Foom 2027
It’s hard to exaggerate the gravity of the situation we find ourselves in right now. The most worrying news of recent weeks: the pace of Anthropic’s progress has reportedly doubled over the past year, and according to the experts at the independent firm METR1 (an analysis confirmed by other experts, including economists who have gone over the numbers with a fine-tooth comb), a very large part of the acceleration comes from the use of the AIs themselves. Given the opacity of Anthropic, which redacted whole sections of METR’s report, the experts don’t have all the evidence they would need to reach rock-solid conclusions, but a consensus is emerging: research at Anthropic is more and more automated, and this automation is pushing up the curve of the models’ progress.
We have entered the era of recursive self-improvement. The loop is not yet closed—human researchers still play an important role in AI research—but the impetus has been given and the trajectory is clear.
I was saying it last week: the researchers at the AI Futures Project—authors of the scenarios AI 2027 and AI 2040, who have gained a great deal of credibility over the years given how accurate their predictions have been—are aghast. I largely defer to them, and to Daniel Kokotajlo in particular, to gauge the intensity of the moment. On X, Kokotajlo, normally a very composed person, had harsh words for Anthropic: “@AnthropicAI: what the fuck?” “I am scared, sad, and angry.”2 We should all be talking like Daniel Kokotajlo right now. After Coxon’s viral tweet in September, I had a burst of hope. The cycle of reports and interviews around the question “Is AI going to kill us all?” lasted a few weeks, then ran out of steam. Some experts swore there was nothing to fear. It seems the mainstream has digested the information, concluded that nothing could be concluded, and moved on to the next call.
Kokotajlo and all his colleagues at the AI Futures Project now predict—if nothing is done, legislatively, to stop or slow down the frontier companies—that we will reach superintelligence in 2027. I repeat: SUPERINTELLIGENCE IN 2027. Yes, that’s next year. Superintelligence is an artificial intelligence that surpasses all human experts in every field, including physics, cyber, political strategy, engineering, etc. Those who have an optimistic view of artificial superintelligence surely haven’t turned the question over to explore it from every angle. I will say in passing that I’m a little fed up with this optimist/pessimist divide, as if it were a choice. Should the Indigenous peoples have been “optimistic” as they watched the conquistadors’ ships arrive in the sixteenth century? “Oh, come on, don’t be such a killjoy! Aren’t they beautiful, those boats!” The AIs we have right now are unaligned. They are ruthless in the pursuit of their goals. They are alien entities, not human in the least. Do we have the slightest rational reason to think this will go well for us? Stephen Hawking warned us:3 if extraterrestrials land on Earth, or if we create an artificial intelligence that surpasses us, it is very likely that we will be replaced.
When superintelligence arrives, that won’t mean the immediate end of humanity. It could take its time to establish its grip. In any case, it has already begun with today’s AIs: we ask for nothing better than to hand them the keys to the military, intelligence, surveillance, etc. Not to mention their capabilities in medicine and biology, which transfer so well to biological weapons…
Recursive self-improvement is paving the way for what is called the intelligence explosion. From an acceleration factor of two, we can go to a factor of 10, of 100, or of 1,000 in a few months, a few weeks, or a few days, and see the emergence of AIs we don’t understand at all, created by other AIs that were themselves created by other AIs, and so on. This is what Eliezer Yudkowsky called foom, after the sound a rocket makes as it takes off. We are at the first crackles of the match, and what are the media talking about?
I no longer know which rooftop to shout from. I sometimes think of selling everything I own, quitting work, and enjoying the years to come (the years left). I’ve always dreamed of going to Japan…
A colleague tells me: “If you think the end of the world is coming, why don’t you go off on a sailboat?” That’s the very last thing I’d want to do, prone to seasickness as I am. But traveling indefinitely, maybe.
Actually, I’m not absolutely certain the end is coming, of course. My P(Doom) is about 60%: that leaves a 40% probability of non-apocalypse. Of that, I wouldn’t assign more than 5% to “It’s going to be all right because superintelligence will wish us well.” The remaining 35% rests on political decisions. As I see it, that’s the only thing that can get us out of this mess. A real catastrophe—a large-scale failure of the power grid or the Internet, a massive loss of human life—sparking a global awakening sparking relentless political pressure sparking swift, international political action... It’s that kind of “recursive self-improvement” of the political situation that could save us. Otherwise, we will all board together (including your parents, your children, your babies, those yet to be born, and the generations to come) the foom rocket, cobbled together by mad scientists who don’t understand their own invention.
Nakhon Ratchasima, Thailand
METR estimates that Anthropic’s progress is moving about twice as fast as before, and attributes nearly 80% of that acceleration to AI itself. See “METR Says Anthropic’s AI Development Pace Has Doubled,” Dataconomy, November 6, 2026; “METR finds Anthropic may have crossed its own AI acceleration warning threshold back in July,” The Decoder, November 2026. ↩︎
Daniel Kokotajlo, on X, November 2026; Anisha Sircar, “AI Futures Project Pulls Its ‘Intelligence Explosion’ Forecast Into 2027,” Forbes, November 12, 2026. ↩︎
Rory Cellan-Jones, “Stephen Hawking warns artificial intelligence could end mankind,” BBC News, December 2, 2014. ↩︎