What's this about?
Mollick raised a policy question K3 makes urgent: frontier proprietary models go through government-adjacent safety vetting (US/UK AI safety institutes) before release — but how does that work for open weights? K3 has no model card yet, open models are easy to jailbreak, and once weights are published they can't be recalled. He notes governments still have levers (conditions on companies doing business with them), but concedes policy here is "emergent" — nobody knows.
What exactly is "pre-clearance"?
Before releasing top models, major labs let external safety institutes test for dangerous capabilities (bioweapons help, advanced cyber offense). That works because API access can be gated or revoked. Open weights break the model: anyone can download, strip the safety training, and redistribute. The question is whether open releases claiming frontier capability will face equivalent vetting — and whether China's labs would ever participate.
Why does it matter?
K3 isn't claimed to be at Mythos/Sol level, but Mollick's point is someone's open model will be soon. The governance gap between "vetted API models" and "unvetted downloadable weights" is widening exactly as open models close the capability gap.