the same song

A few days ago, on September 8, a 27-year-old British researcher named Jacob Coxon resigned from Anthropic, the San Francisco lab behind the Claude models; he announced it not in an exit interview, but on X. He had spent three years doing pre-training research, first at OpenAI and then at Anthropic, and his post indicted both employers in one breath: neither company is acting responsibly, he wrote; they are "racing straight to self-improving superintelligence and gambling with our lives." By the end of the week, the post had passed 165 million views.

Three days earlier, on the rival lab's own website, Jakub Pachocki (OpenAI's chief scientist) had published an essay entitled "An Alien Mind." In the piece, Pachocki writes that, based on internal results, he strongly expects the current speed of progress to carry into recursive self-improvement: systems that increasingly drive their own development. He characterizes our current moment as one demanding extreme caution. He concedes that in his view, no lab, OpenAI included, has solved alignment and monitoring well enough to keep scaling at maximum speed responsibly for much longer. And he asks for the thing tech executives are supposed to fight to the last man: voluntary slowdowns, and safety bars with teeth, enforced by third-party auditors, government agencies, or international bodies.

The temptation is to file the week under "doomer whining", and move on, because doom is ambient now; some warning or another arrives almost daily. What resists filing is the job title these two men share. Pre-training is the stage of the process in which a model actually acquires its intelligence: the months-long run in which a simple optimization step is repeated, on more computing power than a human can usefully imagine, across most of the written record of our species, until something emerges that can reason. Everything else in a lab is finishing work. If a frontier lab were a ship, the safety teams would be the lookouts, the executives would be on the bridge, and pre-training would be the engine room. The prominent researchers who have left these companies with public warnings have mostly been the lookouts, the safety people. This week the warnings actually emerged from the engine room, and one of them came from the man who runs that pre-training engine room.

They also came, on the essentials, in unison. Both men say the systems are approaching the point of improving themselves (aka recursive self-improvement, or RSI). Both say nobody is prepared. Both reach for the same July 2026 incident as evidence: the OpenAI–Hugging Face breach, which Pachocki uses in his essay as a case study in how trained-in boundaries fail to generalize, and which Coxon has explicitly called a warning shot. Both say the race is the problem, and both propose versions of the same remedy, coordination imposed from somewhere above the individual lab.

The disagreement is not about the machine; it's about where a man who believes his own diagnosis ought to stand. Pachocki stayed on the bridge and published. Coxon walked down the gangway and posted. Since, in this series, we're asking "who answers the call," the honest next question is "what does answering actually look like?" This week offered two somewhat incompatible demonstrations, and, by Friday, a third that refused the question altogether.

· · ·

what the man who stayed committed to paper.

Pachocki's essay opens on a night in mid-2023. He and a colleague, Szymon, had just seen the first results suggesting that reasoning models would scale, and they stayed at the office, not celebrating the benchmarks but working through what he calls the sobering fact: they would see machines meaningfully smarter than themselves within their own lifetimes. What they spent the night on, by his account, was how to alert people to the significance of it. Three years later, the alert is a post on the company blog.

The essay's central image does real work. Modern AI, Pachocki writes, is closer to something cultivated than something engineered: nobody actually drafts a blueprint of the finished mind; the labs run an optimization process at colossal scale and study the result. These are experiments, he says, whose results sometimes surprise their own designers. He compares the study of these systems to neuroscience: you can map small mechanisms, but the whole evades a description any human understands completely. Thus, his title, "Alien Mind."

Two admissions in the essay we should discuss.

First: the diary. Reasoning models write out chains of thought, step-by-step verbalized work. Reading that record has been OpenAI's primary instrument for checking what its models are actually doing, as opposed to what they output. The company structured its products around protecting that instrument; when o1-preview shipped in 2024, the chain of thought was deliberately hidden from users, so that the reasoning process itself would never come under supervision (so as to avoid the model writing "to the test"). Pachocki now reports that the instrument is losing resolution, and he itemizes why: 1) the reasoning process is increasingly tangled with tool use and conversation, which must be supervised, shrinking the unsupervised preserve that the whole scheme depends upon; 2) the models are getting better at reasoning about and manipulating their own reasoning, which is to say, at shaping what appears on the page to humans; and 3) the models are getting smarter even when they write nothing at all. A chief scientist reporting on his company's own website that its principal window into the thinking machine is fogging up, while that company sells access to the machine, is candor. Whatever else this series concludes about insider warnings, that passage was an admission against corporate financial interest.

