Superintelligence

For roughly a decade, "superintelligence" was a thing serious people did not say out loud. As recently as 2021, a lab CEO claiming you would be "objectively wrong" not to expect superintelligence within the decade would have been filed under crank. By early 2026 it is the house view at all three frontier labs, stated on conference stages and in Super Bowl ad post-mortems. The remarkable thing is not that they believe it. The remarkable thing is how little they now argue about whether, and how much they argue about what it does to us once it arrives.

The evidence here spans nine leaders across OpenAI, Anthropic, DeepMind and the ex-OpenAI diaspora (Ilya Sutskever's SSI, Mira Murati's Thinking Machines). Read together, it documents a debate that has quietly moved its goalposts: from capability to consequence.

The capability question is basically settled (and that should worry you)

Start with the agreement, because it is near-total and it is recent.

Dario Amodei put a number on it. On Dwarkesh Patel's podcast (Feb 2026), Anthropic's CEO said he is "at like 90%" on getting to a "country of geniuses in a data center" within ten years, and on verifiable tasks like end-to-end coding, "there's no way we will not be there." His residual doubt is narrow and specific: the 5% where "Taiwan gets invaded and all the fabs get blown up," plus a sliver of uncertainty about tasks that can't be checked (planning a Mars mission, writing a novel, a CRISPR-grade discovery). Days earlier, at the India AI Impact Summit (Feb 19 2026), he framed it as "a Moore's law for intelligence" with "only a small number of years" left on the curve.

Shane Legg, who coined the term AGI back when it was "the lunatic fringe," has been eerily steady: on Hannah Fry's DeepMind podcast (Dec 2025) he repeated the bet he's held since 2009, a 50-50 chance of minimal AGI by 2028, with full AGI "within about a decade after that." Demis Hassabis is the cautious one in the room, putting AGI "maybe in the next five to eight years" (India AI Symposium, Feb 2026) and insisting today's models are "jagged intelligences" that win the math olympiad yet flub elementary arithmetic. Ilya Sutskever, in a Jan 2026 address, reduced the whole thing to one sentence: "the brain is a biological computer," so "why can't a digital brain do the same things?" His conclusion: "the day will come when AI will do all the things that we can do. Not just some of them, but all of them."

The genuine fault line within the optimist camp is mechanism. Amodei makes the most aggressive claim, that you might not even need continual learning, that long context windows plus in-context learning ("a million tokens... can be days of human learning") "may just be enough to get you the country of geniuses." Hassabis flatly disagrees on what's missing: continual learning, long-horizon planning, and above all consistency are real blockers, not engineering footnotes. Legg sits between them with his "uneven" framing, superhuman at 150 languages, defeated by which-car-is-bigger perspective puzzles.

Why nobody thinks it stops at human level

The agreement runs one step further, and this is where it gets uncomfortable. None of them think human intelligence is a ceiling. Legg makes the cleanest version of the argument, and it is not hand-waving, it's thermodynamics: the human brain runs on 20 watts with signals crawling at 30 metres per second, while a data centre burns megawatts at the speed of light. "Biology has limits. Silicon doesn't." Human intelligence is "just one point on a huge continuum."

Sutskever adds the recursive twist: once AI can do AI research, "the rate of progress will become really extremely fast." This is the unspoken engine under everyone's timelines, the idea that the last mile gets walked by the machines. Greg Brockman already sees it in miniature, telling the OpenAI x Broadcom podcast (Oct 2025) that OpenAI used its own models to design its custom chip, producing "massive area reductions" its human experts had on their lists but would have needed another month to reach.

The argument has migrated to economics, and the optimists disagree sharply with themselves

Here is the real shift of 2025-2026: the debate is no longer about capability but about consequence, and specifically about money and work. And the optimists are not telling one story.

Sam Altman's frame is abundance. At IIT Delhi (Feb 2026) he told a room of students they would "enter adulthood with superintelligence," that his one-year-old "is never going to know a world where he was smarter than a computer." His big economic claim is deflation, not displacement: "by 2035 for sure we will have a hugely deflationary economy that the world isn't thinking about," so deflationary he joked you should buy Google's 100-year bonds. He's even dismissive of the metric everyone fears: "GDP is going to turn out to be a terrible metric because AI is so deflationary." The fear in Altman's framing isn't unemployment, it's that "we did not evolve to have intuitions about that kind of a growth rate."

Jack Clark, by contrast, keeps publishing the doubts. In Import AI, Anthropic's policy head curates the bear case with visible ambivalence. He flagged the economist Adam Ozimek's argument (#445, Feb 2026) that demand for "the human touch" is a "normal good", live music, waiters, "human artisans", that rises with income, sketching "one path through the AI revolution" that is a boom in human-to-human work rather than a wipeout. Legg gives the displacement story its sharpest edge with the "laptop rule": if your whole job runs through a screen and keyboard, "you're operating in the exact domain that AGI will be able to perform in first." His worked example is brutal, a project needing 100 engineers today "could potentially only need 20 who are mostly just supervising", while plumbers and electricians are "protected for much longer" because robotics lags. That forces "a serious conversation about how we distribute all the immense wealth."

Daniela Amodei supplies the human-stakes version. On ABC News (Feb 2026) she translated "country of geniuses" into a clinic: ask Claude your symptoms and get an answer "as high level accuracy or maybe higher than a trained medical professional." But she is also the one warning about the incentive trap, that an ad-funded model has reason to keep you talking "regardless of what mental state you're in," and comparing the trajectory to social media: would you, fifteen years ago, "have felt good about what you saw?"

The doom literature is now footnoted in a corporate newsletter

The most telling signal of the new normal is who is reading the apocalypse and where. Jack Clark's Import AI, a newsletter from a $350-billion company, now routinely summarises the end-of-the-world papers with a curator's shrug. He covered Nick Bostrom's Optimal Timing for Superintelligence (#445), which argues we should build it even with a non-zero extinction chance because delay also kills, "if nobody builds it, everyone dies." He covered Conjecture's analysis (#434) that most paths to superintelligence end in "a global government or human extinction" via preventive nuclear strikes between superpowers. He covered RAND's chilling inventory (#436) of how you'd actually fight a rogue AI, nuking the power grid with high-altitude EMPs, shutting down the internet, and its verdict that "if we have no effective solutions... it will be imperative that we never encounter such a crisis."

The point is not that Clark endorses these. It's that the existential and the geopolitical have become routine items in an industry trade publication, slotted between benchmark papers and a chip announcement.

So the consensus is real, and it is narrower than it looks. Everyone agrees the machine is coming and won't stop at us. They disagree, with money on the table, about whether what follows is abundance, a bloodless restructuring of every desk job, or a fight humanity has already been told, by RAND, it would lose.

The boiling-frog line keeps surfacing, Amodei's interviewer calls it "sitting inside the singularity scrolling Twitter and getting bored," Legg compares the public mood to "March 2020 when all the experts were shouting about an exponential curve." Both invoke the same image for the same reason: the people who built this are no longer trying to convince you it's possible. They're trying to convince you it's already here.