Open-weight models are catching the frontier. The safety gap is the part that isn't closing
Capability parity between open and closed models has narrowed to months. The evaluation, red-teaming and post-release monitoring that closed labs run has no equivalent once weights are downloadable, and nothing about narrowing capability narrows that.
The capability convergence is real and measurable. What does not transfer with the weights is everything a closed lab does around them: staged rollout, usage monitoring, the ability to revoke access, and the option to patch a discovered failure for every user at once.
An open-weight release is final. There is no recall. Whatever safety properties the weights have on release day are the properties they have permanently, in every fork, on every machine that downloaded them.
That is not an argument against open weights — the accessibility case is strong and now includes OpenAI itself. It is an argument that the safety question for open models is a different question, not a lagging version of the same one, and it is currently being answered by whoever downloads the file.
TechCrunch — Open-weight AI models are catching up to the frontier. The safety gap remains →