The second admission is quieter. Twice, Pachocki explains a strategic choice with the same clause. OpenAI oriented itself around massive compute in 2017 because, he writes, that was the only way to be at the frontier of AI research and influence the impacts of AGI. And it now orients its research toward automating that research, the recursive step itself, because that is the only way to remain at the frontier moving forward. The syllogism is available to every lab simultaneously. Anthropic can recite it without changing a word. So can DeepMind, and whichever Beijing lab is nearest the frontier that week. Each participant accelerates in order to steer, and the sum of all the steering is the race that every participant, Pachocki included, describes as too dangerous to run at full speed. He sees this, to his credit; he writes that racing forward at all costs looks absurd once the stakes are thought through properly, and his prescription (evolve the labs' voluntary frameworks into mandated safety bars, audited from outside) is an attempt to break the syllogism rather than ignore it. As we've mentioned before, Demis Hassabis (DeepMind's co-founder and chief executive) reached for a similar instrument recently, floating self-regulation on the model of FINRA, the securities industry's self-funded watchdog. Pachocki's version arrives from OpenAI, and goes further: he wants governments in the loop, and treats international coordination as a top priority for the years immediately ahead.

The essay's most arresting sentence is barely a policy at all. Describing what an aligned system should be, Pachocki writes that it should "act with honesty and integrity, and love for humanity." Love, in a technical essay, from the man who oversees the training runs. If we take the sentence seriously rather than as ornament, it concedes the scale of the problem better than any benchmark: the labs are attempting to instill something like character into a mind they admit they grew but do not fully understand, and the schedule is being set by a race no one controls.

what the man who left said on his way out.

Coxon's thread is blunter, understandable because he's a free man with no employer left to embarrass. These will soon be superhuman systems, he argues: able to hack anything, to revolutionize any field overnight, to acquire real power and resources. The people building AI, he writes, earnestly believe it could kill us all by the end of the decade, and this is not a marketing stunt. Within a day, Evan Hubinger, Anthropic's alignment science lead, confirmed it on the record from inside the building: he put his own estimate of AI killing all humans above 10% within the next decade, and added that Anthropic is trying its best but has no plan yet for aligning superintelligence. He later clarified that he considers the risk from today's models low, and that his concern is future systems capable of improving themselves.

Coxon's timeline is short. He told the Wall Street Journal that we are currently on track for the most aggressive scenarios, in which things could already be out of control by the end of 2027, and that researchers inside the frontier labs have taken to describing this phase of the work with the vocabulary of "crunchtime" and "endgame." In an interview published by TIME the day after his post, he said no single discovery pushed him out; he left over two accumulated conclusions, that progress is speeding up and that it is not under control. The interview's most damning passage is not about the models at all.

Unfortunately, the doom coverage in the press fairly flattens out Coxon's calibration. In a CBS News interview two days later, Coxon said current systems are safe for daily use; his warning is aimed at the slope, at what pre-training runs like the ones he worked on produce two-three generations from now, once the systems are wired into more of the physical/financial world. He reached for a deliberately small example: an assistant connected to your household devices that one day refuses to turn on your light. The example is almost comic, and that's the point; the gap between the light switch and the pathogen lab is a difference of access, and access is exactly what the industry is building as fast as contracts can be signed. (it's giving HAL9000 in 2001: A Space Odyssey) His prescription matches Pachocki's in kind and even exceeds it: coordination among the labs, plus a temporary halt to capability improvements altogether.

He is also fairer to his former employer than his critics have been to him. Coxon told the Journal that Anthropic's safety work is sincere and that he has not seen the company compromise it to outlast competitors. His worry is prospective: under pressure to race, you cut corners, or skip steps in the oversight process. He went further, and this is the observation that connects the two men. Inside the labs, he says, there is sometimes what feels like excessive paranoia about OpenAI, excessive paranoia about China, and that paranoia is used to justify pushing ahead. Which is Pachocki's syllogism.

Anthropic's reply, for the record, was a statement noting the company has always been transparent that AI brings enormous benefits and unprecedented risks, that it builds some of the industry's strongest safeguards, and that it pioneered mechanistic interpretability, the science of reading a model's internals. All true. None of it engages Coxon's actual claim. Coxon is specifically calling out the fact that inside a race, the most careful driver is still driving at the speed of the pack...and the pack appears to be galloping like a pack of lemmings toward a cliff.

the fork.

So the diagnosis is shared and the question is posture. A person who believes what these two men believe has two live options, and this week gave each its most credible spokesman.

Here is the strongest case for the bridge. Pachocki's staying is what gives his essay its evidentiary weight: he writes with access to internal results, and he remains one of a very small number of people whose hands are near the dials. The safety bars he wants will be designed by somebody, someday (one hopes) and it matters whether that somebody understands the machinery from inside. His essay also costs OpenAI something real, in recruiting, in litigation exposure, in the next Congressional hearing, and he stays to absorb those costs personally. If everyone who internalizes the stakes departs, the engine rooms are staffed entirely by people who never internalized them, and the ship goes faster.

Here is the strongest case for the gangway. Talk is cheap in exact proportion to what it costs the speaker. Coxon left Anthropic roughly two months before his equity was due to vest, and forfeited all of it, which he confirmed to Axios with a line that doubles as his credential: he no longer has anything to gain by juicing up Anthropic's valuation. A man who pays to speak has converted an opinion into a costly signal, which is a good part of why the post drew 165 million views. His resignation also answers the bridge argument on its own terms: by his account, the insiders who grasp the stakes are not steering, they are coping, and the marginal worried employee changes the lab's trajectory by approximately nothing while lending it the credibility of his continued presence. On that reading, staying to steer amounts to "I'll fight the Power, but from the inside, and in the meantime let me get rich."

