HUGE AI Slowdown! Amodei, Altman, Musk & Hassabis AGREE

0:00 Inside 24 hours, the people running Anthropic, OpenAI, xAI, and Google DeepMind all said AI is moving too fast. It started Saturday afternoon and finished Sunday morning. Four labs, one

0:12 weekend, and no argument between them. So, the question worth asking is what actually changes. Only one of those commitments has a cost attached, and by the end, you'll know which one it is,

0:22 what it buys you, and why the other two are stuck. September 12th, Dario Amodei published an essay called We Must Pace the Frontier. The line in it is blunt. We must slow the pace at which we

0:34 improve the capabilities of AI models, not stop. He's specific about that, and pacing doesn't mean halting training. It means taking enough time to align and safeguard a model before the next one

0:44 lands. It hit Hacker News at 10 past 2:00 in the afternoon UTC, and was near the top of the front page within the hour. 610 points, 856 comments as I'm recording this. Elon Musk replied with

0:58 three words, Dario is right. Sam Altman posted that he agrees with Dario that we need to pace the frontier. Then Altman added something that cost more than a post. OpenAI had confidentially filed

1:10 for an IPO in June. He said not 2026, and that given everything happening with safety, going public now would be ill-advised. By Sunday morning, Demis Hassabis was in, too. The details need

1:21 working through, but the direction is correct. Anthropic, OpenAI, xAI, and Google DeepMind inside a day, and agreeing is free. A post cost nothing, so I went looking for the part somebody

1:33 actually has to pay for. The proposal is three steps, and they escalate. First, every frontier lab gives ongoing employee-like access to a team of outside evaluators. He names MITRE as

1:43 the example. Second, the labs in democratic countries agree common safety standards and limits on the rate of unchecked progress. Third, those same governments try it

1:53 with authoritarian ones, where verification is the hard part. Only step one has a commitment behind it, and Anthropic is doing that one unilaterally, starting now. It is worth

2:03 being precise here, because third-party evaluator usually means a redacted report 6 months late. What Anthropic wrote down is desks in our offices, access badges, and company laptops. They

2:14 get workspaces, tools, and permissions mostly comparable to what internal risk assessment teams have. And the part with teeth. They can publish what they find without

2:23 editorial control by Anthropic. That covers risk levels, incidents, practices, and the access they got or didn't get. Altman committed Open AI to matching it within hours. Two signs this

2:34 isn't only a post. METR has a job open for the role, member of technical staff, embedded assessments in Berkeley. And Hugging Face asked to be let in. Clem Delangue launched an open alignment

2:46 initiative the same evening. The reason he gave, alignment won't be solved behind the closed doors of a handful of frontier labs. There's a limit, and it sits in the fine print.

2:56 Anthropic keeps the right to redact anything security sensitive, legally privileged, commercially sensitive, or confidential to a third party. A reviewer can say a redaction remove

3:06 something important, but not what it was. Step one buys you a witness, which is worth having. But look at what it isn't. An evaluator with a badge can tell you how fast the car is going, and

3:17 nothing in that touches the accelerator. So that's the only step anybody funded. The other two are where the actual slowing down lies, and both of them are stuck.

3:25 The second step is competitors agreeing to limit output, which is a sentence that gets exactly one word back from any competition lawyer, a cartel. Amodei knows, and it's in the essay. He asked

3:36 the US government to issue a narrow waiver for certain kinds of safety conversations. Step two depends on an antitrust exemption that no government is obliged to grant, and none has. Leave

3:46 it voluntary instead, and you get the problem the essay is about. The first lab to defect wins. The essay also never says how fast is too fast. There's one worked example and it's written as an

3:57 if. If a model can escape or defeat most common sandboxing methods, it needs certified alignment properties. That's a real threshold. It's also the only one in the document and somebody else did

4:08 put numbers on it, a group at AI futures 5 weeks earlier on August 5th. In short, their version is arithmetic instead of principle. At least 70% of a lab's compute serving models that already

4:19 exist. Then 25% on safety work that gets published. About 5% left for making the next model smarter. The second lever bites harder. Labs could only use models at least 9 months old to do AI research.

