White House Will Ad Hoc Decide Who Can Individually Access GPT-5.6

ai-policyopenaiwhite-housemodel-deploymentexport-controlsai-governance

Abstraction: Critique of ad hoc White House frontier-model access regime

Key points:

New de facto policy: White House approves GPT-5.6 access "customer by customer" in an opaque ad hoc way — Samuel Hammond: "zero AI regulation to CFIUS-but-for-API-access in about a week"; Altman calls it not their "preferred long term model."
Zvi calls it "maximally Not The Way" but notes it at least isn't Anthropic-specific persecution, and isn't the old do-nothing/see-no-evil posture.
Andrew Curran: this slows releases not internal training, so the public-vs-internal capability gap widens; US is ~9 months ahead of China, staggered releases could close that; raises odds of Chinese-model bans and Nvidia export tightening.
On Fable: Zvi reads Polymarket as implying most of the ~50% July return chance is a KYC implementation (~12%) rather than restrictions lifting; Dario replaced by Tom Brown in negotiations, undermining "must only talk to Dario."
Part 2 (blame game): anti-regulation figures (Andreessen, Sacks) who insisted models would commoditize and no preparation was needed now blame safety advocates; Zvi's Covid-19 parable — warning about a problem doesn't make you responsible when unprepared government responds badly.
Defends earlier qualified support for the June EO: the window for "hard-to-abuse" regulation (SB 1047, Biden EO) closed; the EO formalizes rather than causes the ad hoc regime.

Connections: Zvi Mowshowitz · Openai · White House · Anthropic · AI Policy · AI Governance · Model Deployment