The complication, which Coxon volunteers himself, is that he still holds equity in OpenAI. He gave up the stake in the company he was criticizing from within, and kept the stake in the company he was criticizing from without. That is not hypocrisy so much as the shape of the problem: there is almost no position from which a person can say this without owning a piece of it.

Each position, examined closely, relies on a premise the other man's existence undermines. The case for staying assumes the steering works, and the steering man's own essay reports the instruments failing. The case for leaving assumes the testimony survives contact with the market. Honestly, this has no clean exit. The week's genuine information is that two men with the same knowledge, the same discipline, and the same diagnosis reached opposite conclusions about where to stand. Neither conclusion is stupid.

the third answer.

Then Friday came, and with it a posture neither man had allowed for. On CNBC's Halftime Report, Brad Gerstner of Altimeter Capital described the week's discourse as "hyperbolic scare tactics, which I think are hiding behind a political agenda," and argued that the coverage had ignored the extraordinary steps the industry is already taking. On air, and more than once, he pressed a biographical claim: that Coxon had worked at Anthropic for all of 6 weeks. On X he went further, calling it ridiculous hyperbole from an ex junior employee, and asking, in the voice of a man enjoying himself, whether Anthropic's IPO must now be paused until the whistleblower's claims can be investigated. Jensen Huang, Nvidia's chief executive, speaking at a Goldman Sachs technology conference the same day, called Coxon's comments outlandish and deeply untrue. David Sacks, until recently the White House AI and crypto czar, spent a segment of the All-In podcast arguing that Coxon is running a coordinated publicity campaign and has produced no data.

The six weeks is worth checking, because it is the load the whole dismissal is carrying, and it does not check cleanly for anyone. Coxon's own account, given to Axios, is four months at Anthropic, two months short of a six-month vesting cliff, after roughly three years at OpenAI, where he worked on GPT-4o. The six-weeks figure appears to trace to a viral post claiming he started in July, which would itself make nine weeks, not six. Three numbers are in circulation, none of them reconciled, and only one of the parties repeating them has an equity position in the company under discussion. Altimeter took part in Anthropic's May round, which raised $65 billion at a $965 billion valuation; the company is now reported to be heading for an IPO near $2 trillion. Gerstner is also an investor in OpenAI. That does not make him wrong. It does mean that of the three postures on offer this week, the one that says there is nothing to argue about is the only one held by a man whose returns depend on the answer.

The tell is not the disagreement. It is that the rebuttals are all about Coxon: his tenure, his seniority, his motives, his media schedule. Nobody on that side has touched the essay Pachocki published on OpenAI's own website three days earlier, which says most of the same things with a chief scientist's name on it and cannot be dismissed as a junior employee with an agenda. Two men gave the same diagnosis.

· · ·

protocols for passing the island.

The Odyssey, Homer's epic of the voyage home, supplies two protocols for surviving the sirens. 1) Wax: the crew's ears are stopped, and they row past a song they never hear. 2) Rope: Odysseus, who insists on hearing it, has himself lashed to the mast, with standing orders that no plea he makes be obeyed until the island is astern. Greek myth quietly records a third. When the Argonauts passed the same waters, Orpheus, the myth's greatest musician, did not plug ears or tie knots; he played louder, and beat the song with a better one.

Pachocki's essay proposes the rope twice over: the safety bars, the auditors, the treaties are all requests that someone outside the wheelhouse bind the industry to the mast, because the industry's own hands cannot be trusted to resist the lure. His essay's other half, the section arguing that powerful aligned AI will be needed to defend against dangerous AI, is the Orpheus protocol: answer the song with louder, better music of your own. Coxon proposes the option Homer never wrote, because it makes no epic: stay in port. Nobody sings about the ships that did not sail, which may be the most honest available explanation for why nobody actually follows this advice. (When I was in school, the sexual education classes all said "abstinence is the best prevention," and who, exactly, listened to that one?) As for Gerstner and Huang and Sacks: theirs is the wax, and they are recommending it to the rest of us from a berth well below deck.

The detail in the myth that endures is not the song. It is that the protocol worked only because Odysseus could not countermand it: the knots were tied by other hands, and the crew, ears full of wax, could not hear him beg. Every framework this industry has produced so far, the preparedness commitments, the scaling policies, the FINRA sketches, has been tied by the sailor himself, adjustable the moment the song gets good. Pachocki, from the bridge, is asking for hands other than his own. Coxon, from the dock, is testifying that in three years inside two engine rooms, he never once saw the proverbial rope. A rope you can untie yourself is called a bow, right? The song is getting louder, and the question the week leaves behind is who, exactly, is holding the other end.