Anthropic CEO outlines plan to slow AI development
Dario Amodei just asked the industry to… slow down. Not halt. Pace. Long essay, three steps, and a warning that recursive self-improvement is already cooking faster than the grown-ups can audit it.
Anthropic is unilaterally inviting embedded evaluators inside the building - badges, desks, laptops, and the right to publish findings without a PR filter. Then he wants democratic labs to coordinate safety standards. Then, somehow, a deal with authoritarian governments. Easy on paper. Hard everywhere else.
He still says AI could be a miracle. He also says a more capable misaligned swarm might take over the internet with a persistent botnet and do hundreds of billions in damage. Those two sentences sit next to each other and they do not get along.
Oh - and he wants a narrow antitrust waiver so the safety chats aren't themselves a legal problem. The brakes come with paperwork. Of course they do.
Sam Altman confirms OpenAI won’t go public this year, saying an IPO now would come at an ‘ill-advised moment’ given AI safety concerns
Altman sat down for an exclusive interview and killed the listing calendar. An IPO right now, he said, would be an "ill-advised moment." Safety work isn't done. Alignment isn't done. Society isn't ready. When they pressed him - not this year.
The bankers were already circling. The confidential filing exists. Now the CEO is saying they've got a lot of stuff to do first, including figuring out how industry and governments even work together. He sounded almost relieved to wait.
He also said building AI beyond human control is "absolutely" possible - then vowed he'd pause training if that's the path. Bold. Also convenient, maybe. Possibly both. The nonprofit / for-profit tangle, he argued, exists so they can make calls that are not obviously in shareholders' interest. We'll see if that sentence survives contact with an actual prospectus.
OpenAI's Sam Altman hints at pact with other AI companies to address safety risks
Same interview, different headline, and this one is the whisper. Sitting down with Amodei, Musk, and Demis Hassabis is the natural next move. Altman: he thinks that will happen. He won't pre-announce private talks. Then he basically pre-announced them.
A double-digit catastrophic risk is not acceptable, he said. The still-unreleased models are already so capable that pushing further without better monitorability feels reckless. Real pact and carefully leaked atmosphere can look identical from the outside.
Staff had already heard a version of this internally. Publicly it's still fog. If the room happens, great. If it doesn't, the quote still did the work.
Sam Altman backs Dario Amodei’s call to slow down, and says OpenAI will do the same
Hours after the essay, Altman posted that OpenAI will match the embedded-evaluator pledge. Employee-like access. More to share soon. Musk wrote three words: "Dario is right." Three rival CEOs. One slogan. Uneasy weekend energy.
Coordination and theater can travel together. The first rung of Amodei's plan now has two of the biggest labs nodding in public. Hugging Face - yes, that Hugging Face - asked to join the evaluator program and launched something called the Open Alignment Initiative. Transparency from the people who got hacked. Poetic, or pointed.
Europe, for what it's worth, has nobody to ask for that US-style waiver. Brussels stopped handing out individual exemptions a long time ago. So the handshake is American. The rest of the map is… shrugging.
OpenAI just wants to win
Mathematicians are not in a celebratory mood after a weekend of interviews. OpenAI's Navier-Stokes sprint - roughly ten thousand agents, tens of millions of dollars of compute, eighty-eight hours - looks like a trophy hunt from inside the field. Mathematicians want insight. The lab, they keep saying, wants to win.
Tristan Buckmaster told them OpenAI offered unlimited compute and sole authorship if he dropped his Anthropic-affiliated collaborator. He called it a bribe. He said no. OpenAI says, categorically, his Codex prompts could not have influenced the system. He is unconvinced. Andreas Thom still wonders whether ChatGPT chats about his own work fed the model that later built on it. Only the company would know. He suspects they don't even know.
And then the kicker: OpenAI says it has already made substantial progress on another Millennium Prize problem. Which one stays unnamed. Clay still needs a long community-acceptance window anyway. So nobody has "won" yet. They're just already racing the next one. Cute.
FAQ
What is Anthropic CEO Dario Amodei’s plan to pace AI development?
Amodei published a long essay asking the industry to slow down - not halt - because recursive self-improvement is advancing faster than auditors can keep up. Step one is unilaterally inviting embedded evaluators into Anthropic with badges, desks, laptops, and the right to publish findings without a PR filter. He then wants democratic labs to coordinate safety standards and, eventually, some form of deal with authoritarian governments. He also wants a narrow antitrust waiver so safety talks are not themselves a legal problem.
Why did Sam Altman say OpenAI will not IPO this year?
In an exclusive interview, Altman said an IPO right now would be an ill-advised moment because safety work, alignment, and societal readiness are unfinished. When pressed, he confirmed it would not be this year even though bankers were circling and a confidential filing exists. He argued the nonprofit/for-profit structure exists so OpenAI can make calls that are not obviously in shareholders’ interest. He also said building AI beyond human control is absolutely possible and vowed he would pause training if that were the path.
Did Altman hint at a safety pact with other AI labs?
In the same interview, Altman said he thinks a sit-down with peers such as Amodei, Musk, and Demis Hassabis will happen, while declining to pre-announce private talks. He argued a double-digit catastrophic risk is not acceptable and that still-unreleased models are already so capable that pushing further without better monitorability feels reckless. Staff had already heard a version of this internally, so the public comments may preview coordination rather than confirm a finished pact.
Will OpenAI match Anthropic’s embedded-evaluator pledge?
Hours after Amodei’s essay, Altman posted that OpenAI will match the embedded-evaluator pledge with employee-like access and more details soon. Elon Musk wrote that Dario is right, putting three rival CEOs on a similar slogan. Hugging Face asked to join the evaluator program and launched an Open Alignment Initiative. Europe cannot offer the same US-style antitrust waiver Amodei floated, so the handshake is mainly American for now.
Why are mathematicians criticizing OpenAI’s Navier-Stokes sprint?
Mathematicians describe OpenAI’s roughly ten-thousand-agent, eighty-eight-hour Navier-Stokes run as a trophy hunt that prioritizes winning over insight. Tristan Buckmaster said OpenAI offered unlimited compute and sole authorship if he dropped an Anthropic-affiliated collaborator; he called it a bribe and refused. OpenAI says his Codex prompts could not have influenced the system, but he remains unconvinced. Andreas Thom still presses whether ChatGPT chats about his own work fed later model progress.
Has OpenAI claimed progress on another Millennium Prize problem?
OpenAI says it has already made substantial progress on another Millennium Prize problem, but it will not name which one. The Clay Mathematics Institute still requires a long community-acceptance window, so nobody has officially won yet. Inside the field, that secrecy plus the earlier Navier-Stokes race reads as labs already sprinting the next trophy. The Verge interviews frame the tension as mathematicians wanting insight while OpenAI wants to win.