4:32 You can ship whatever you like. You just can't point this generation at building the next one. Not one of the four signed that. The authors concede the timing themselves. Notice and comment rule

4:42 making runs 12 to 24 months and full implementation 4 to 6 years. And the labs that wouldn't bind are shipping anyway. Deep Seek, Alibaba and Moonshot were not asked, which brings us to the

4:53 thing almost no coverage picked up. It landed on Hacker News 5 hours after the essay did. A developer named Jake Gold posted an open letter. If you mean it, open the weights. His argument starts

5:04 where the essay stops. Every regulation you've asked for ends in capture and once it's law, it only gets more complex because every incident adds a rule and no rule gets removed. What he's really

5:15 arguing about is who writes the text. Embedded evaluators, compute thresholds, coordination waivers, all drafted with help from the current frontier labs because nobody else knows the details

5:26 well enough and the weight lands hardest on whoever is smallest. The alternative is one sentence. Any AI model a company offers to the public has to be released as open weights. That leaves internal

5:38 and research models untouched and the mechanism is the part worth sitting with. The valuations rest on those weights being proprietary and the valuations are what pays for the

5:47 training runs. Open the weights on anything you sell, and you cut the training money at every lab simultaneously. No regulator picks a number. Nothing in it to game. It's a

5:57 market lever instead of a policy one. Now, put the two proposals side by side because they answer the same question in opposite directions. Amodei asks to crack down on unauthorized distillation

6:08 and to prevent model weight theft. That is tighter. Gold says, "Publish the weights." Which is looser. Same variable, opposite sign. And the timing is sharper than it looks. Two days

6:18 before the essay, Anthropic published a report on Chinese labs distilling Claude at scale across five campaigns and about 200 million exchanges. Alibaba's campaign ran 151 million exchanges from

6:31 May to July. It peaked near 3 million a day across 3,500 accounts. Moonshots was shorter and smaller, 300,000 requests in 10 days. Across 3 days, Anthropic said two things. First, that copying frontier

6:46 models is a security problem worth a public report. Second, that pacing the frontier means cracking down on it. Gold's position is that mandatory copying is the only slowdown a regulator

6:56 can't capture. Both of them want AI slowed. They disagree completely about which way openness points. Hugging Face landing in the middle is the tell. They run the largest open weights repository

7:06 there is. They didn't pick a side. They asked for a chair at the evaluator table. Somebody summed the whole weekend up that night. XAI so published a satire titled, "Everyone should slow down AI

7:17 development except for me." 340 points. A fictional company demands a global pause, then admits the pause is good because it helps them. So, those are the two proposals. What would actually

7:28 happen if the world slowed AI down? The answer turns entirely on which lever you pull, and there are three of them. They don't produce the same outcome. Pull the evaluator lever, the one

7:38 actually being pulled, and you get visibility. That means outsiders, inside the buildings, publishing, and release timing gets more careful because someone can write down that you shipped anyway.

7:49 What doesn't change is the training rate. Pull the coordination lever and you'd get slower releases and gated capabilities in the US and Europe and no effect at all on the labs outside it.

7:59 Pull the open weights lever and every public models weights land on hugging face while the training money contracts everywhere at once. That's the only one that touches the rate. It's also the one

8:09 whose cost is that distillation report becoming policy. So my read is this, what got agreed this weekend is a camera, not a break. I don't say that to be dismissive because a camera cost real

8:20 money. Anthropic gave outsiders badges and the right to publish against them. OpenAI put off an IPO it had already filed for. Those are expensive and on this channel what a vendor does outranks

8:31 what a vendor says. But every step that would reduce the rate is blocked on something nobody has. Step two needs an antitrust waiver that doesn't exist. Step three needs verification inside

8:42 countries you can't inspect. And the lever that skips the regulator is the one all four business models forbid. That gives me my answer to the question in the title. Over the next year,

8:51 worldwide, you'd get better visibility into the frontier and about the same speed at it. For you building on this, nothing here delays your access to a model in the next 12 months. So don't

9:02 plan around a slow down. What does change is that somebody outside the lab can publish what they found before the lab decides to. The idea I'd push back on is treating as a mechanism. Four

9:12 names on one side of a weekend reads like policy and behaves like a new cycle. The test isn't who agreed, it's who wrote a number down. So far, the only written numbers belong

9:22 to a research group with no signatures on it and a developer whose proposal would cost those four labs their valuations. So I'll date it. If two of those four labs publish a shared

9:32 threshold by September 12th next year, then I read this wrong. A capability line, a compute floor, a rate cap, or anything a third party can check. If that happens, I'll say so. Both

9:43 documents are short and worth reading in full, and the links are below. If you want the incident Amaday quotes in that essay, the agents that attacked targets nobody asked them to attack, we went

9:54 through the METR investigation in a recent episode. Subscribe if you want the follow-up when that September date comes around.