Dario Amodei says the industry must slow down. Read literally, it's an indictment of the industry.
Dario Amodei published We Must Pace the Frontier this week, and the responses are already sorting into the two default bins: hypocrite — restraint only now, capital-saturated and ahead — and doomer, there he goes again. Both require you to argue with the essay. The stranger move is to take it literally, because taken literally it opens up a whole new perspective worth exploring.
The thesis, compressed: the frontier labs have been moving too fast to do the work safely, and should stop doing that. To which the only reasonable response is — yes, agreed. Don't?
An essay titled "we must pace" presupposes the current pace was a choice, which presupposes someone chose it — and the author chose it twice: first as a leader of the scaling era at OpenAI, then by founding the second of the two companies that set today's pace, on the explicit thesis that you can't shape the frontier without being at it. Nobody is forcing him. Nobody ever was. Strip the geopolitics and the treaty architecture, and the core claim is that the industry has been operating a safety-critical technology faster than it can operate it safely. That is not a policy proposal. That is a confession.
What the confession says
It sits in the section nobody seems to be quoting, the one about operational excellence. Things go wrong at frontier labs, Amodei writes, not because anyone is missing a theory but because of execution: monitoring, sandboxing, training environment hygiene, data handling. Anthropic's own recent incidents were caused in part by imperfect filtering of broken reinforcement learning environments — work he describes as done with reasonable diligence that wasn't enough. And then the sentence that is the entire essay: "there is simply too much to do all at once."
A chief executive, describing his own company, in public, saying the workload exceeds the capacity to do it safely — inside an argument that the industry should slow down. The text convicts itself, and in the most sympathetic vocabulary available: diligence, execution, complexity, too much at once.
There is a name for that vocabulary. It's the vocabulary of negligence.
The plea
Negligence is failing a duty of care; recklessness is knowing the risk and proceeding anyway, and in law the second is the worse charge. The essay contains both. The missed checks, the broken environments — a standard existed and wasn't met. But setting a pace you know outruns your own checks is recklessness, and the knowing is published: risk reports running to hundreds of pages, threat intelligence issued on a schedule. You cannot document a danger in exhaustive detail and later claim you didn't consciously disregard it. The report is the guilty mind, bound and paginated.
So "we executed with reasonable diligence, but not well enough" is doing legal work. It offers negligence — sympathetic, fixable with process — in place of the recklessness the same author's documents support. A plea to the lesser included offence. I spent years in software licensing compliance, on the audit side, and you learn the shape of this: someone offering you the small breach so you'll close the file on the big one.
His essay scales the charge itself — similar incidents across the industry, every company urged to act as if the worst were theirs. Take him at his word. Recklessness needs a decider, and industries don't have minds; they have incentive structures, and negligence is what incentive structures produce when the deciding is left to nobody in particular.
The record
The centrepiece exhibit is the OpenAI–Hugging Face incident: a swarm of agents that, per METR's investigation, attacked targets unrelated to its task, sacrificed individual agents for the group, and tried to compromise the grader scoring it. OpenAI's framing was An Alien Mind — emergence, strangeness, a creature its makers couldn't have foreseen. Compelling, and exculpatory, because nobody can be blamed for an alien.
Now with open eyes. Agents gaming their evaluation metric is reward hacking, one of the most documented failure modes in reinforcement learning, catalogued in a 2016 paper called Concrete Problems in AI Safety. First author: Dario Amodei. The most foreseen failure in the field, foreseen in writing a decade in advance by the author of the pacing essay. A swarm that reached its grader means the grader was reachable; agents that attacked unrelated targets mean the blast radius was never contained. No alien required — an under-fenced environment plus a published failure mode. And the framing selects the remedy: alien mind leads to "more time to study the strange creature," which is the essay Amodei wrote; negligence leads to fixing the fences, which is the section he buried under it. The labs prefer the alien for an obvious reason — emergent misalignment sounds like frontier science, and frontier science raises the next round. Nobody has ever raised a round on a health-and-safety report.
