Taiwan sovereignty questions. These aren't bugs — they're the RLHF layer doing what RLHF layers do, expressing the values of the people who trained them. If your on-prem RAG stack in 2027 is running a Qwen 3.5 derivative because it was the strongest open model when you froze your deps, you've quietly imported those defaults.
The counterargument, which Tan gestures at but doesn't fully engage, is that distillation-based open-weight labs would be structurally dependent on frontier labs continuing to serve API traffic to entities that are provably distilling them. OpenAI has already banned specific accounts caught doing large-scale distillation runs. The obvious response is that once you have the distilled weights, you don't need continued API access — you have your student model, and you can iterate from there with synthetic self-play. DeepSeek has demonstrated this is a viable path.
The deeper question is whether "open-weight" as a category survives if the frontier labs decide the ToS enforcement gap is intolerable and start rate-limiting or geofencing aggressively. That's the endgame Tan is implicitly betting against — that the frontier labs will keep the taps open because the alternative (a closed API economy where distillation is provably impossible) hurts their own revenue more than it helps their moat. He's probably right, but it's not a certainty.
If you're picking an open-weight base model for a production fine-tune today, the honest read is that the Chinese labs have a 6-12 month lead on capability-per-parameter and it's widening. Qwen 2.5-72B beats Llama 3.1-70B on most reasoning benchmarks. DeepSeek-V3 is competitive with GPT-4-class models at a fraction of the inference cost. If your compliance team will let you ship a Chinese-origin base model, you're probably already using one. If they won't, you're paying a capability tax to stay on Llama or Mistral.
The practical implication of Tan's pitch, if American open-weight labs actually take it, is that the capability gap on Hugging Face closes in 2026-2027. You'd see a Llama 5 or an AI2 Tulu 4 that was openly distilled from Claude or GPT-5 outputs, ships under a permissive license, and matches or beats the Chinese baseline. That's the world Tan wants. It requires American labs to accept some ToS-violation risk in exchange for capability parity — a trade the current cohort of legal-and-comms teams has been unwilling to make.
For infra teams: assume the local-inference model landscape is going to fragment along geopolitical lines faster than most vendor roadmaps admit. Any 3-year contract that assumes "open weights" is a stable category should have exit ramps. Fine-tuning pipelines should be portable across base models — the base model is going to change more often than your training data does.
Tan's pitch will get louder before it changes anything. The first American lab to openly announce a distillation-based open-weight release — with the ToS violation as an explicit strategic choice rather than a plausibly-deniable accident — will take heat from the frontier labs, cover from the national-security crowd, and market share from the Chinese incumbents. Whoever moves first probably wins the category. If nobody moves, the default local model in your 2027 stack ships from Hangzhou.
I agree. The frontier models are based on training data from tons of copyrighted work. Some of that work was obtained illegally, even. They could not exist without strip-mining the commons. The labs have no moral or ethical ownership to the end result, and others should feel free to treat any compan
> He also notes that the proprietary AI labs didn’t ask permission when they vacuumed up as much human knowledge as they could to train their models.I think this should desactivate the moral high ground from which Anthropic is trying to speak. That they would want to make distillation orderly IMH
I think OpenAI and Anthropic will go bust, or at least be scrapped for parts in the next 5 years or so. It's clear that the extreme cost used up for training is impossible to recoup, as inference is already being subsidized.It's also clear that, as Tan indicates, open-weight models will be
Controlling what users and customers do with API calls to closed weight models feels constraining, and there’s a role government can play here to normalize the fact that access to intelligence that was trained on broad public access data should itself also be more a form of a public good than someth
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
https://archive.ph/BnceE