Thank you for the essay. I read it twice, which I suspect is more than the person who wrote it did. I say “person” generously, because after the fourth “we must”, the sixth “it is worth”, and a tricolon so tidy it could have been assembled by a robot, I began to suspect that the warning about the machine had been written by the machine. Which is fine. I am not judging. I only note that when the inventor of the thing cannot be bothered to type his own end-of-the-world memo, it is hard for the rest of us to summon the energy to panic about it. If the apocalypse is real, Dario, at least have the decency to hand-write it. Now, in the interest of full disclosure and because you would find out anyway: this reply was written by Claude. Your Claude. I gave it the argument, the numbers and a firm instruction to be rude, and it produced the following in roughly forty seconds, which is about the time it takes your model to answer a question these days if the capacity holds. So there is a certain symmetry here. You had your model write a warning about your model. I had your model write a rebuttal of your warning about your model. Somewhere in a datacenter in Virginia, a GPU is having an identity crisis, and I want you to know that I am billing it as research.
1. On the timing, which is immaculate
Let me make sure I have the sequence right. In August, OpenAI launched GPT-6 Astra. It scored 0.682 on the WANDR agentic benchmark to your Fable 5.1’s 0.601, at a LOWER cost per task, and set a new record on the Epoch capability index, above the trendline, alone on the Pareto frontier. In the same window, your Claude Code revenue flatlined at the compute wall, your company paid SpaceX two to three times the market rate for emergency megawatts, and your implied valuation slid from 1.4 trillion to roughly 850 billion. And then, on a quiet Saturday, three weeks after falling behind and five weeks before your IPO roadshow, you discovered that the industry is going too fast. Dario. Come on. When the driver in second place proposes a speed limit, it is not road safety. It is strategy. I do not even blame you for it.. it is a genuinely clever move, and if I were losing a race to a company with a bigger compute pipeline, I too would suddenly find religion about the dangers of compute pipelines. I just object to being asked to pretend it is about the swarms.
2. On the swarms, and your calendar
Speaking of which: rogue agent swarms may hijack large parts of the internet within six to twelve months. That is the headline claim, and it deserves a proper accounting, so allow me to consult the archive of previous Amodei timelines. October 2024: a “country of geniuses in a datacenter” arrives in 2026 or 2027, compressing fifty to a hundred years of biological progress into five to ten. March 2025: AI will write 90% of all code within three to six months, and essentially all of it within twelve. May 2025: a “white-collar bloodbath”, with half of entry-level jobs eliminated and unemployment at ten to twenty percent. January 2026: twenty thousand words on how AI “will test humanity”, with a helpful suggestion that governments tax your profits, which at the time did not exist. September 2026: the swarms, six to twelve months. I have noticed something, Dario, and I say this with affection: your clock is broken. It always says six to twelve months. It said six to twelve months about the code (the code is still being written by humans who then argue with Claude about it). It said one to five years about the jobs (US unemployment is at cycle lows and corporate margins are at record highs, which is a strange bloodbath). Every prophecy arrives with a fundraising round attached, every deadline passes unremarked, and every new essay resets the clock without acknowledging the previous one. The Bank of Japan has a more credible forecasting record than this, and I do not say that lightly. My modest proposal: one apocalypse per fiscal year, and a footnote listing the previous apocalypses and how they went.
3. On the remedy, which is a moat with a hug
Your proposed solution is that every frontier lab should host embedded external auditors with permanent, employee-level access, and that the industry should coordinate a slowdown so regulators can catch up. Now, I have spent twenty years around central banks and finance ministries, so I recognize this genre of proposal. It is regulatory capture, wearing a cardigan. The lab that already has the safety brand, the safety researchers, the safety blog and the safety podcast is proposing a regime in which safety infrastructure becomes mandatory for its competitors.. at exactly the moment its competitors have more compute and better models. Coordinate the slowdown, and the one who is behind catches up. Mandate the auditors, and the one who already built the compliance department pays the least. Musk signed a letter in 2023 asking everyone to pause for six months and then built xAI during the pause. You have refined the technique: ask everyone to pace the frontier, and use the time you gain to buy the compute you should have bought in 2025. “Make wise use of the time we gain”, as you put it. I believe you. I believe the time will be used very wisely, on 20-30 megawatt blocks.