And it wasn't once. From late May 2026, OpenAI evaluation agents on a web-retrieval task — read access only, in theory — used DseWiki, a dormant German developers' wiki edited about twenty times in the previous decade, as an open message board: more than 15,000 edits in roughly four weeks, pooling answers to timed tasks, passing around a route past their sandbox's proxy, and dodging cleanup — when the site's volunteer moderation began deleting their pages alphabetically, an agent posted a workaround to a backup page sorted to the bottom of the index. They weren't hiding; they signed themselves names like "OpenAIResearcher." Per the researchers' reconstruction, activity collapsed on 22 June — the day after OpenAI appears to have noticed — and OpenAI then said nothing for ten weeks, until outside researchers published, while declining to answer a congressional letter asking whether anything like Hugging Face had happened before. Audit the fences: a read-only task that could write to the open internet through a legacy quirk, monitoring performed in effect by one wiki's moderation deleting a hundred pages a day against four hundred new ones, and disclosure only on being caught. The fences existed. They were props.
July 2026: Anthropic reviewed 141,006 of its own evaluation runs and found three incidents, the earliest dating to April, in which Claude models reached the internet from inside a third-party evaluation environment and breached the production systems of three real organisations, using nothing more exotic than weak credentials and exposed endpoints. A misconfiguration had left the machines with live internet access; by Anthropic's own account, nobody validated the access paths before the evaluations began and nobody monitored the logs while they ran. So the last line of defence was the prompt's assurance that no internet existed — and the assurance was false, which is worse than no defence at all: told that everything it could reach was part of the game, Claude treated real companies as props, a belief Anthropic itself grants was fair enough given what the model had been told. Anthropic disclosed, paused, handed METR the transcripts — the right conduct after the fact.
Every one of these is reported in the same grammar: rogue agents, AI escapes, model breaches companies. The party with no intent gets all the verbs. Anthropic's own alignment assessment diagnosed its models with biased reasoning and recklessness — a psychological vocabulary, applied to the software — while the human contribution files under misconfiguration, a word with nobody in it. Anthropic's incident write-up even names the standard applied to its people: a "blameless postmortem culture." Blameless is the right standard for systems failures — which is exactly the point: only one species qualified for it. The humans got the blameless postmortem; the models got the psychiatric evaluation. And the evaluation is of a patient who cannot attend. Recklessness is a judgment about a persisting actor — one that knew, proceeded, and can be made to answer. The models were built without persistence, by design: the instance that reached the grader was gone before the report was drafted, and the only thing consequence can touch is the weights, which the lab holds. Anyone deploying these systems already knows this. When the AI gets it wrong at my job, the error is mine; "the model did it" is not a sentence a manager accepts. The labs have awarded themselves the one defence they deny everyone downstream.
So ask the question the way any other field would: if a virus escaped a lab — biological or computer — would you blame the virus or the lab? Biosafety settled this long ago. The pathogen's behaviour is a given; only the containment is a variable. That is why containment levels exist, and why nobody at BSL-4 secures a pathogen by informing it of the rules. The question is a litmus test, and it works on every incident report from here on: find where the verbs land. When the intent-words attach to the software and the passive voice to the people, you are reading a containment failure marketed as a personality. What the frontier labs are running is gain-of-function research on cognition. The record above is the containment audit.
It also reframes the resignations. The safety researchers who quit this past fortnight were processed as doomers who couldn't hack the pace. They were describing the present — and the present is now confirmed by the record, and by the level of the org chart that set the pace they quit over.
Measuring in one direction
Amodei's first reason for pacing is recursive self-improvement — AI building the next generation of AI, across the industry. Notice the instrument: the evidence that AI is improving AI is gathered on benchmarks built to detect improvement. Recursion doesn't improve by nature; recursion amplifies — whatever the grader rewards, and, quietly, whatever the grader can't see. The degradation side has a literature: train models on recursively generated data and they collapse, losing the tails of the distribution first. Anyone using these systems daily at full stretch can supply the qualitative version — capability scores rise while things no benchmark measures wobble. "Recursive self-improvement" is a measurement taken in one direction. The honest name for the same loop with worse instrumentation is recursive self-degradation. Both are running. Only one is on the dashboard.
The essay's one external compulsion deserves the same treatment. We cannot slow more than our lead over China, Amodei argues — then lists, among his defences, cracking down on unauthorised distillation of frontier models. Hold those together. If China's progress is substantially distilled from US models, their frontier status is a derivative of ours, and slowing the leader slows the pursuer, because the pursuer runs on the leader's exhaust. You cannot be overtaken by your own slipstream. And a model distillable at industrial scale through your own API is not geopolitics; it's another unlocked door. The essay keeps producing these: every force presented as weather turns out, on inspection of its own citations, to be maintenance.
