'Six months to live for open models': a warning read closely

The claim

Lambert's essay is a forecast with a deadline. He argues that within roughly six months, US policy will move against open-weight models at the frontier, and that the movement is coming from two directions at once. One is a set of discussions about restricting open models above a capability threshold. The other is the debate about distillation, meaning the practice of training a model on the outputs of a stronger one, which has become a proxy for the fight over Chinese-origin models.

He grounds the timeline in specifics. There are discussions at the White House of a new executive order to manage open models, with Chinese-origin models and government use as the targets. There was a June 9 meeting where representatives of frontier companies, including Reflection AI, discussed carving out exemptions for open models based on capability level. And the threshold he expects a rule to land on would ban open-weight releases beyond the capability range of GPT 5.5, Claude Opus 4.8 or GLM-5.2. His summary is that open models are staring down the barrel of policy action that could make them a permanent second-class citizen.

The distillation argument and who benefits from it

The sharpest part of the essay is the section on distillation, and it is sharp because it names an incentive rather than a principle. Lambert characterises the campaign against Chinese models that train on the outputs of US labs as regulatory capture. A ban on those competitors would give the labs pushing for it substantial economic security, and it would do so without the labs having to solve the problem on their own side, which is that their APIs leak. He cites the episode in which people organised on Discord obtained unauthorised access to Anthropic's Mythos model while it was in a limited beta, and draws the obvious conclusion. If the technology is as powerful as its makers say, they should be able to secure their API.

We think this argument is right on the incentive and incomplete on the remedy. It is true that a company asking the government to prohibit a competitor's training method is asking for a subsidy. It is also true that trillion-parameter models are now being distilled in ways nobody has a clean way to detect, so a rule against distillation would be enforced against whoever is easiest to point at rather than whoever did it. What the essay does not do is separate the question of whether distillation from a closed API violates a contract, which is a matter between two companies, from whether the state should prevent open models from existing, which is the thing that would hurt everyone else.

The proposed fix

Lambert's remedy has three parts. US companies with the resources to do it, and he names Microsoft and Meta, should release competitive open models, so that the story in Washington stops being that only China builds them. The open-source community, which he describes as diffuse, needs to organise into a coalition that can actually show up when rules are being drafted. And the whole thing has to be coordinated internationally, because a unilateral US ban does not remove the models from the world, it only removes American participation in them, while people outside the rules keep access. His phrase for the alternative is speedrunning dystopia, with a US tech industry that ends up looking far more like a Chinese system with control.

The first part is the load-bearing one and the least in Lambert's control. Whether a US lab ships a frontier-adjacent open model in the next six months depends on decisions already being made inside those companies, and the essay is in effect a public plea to the people making them.

Reading it from a lab that needs open weights

We are a non-profit research foundation. Almost everything we publish requires weights we can load, activations we can read and training data we can at least partially inspect. Interpretability on a closed API is not interpretability, and an evaluation of a model we cannot run at fixed precision and fixed decoding settings is an evaluation of someone else's serving stack. So a rule that stops frontier open weights at a capability threshold does not slow our work down, it caps it, and the cap moves further from the frontier every time the closed models improve.

That gives us an interest, and we want to be honest that it colours how we read the essay. Lambert's forecast could be wrong in either direction. The White House discussion might produce nothing, or it might produce something narrower than a capability ban, such as procurement rules for government use that leave researchers alone. Or the threshold could land lower than the models he names. We do not have independent information about the meetings he describes, and neither will most readers, so the honest position is that the essay is a well-sourced warning from someone close to the process rather than a confirmed plan.

What we can say from our own experience is that the second-class citizen outcome does not require a ban. It only requires the frontier open models to be a generation behind and to stay there, because then every result on an open model comes with the caveat that it may not transfer to the systems people actually use. That gap already exists. A rule that fixes it in place would be worse, but the essay's six-month deadline is a deadline for the policy, not for the underlying problem, which has no deadline at all.

What we would do with the next six months

If the coalition Lambert asks for is going to exist, its most useful product would be evidence rather than advocacy. Nobody in the policy discussion he describes has a document that shows what research was done on open weights in the last two years that could not have been done on an API, with the results and the models named. We can write that document for our own work and we expect other groups can too. A rule drafted to prevent capability diffusion should have to weigh that against something concrete.

The other thing we would want is a test of the essay's own premise. If a US lab does ship a competitive open model before the deadline, we will learn whether that changes the policy conversation the way Lambert expects. If nobody does, we will learn how much of the open-model ecosystem's political position depended on companies that were never going to defend it. Either result is worth having, and both arrive within the window the essay sets.