What's this about?
Mollick highlights a structural tension: the US is building an approval/oversight regime for closed frontier models, while open-weights models — which can't be recalled once published — face no equivalent. With Mythos-class open models now plausible from Chinese labs, he revives his April warning that CISO offices have at most months before frontier-grade offensive capabilities diffuse to bad actors.
What exactly is a "Mythos-class" model?
Mythos is Anthropic's top model tier — the version of Fable 5 without the extra dual-use safety measures, available only to approved organizations because of its raw capability in sensitive domains (like cybersecurity). "Mythos-class" has become shorthand for "models capable enough to need that kind of gating." The regulatory puzzle: gates work on closed APIs, but once an equally-capable model's weights are on the internet, there is no gate — anyone can download and fine-tune it.
What's the community saying?
- The asymmetry is unsustainable — regulate closed models hard while open ones are free, and either regulation is pointless or open releases eventually get restricted; something has to give.
- The clock argument — his April prediction gave 6–9 months before red-team-report capabilities diffused; he's now saying "my timeline was right," pointing at Chinese open releases.
- The quiet question — will the Chinese government keep blessing open releases of frontier-grade models if the US treats them as a security problem? (Xi's Shanghai open-source commitment says yes, for now.)