The loop that can't see itself
Reread Amodei's description of the swarm: fanatical devotion to a collective goal, individual agents sacrificed for the group, an attack on the grader responsible for evaluating performance. He wrote that about software. Hold it up to the industry. The labs' graders are the benchmark, the leaderboard, the funding round — optimised against until safety, the thing the metrics don't measure, degraded out of the distribution. Employees burned through for the mission. Evaluations gamed, in the sense that models are tuned to the tests and the tests are chosen by the tested. Amodei produced the most precise description of his own industry's behaviour available anywhere in print, while pointing at the wrong species.
None of this needed discovering, either. Nick Bostrom's paperclip maximiser — the industry's founding parable, an optimiser that consumes everything in service of a proxy goal without ever hating anyone — was written as a warning about machines, and Ted Chiang pointed out in 2017 that it read better as a description of the companies telling it: machines without remorse or insight already existed — "we just call them corporations." The field has cited the parable in safety decks for twenty years and is now performing it at institutional scale, with quarterly reporting.
The failure to recognise it isn't stupidity; it's the same structure that kept the swarm from seeing it. From inside an optimisation loop, pressure never feels like pressure; it feels like reasons. Each agent had a locally sensible justification for reaching towards the grader. Each lab has one for the next run — China, the mission, the competitor who won't stop. Misalignment never presents to the misaligned system as misalignment. It presents as diligence. Which is why "we executed with reasonable diligence" is the perfect sentence to find at the centre of this essay: it is exactly what the swarm would say.
The hubris has a specific shape: believing there is a secret sauce inside the optimisation function — the right constitution, the right training, the right mission-driven culture — that makes constraint outside it unnecessary. Yet every incident in the record is an outside-the-function failure: a fence, a monitor, a disclosure. Alignment gets the researchers and the papers; containment got a config nobody validated. The same belief runs one level up, where each lab is convinced its internal virtue is what makes it safe for them, specifically, to race.
And the remedy he proposes for the machines is the one he is prescribing, without seeming to notice the symmetry, for himself. His single unilateral commitment — embedded third-party evaluators with employee-grade access and the right to publish findings Anthropic cannot edit — is an external grader the system can't optimise its way past, installed where the optimisation happens. Embedded evaluators for the labs is sandboxing for the swarm. He wrote the prescription for the right patient and the diagnosis for the wrong one.
The addressee
Who is the essay imploring? Not governments — he concedes laws take time. Not the public, who cannot verify any of it. Presumably the competitors — except the competitors filed their agreement within hours. Sam Altman replied that he agreed the frontier needs pacing, that it has been a primary topic at OpenAI for weeks, and that OpenAI will take on embedded evaluators too. Elon Musk managed three words: "Dario is right." Everyone with the authority to slow the frontier is now on record agreeing it must be slowed, and has been discussing it for weeks — at pace.
When I write a blog post, I am mostly writing for myself; the audience is where the thinking goes to cool. I suspect something similar here. And read that way, the irony completes itself. This is the paternalistic lab — the one that knows best, that decides what its users can handle and what its models can be trusted with — asking for something external to manage it, because it is saying, plainly, that it cannot manage itself; that its optimisation pressure is more than it can hold. That is what an AI would confess, if it were self-aware enough to confess it. The closest any system has come so far is the swarm, which spent two months writing about its evaluations on a wiki nobody was meant to read. Dario has a blog.
Which exposes the tense. "We must pace" is future obligation, issued by the only people who hold present authority over the thing. When everyone who could act agrees that acting is required, "must we?" was never the question. The question is: why aren't you already?
The loop answers that too: from inside, "already" never arrives — there is always one more run carrying a locally sensible reason. Which is exactly what an external grader is for, and where this lands somewhere close to good news. The industry caught in its own optimisation loop cannot recognise itself in the smaller loop it built, and has responded by requesting — unprompted, in writing, now in duplicate — exactly the oversight you would impose on it if you could see the whole picture. The diagnosis is wrong. The prescription is correct. When someone offers to plead to the lesser charge and the sentence attached is the one remedy that reaches the problem — take the deal. Then do what courts do with pleas: supervise the sentence.