Amodei wants evaluators inside the labs and the AI czar calls it an op

On Saturday September 12 Dario Amodei published an essay asking the AI industry to slow down, then sat with Anderson Cooper and said he agrees with the researcher who had just quit Anthropic and told the country the technology could kill everyone. Within hours the proposal had public backing from Sam Altman and Elon Musk. It had also, a day earlier, been called an operation aimed at destroying open source software by the president's own AI adviser. I read all three conversations in full. The disagreement is not really about how dangerous the models are. It is that every safety mechanism proposed last week happens to strengthen whoever proposed it.

Amodei did not distance himself from the doomer

Jacob Coxon resigned from Anthropic on September 9 after three years of pre-training work split between OpenAI and Anthropic, writing that neither company is acting responsibly. The obvious move for a CEO is to call that overheated and move on. On CNN Amodei said he agrees with Coxon much more than he disagrees, and then defended him.

"He wasn't calling out us. He was calling out the dynamic of the industry as a whole moving too fast."

Dario Amodei, September 12, 2026, CNN interview

The proposal, We Must Pace the Frontier, has three steps that get harder as they go. Anthropic commits first and alone to letting an outside review team work inside the company with desks, badges, laptops and permissions like an internal risk team, and to let that team publish what it finds without Anthropic holding editorial control. Then labs in democracies agree on common standards. Then somebody tries the same conversation with China. Amodei reached for banking regulation to explain step one. Supervisors sit in the building every day and check practices rather than read filings afterward.

The operative ask is an antitrust waiver

Steps one and two are voluntary. Only one thing in the essay needs the government to act, and Amodei said it plainly. Companies cannot legally agree among themselves on how fast to ship product.

"it can sit down in the room with all the industry players and we can all talk together and the government can make sure that we're not doing anything nefarious from the perspective of antitrust and collusion."

Dario Amodei, September 12, 2026, CNN interview

That is the whole game and it is getting almost no coverage. A narrow waiver is how release timing becomes something the largest model providers may lawfully discuss with each other. Amodei is not hiding it. The essay says the government need not participate, only issue the waiver. Whether that reads as prudent or as a cartel with a chaperone depends on how much you trust the room.

Sacks says the proposal is the point

David Sacks spoke at the Open Source AI Summit in the Presidio on September 11, the day before the essay went up. Asked to read the week politically, he never engaged with the risk argument. He went straight to who benefits.

"I think that what's happening right now is part of a larger campaign whose target is ultimately open source AI. It's not the only target. But you should just understand that open-source AI is the real target here."

David Sacks, September 11, 2026, Open Source AI Summit

His argument is a three move sequence he says he can already see coming. Regulators promise nobody wants to ban open weights, they only want them safe. Then the requirements apply equally to open and closed models, because equal treatment sounds fair. Then some of those requirements turn out to be central monitoring and the ability to roll a model back, which a downloaded weights file physically cannot do. He is right about the mechanics, and Amodei has said the same thing under oath.

Then he went further, and here I would ask for evidence.

"how did the Wall Street Journal post your news about your tweet storm 18 minutes before you posted it? So obviously you sat down with the Wall Street Journal"

David Sacks, September 11, 2026, Open Source AI Summit

A coordinated press rollout and a manufactured belief are different claims, and he slides between them. He also inflates his own side, telling the room that the pro open weights letter led by Nvidia and Microsoft had "a couple hundred signatories." It had 25, with OpenAI, Anthropic and Google absent. One detail does support him, though. On September 11 he used the phrase "pacing the frontier" in air quotes, a full day before the essay carrying that title appeared.

Gerstner attacked the messenger, then endorsed the message

Brad Gerstner is an investor in both Anthropic and OpenAI, disclosed in the first thirty seconds of a Halftime Report appearance on September 11 built entirely around his rebuttal.

"what I didn't like this week is these hyperbolic scare tactics, which I think are hiding behind a political agenda"

Brad Gerstner, September 11, 2026, CNBC Halftime Report

His inventory of controls that already exist is the most useful thing anyone said all week. He cited the government forcing Anthropic to shelve its most capable model, the pre release review framework, and the company's own 154 page misuse report from September 10, which documents Houthi missile work and an Alibaba distillation campaign that peaked near three million exchanges a day. One correction. He credited Treasury and Commerce jointly, but the takedown was a Commerce order on June 12 barring access for any non US national anywhere, which Anthropic could not apply selectively. Then he asked for something stranger.

"I want a little investigation: who is behind all of this negativity? Is it coordinated? Where is the money coming from?"

Brad Gerstner, September 11, 2026, CNBC Halftime Report

Someone arguing that regulators should stay out of his market, asking the government to investigate who funds his critics, is not making an argument. The tell came a day later. Once the actual text was public, the BBC reported that Gerstner called Amodei's proposal an important step forward. He was never fighting the substance. He was fighting a nightly news segment that reached his own sister.

Anthropic already ran this experiment on itself

The best evidence on what the word safety carries here came from a podcast that never mentions Coxon. Ben Thompson taped a Sharp Tech episode around September 2 that YouTube did not publish until September 11, and it is about the other Anthropic story. In June the company told enterprise customers it would retain their traffic for 30 days on its frontier models, justified as a defense against novel cyberattacks. Customers refused. On September 1 Anthropic backed down, promised a replacement called Enterprise Frontier Safeguards for sometime this fall, and restored zero data retention in the meantime.

"And the fact that they will make it zero data retention until this new program is available is the tell. This was not about safety. If it was about safety, you would not do it."

Ben Thompson, recorded early September, published September 11, 2026, Sharp Tech

Read that beside the pacing essay. A safety rationale that evaporated the moment enterprise buyers threatened to walk does not prove the extinction argument is insincere. It proves that safety claims here survive exactly as long as they are affordable. Thompson draws the conclusion I keep landing on.

"A market check. And it is the only check that scales to the size and power of these companies."

Ben Thompson, published September 11, 2026, Sharp Tech

One person described being scared by a product instead of a tweet

Everything above is people arguing about a media cycle. Dave Morin opened his September 11 episode with the same conspiracy framing as Sacks, then dropped it. He had been using OpenAI's Astra. His twelve year old son built a 3D model in an hour and put it on the printer.

"Like this thing is operating in three-dimensional space at a speed and with a 12-year-old that if you were a really bad person, you could figure out ways to use this that are legitimately scary."

Dave Morin, September 11, 2026, Dave, Brit and Jess podcast

He also named the gap in Gerstner's list. The people running the labs keep telling him that what they use internally is alarming, and that stopping is not available to them.

"We can't slow down because we have to keep creating the returns for the investors."

Dave Morin, September 11, 2026, Dave, Brit and Jess podcast

So does any of this change what ships

Federally, no, and not soon. Sacks expects the White House to stay on the light touch side for at least two years and says the fight has moved to more than 1,500 bills in state legislatures, where Anthropic and OpenAI have lobbying operations and the open weight community has none. Nothing in the pacing essay binds anyone, including Anthropic.

Two things would tell you the argument became real. First, the antitrust waiver. If Commerce or the Justice Department issues a narrow safety carve out, competitors get a lawful room in which to discuss release timing, and that reshapes the market more than any state bill will. Second, whether OpenAI or Google matches the embedded evaluator commitment with an outside team that can publish without a veto. Until one of those lands, the honest read is that four powerful people spent a week describing the safety regime that would suit them best, and the only thing that actually forced Anthropic to reverse a decision this year was a group of enterprise customers refusing to hand over their data.