The Once And Future Fable #5

Abstraction: AI-policy roundup on ad hoc licensing regime and open weights

Key points:

• Mythos 5 restored to 100+ US institutions June 26; Fable and GPT-5.6-Sol still withheld from the public while an ad hoc White House licensing regime operates (CAISI under a stop-work order, OSTP underfunded).
• Slaughter/Trump v. Slaughter (6-3) overruled Humphrey's Executor: President can fire principal officers at will (except the Fed), making an independent nonpartisan Frontier AI Commission much harder; Dean Ball favors private-governance/auditor models regulating the lab not the model.
• "Code is speech" / model-as-speech First Amendment claims (Preston Byrne): Zvi predicts they won't hold, and warns that if courts blocked all use-regulation, government would instead block training or access to sufficiently capable models.
• Anthropic accused Alibaba of massive distillation via ~25,000 fraudulent accounts; Zvi distinguishes this direct fraud/ToS violation from training on already-public outputs; notes Google uses intentional silent output degradation against distillation.
• Open-weight models are "unsafe and nothing can fix this" — the lack of safety is the point; restrictions are cheap to strip (e.g. Obliteratus); Dario warns open source is on a "very dangerous path."
• "Banning open source" is only coherent as the existing de facto ban on serving/open-sourcing a truly frontier model without White House permission.

Connections: Zvi Mowshowitz · Anthropic · Deepmind · Alibaba · AI Policy · Open Weights Models · AI Governance

Source: https://thezvi.substack.com/p/the-once-and-future-fable-5