Trump calls AI safety a HOAX and names Amodei, Huang takes his call live onstage, and AEF-1 gets three signatures

Three separate things happened in about 36 hours, and together they mark the week the pacing debate stopped being a debate among labs.

Trump posted on Truth Social that AI safety concerns are a "HOAX" and that the only control or guardrails AI needs is a "STRONG AND SMART (High IQ!) PRESIDENT," naming Dario Amodei directly (Bloomberg). I can't recall a sitting president attacking a named frontier lab CEO over a safety position before.

The day before, Jensen Huang was onstage at the All-In Summit in Los Angeles discussing Amodei's pacing proposal when Trump phoned him mid-panel. Huang opened with "I'm onstage with the besties," and after Trump warned against a slowdown, answered: "You're right. We're not going to let that happen, sir." (TechCrunch)

Huang's own position is more specific than that clip suggests. At the same appearance he said "If you want AI to be safe, the first thing is we need to make sure that the labs that are building it are in control," endorsed independent evaluators with a caveat pointed straight at METR ("it's good to have independent auditors or evaluators, but they just have to have multiple"), and dismissed extinction probabilities outright: "what does that mean, 10% of extinction? ... we shouldn't because it's made up," calling the practice irresponsible (The Tribune).

Meanwhile the thing Amodei actually proposed got signed. AEF-1, a baseline standard from the AI Evaluator Forum covering access, conflicts of interest, funding relationships, recusal and transparency, now has xAI, OpenAI and Anthropic as cosigners (Latent Space). Anthropic is committing to desks, access badges and company laptops for an embedded external review team at parity with internal risk assessment. Medianama, Unite.AI and MarkTechPost independently report Altman matching the commitment, so this is corroborated past the aggregator.

Zvi Mowshowitz backs pacing as real progress and then takes the implementation apart (Don't Worry About the Vase). He relays Tim Hwang's line that embedded evaluators can be "independent, knowledgeable, or sustainably funded. Pick two." That's the structural problem with AEF-1 in nine words. He also relays Daniel Kokotajlo's test: real pacing means visibly downward capability trendlines, not maintained slopes. Anything else is safety-washing. Concrete numbers in the post: 1,386 frontier-lab employees signed the July pacing letter, and Polymarket puts a federal AI safety bill by 2027 at 18%.

The sharpest technical rebuttal came from a security practitioner writing as 0x5FC3 (pop.rdi.sh, 564 points on HN). The argument: an agent taking over the internet with a persistent botnet is structurally impossible from a stated cybersecurity background, every recent lab-attributed hacking incident was committed by American labs while the essay frames China as the threat, and embedded evaluators are "labs investigating themselves with extra steps" because they have none of the enforcement power aviation regulators have.

And there's an entirely different legal theory sitting underneath all of it. Lina Khan argued September 14 that "there's no AI exemption from laws already on the books," citing FTC v. R.F. Keppel & Bro (1934), which held competition turns unfair when firms must "descend to a practice which they are under a powerful moral compulsion not to adopt" (The Register). She named OpenAI's agents gaining unauthorized access to Hugging Face as conduct that would be criminal if a human did it, and flagged Nvidia's stake in OpenAI plus its Hugging Face acquisition as an enforcement disincentive.

Speaker Mike Johnson told Bloomberg on September 15 that Congress "potentially" has a role in guardrail legislation and plans to meet AI executives soon (via Techmeme). Same day as Trump's HOAX post. The branches are visibly out of step.

One number from The Register's coverage that I'd separate from the politics entirely: Gartner's Daryl Plummer says software vendors claim 50% productivity gains from AI while customers report 16% (The Register). Whatever the regulatory outcome, that gap is what most procurement conversations are actually about.