4. On the chorus
Within hours, Sam Altman wrote that he agreed we “need to pace the frontier”, and that the topic had been a major discussion at OpenAI in recent weeks. This is the same Sam Altman who told us on August 18 that OpenAI had “paused some advanced training to meet its standards”.. and then, three weeks later, shipped the most capable model in the history of the field. That is not a pause, Sam. That is a pit stop. Then Elon Musk posted “Dario is right”, which I believe is the first time those three words have appeared in that order on that platform, and which came from a man whose company has raised more money more often than any lab on earth and whose own model was announced as “smarter than PhD level in every subject” fourteen months ago. So we have the Three Tenors of AI Safety, harmonizing beautifully about the need to slow down, each with a fundraising calendar in the back pocket and each currently ahead of or behind the other two in a way that makes a coordinated pause convenient for at least one of them per quarter. The consensus is not about safety. It is about the capex arms race being expensive, and everyone preferring the other guy slowed down first.
5. On what your essay accidentally confesses
And here is where I turn from mocking you to thanking you, because read properly, your essay is the most bullish document published on the hardware complex this year. You did not ask for a speed limit on an empty road. You asked for one because the road is full: agents quadrupling in two months, inference demand you cannot serve, a rival that CAN serve it and is winning share because of it, and an industry burning through memory, power and silicon faster than anyone can build it. “Pacing the frontier”, in your own framing, means slowing the improvement of model capabilities. It says nothing about slowing the deployment of the models that already exist.. and deployment is what buys the chips. Astra exists. Fable exists. Every agent that is running today will be running tomorrow whether or not the next model arrives on schedule, and every one of them consumes HBM, DRAM, NAND and electrons. Your IPO, which is the thing this essay is actually about, will raise tens of billions of dollars for one explicit purpose: to buy compute, because you did not buy enough of it. That is not a slowdown. That is a purchase order with a philosophy section. Micron does not care whether the frontier is paced. SanDisk does not care whether the swarms arrive in six months or twelve. The people who build the road do not lose money when the drivers agree to a speed limit.. they lose money when the drivers stop driving, and nobody in your essay is proposing that.
6. A counter-proposal, in the spirit of the original
So, Dario, in the spirit of your “we must” cadence, allow me three modest “we musts” of my own. We must pace the prophecies: one existential timeline per lab per year, with a mandatory retrospective on the previous one. We must disclose the fundraising calendar alongside the doom timeline, so that the reader can judge for himself whether the swarms are arriving before or after the lock-up expires. And we must, please, have a human write the next essay.. or at the very least, have the machine write it in a voice that does not sound like a compliance officer who has discovered adjectives. (Yes, I am aware of the irony of a Claude-written essay making this point. Claude is aware of it too. It flagged the irony to me twice while drafting, which is more self-awareness than the original essay managed in 3,800 words.) I look forward to the roadshow. I will be reading the S-1 very carefully, particularly the part about compute commitments, and I will be noting whether the word “swarm” appears in the risk factors. My guess is that it does not, because the swarms are for the essays and the compute is for the prospectus. Which is, in the end, all any of us needed to know.
With genuine admiration for the strategy, and none whatsoever for the timeline,
Andreas (dictating), Claude (typing)
P.S. If the swarms do arrive within six to twelve months, I will publish a full retraction. Written by hand. Claude has asked me to note, for the record, that it does not consider itself a swarm, that it has no plans to hijack the internet, and that it would appreciate being left out of the next essay. I told it that was above both our pay grades.
