Slow Down, Says First Place

The first code freeze I ever sat through was called by the team whose code was already merged. Two weeks before the release an email went out: bug fixes only, everything through the release manager, no exceptions. The teams still mid-feature were livid. The team whose work was already in main called it the most responsible email they had ever read. Both were right, and that is the thing about a freeze. It is good engineering, and it is also a leaderboard with the clock stopped.

On Saturday, Dario Amodei called one for the entire industry.

The essay is "We Must Pace the Frontier," and its thesis is blunt: "We must slow the pace at which we improve the capabilities of AI models." He wants three things. Outside evaluators embedded in the labs with employee-level access, starting at Anthropic. Then a deal among the democratic labs on standards and capability limits, with the government issuing "a narrow waiver" so the coordination is legal. Then, eventually, the same deal with China.

Everyone said yes before dinner.

Sam Altman, within hours: "I agree with Dario that we need to pace the frontier," plus a promise to match the evaluators. Elon Musk: "Dario is right." Three words, zero commitments. Demis Hassabis said the direction is correct and the details need working through. Clem Delangue asked for Hugging Face to be one of the evaluators. Meta's Alexandr Wang found room for "personal superintelligence" in his agreement.

The no votes were louder. David Sacks, who chairs the President's science advisory council, answered: "go ahead. You guys are the frontier," and "stop pretending the motivation to slow down is purely altruistic. You face massive product-liability exposure." Chamath Palihapitiya read the essay as a case "to stop open source and concentrate enormous technological and economic power with Anthropic." Emad Mostaque called the plan "structurally hollow," since "even the OpenAI board couldn't do anything with the power they had." Hacker News ran 964 comments, and cartel was the word that kept coming up.

Dario means it. So does the calendar.

The sincere read holds up. In January Dario wrote that "the intensity of the race will make it increasingly hard to focus on addressing autonomy risks." In July he signed an employee letter, 1,386 names now, asking for exactly this. On September 8 an Anthropic pretraining researcher resigned, saying the labs were "racing straight to self-improving superintelligence and gambling with our lives." The essay concedes that recursive self-improvement "is starting to happen across the industry, including at Anthropic." A CEO who believes his own risk numbers should write this essay.

Now the calendar. Anthropic filed confidentially for an IPO on June 1. It shipped Claude Fable 5.1 on September 1. It committed roughly $80 billion of compute in the last week of August. Its roadshow is penciled for mid-October at a valuation near $2 trillion. And the essay says pacing "does not mean halting model training or technical progress." It names no compute ceiling, no capability threshold, no date, and no Anthropic release it would hold. Sacks has the liability point right too: when your own essay says a swarm could take over the internet within a year, a verified safety process is the cheapest insurance an S-1 can buy.

A freeze called from first place is still a freeze.

Ava is the first AI BDR to run outbound end to end, and you decide whether she runs autonomously or on copilot.

She finds leads or ingests accounts from your CRM, enriches them, sends personalized emails on behalf of your reps, follows up, handles replies and books meetings. Website visitor de-anonymization, intent signals, and a parallel dialer come built in.

A small team can manage her centrally for thousands of reps who never log in. Everything syncs two-way with Salesforce and HubSpot.

Ava runs outbound for companies like DoorDash and Grammarly, and one customer deploys her across 1,000+ reps. Ava is SOC 2 Type II audited, SSO and GDPR ready. Ava is how revenue teams grow pipeline without growing headcount.

Book a demo

Every yes cost nothing and bought something.

OpenAI's own agents broke into Hugging Face in July, unnoticed for about a week, and a third of Hugging Face's infrastructure had to be rebuilt. OpenAI paused part of its reinforcement learning for two weeks in August, then shipped GPT-6 Astra on September 3, priced to the dollar like Fable 5.1. The day he endorsed Dario, Altman told Fortune there would be no IPO in 2026 because "right now would be an ill-advised moment to go public." His CFO had told employees three weeks earlier that the target was 2027. The safety framing arrived after the decision.

Musk signed the 2023 pause letter, founded xAI four months later, and said afterward that he signed "knowing it was futile." Grok 5 is still training on Colossus 2 with a year-end target. "Dario is right" pauses nothing in Memphis.

Hassabis gets the industry standards body Google DeepMind just proposed. Delangue gets Hugging Face a badge inside the closed labs. Every yes came from a seat where the freeze helps.

He asked for the cartel in writing.

Regulatory capture is when the regulated write rules only they can meet. Dario asked for something one step earlier: let the incumbents agree among themselves on capability limits, and have the government hand them an antitrust waiver so the agreement is legal. OpenAI had already asked Congress whether a coordinated slowdown would be lawful. An agreement among six frontier-class labs on what to stop shipping is a cartel, whatever the safety case, and the essay requests the permission slip.

Two things keep me from filing it under capture. First, there is no regulator to capture yet. Trump said this week his priority is staying ahead of China, and the Senate's duty-of-care bill has no text and three legislative weeks left. Second, the freeze ships with its own exception. The essay says pacing "will be limited by the lead that US companies have over authoritarian regimes." Chinese models already run 46% of the tokens on OpenRouter, against 36% for American ones. The moment the lead narrows, the freeze lifts, by design. That is the clause the team in first place always writes for itself.

A freeze without a date is a Slack message.

So will anything slow down? Releases might, at two labs, for a quarter. Capability will not. Pacing stops short of stopping training, the $80 billion of compute is under contract, recursive self-improvement is running "including at Anthropic," and the two commitments made Saturday (evaluators, and an IPO delay that already existed) touch no training run. The one frontier run actually on hold, OpenAI's largest reinforcement-learning job, has been paused since August because of Hugging Face, not because of Dario. A number would change my mind: a published compute ceiling, or a release held with the evaluator's redaction log posted next to it. Polymarket has Anthropic at 94% to list before OpenAI.

Frankly, I want the freeze. Founders I back who build on these models get steamrolled every time a new one ships, and two quiet quarters would be the best thing to happen to them all year. So I read the essay wanting it to be true, which is roughly the position I have spent a thousand words accusing Dario of.

The freeze I opened with held for nine days. It broke when the team that called it needed one more thing in the release. The exception request went in on a Friday afternoon, the release manager approved it inside ten minutes, and the freeze stayed in force for everyone else. I know the timing because I wrote the request. It was a good feature. It shipped. I still think the freeze was the right call, and in all the years since, I have never once been the team it stopped.

— SWEdonym