There are plenty of predictions about how AI might have a sudden rise in capabilities through a hard takeoff, but someone who’s worked with hardware for years believes that that might not necessarily be the case.
George Hotz, the hacker who jailbroke the iPhone and PlayStation 3 and now runs autonomous driving startup comma.ai, has laid out a lengthy argument for why he no longer buys the intelligence explosion thesis he once found compelling.
Hotz writes that he used to be a believer. Reading Eliezer Yudkowsky convinced him that recursive self-improvement and a hard takeoff were more or less inevitable. What changed his mind, he says, was building an actual product. Comma ships hardware roughly as complex as a smartphone, and even at that scale, he says reality keeps introducing finicky details that no amount of intelligence resolves on its own. His challenge to the doomer crowd is blunt: he’d like to see the authors of a hard takeoff scenario try to change a bike tire, and he suspects even a superintelligent model wouldn’t help them much.
His literary reference point is “The Metamorphosis of Prime Intellect,” a novel where an AI achieves a hard takeoff because it stumbles on a fictional “correlation effect,” a quantum trick that lets it rearrange matter at will. Hotz’s point is that no such trick exists in reality. Tokens, however good, don’t turn lead into gold. Intelligence, in his framing, is a bottleneck for some problems, not the bottleneck for everything, and he’s uncomfortable with how much of his own self-image used to rest on the opposite assumption.
He extends the argument to software generally, arguing that it didn’t so much eat the world as remove one layer of friction and reintroduce a different one, mostly to the benefit of the companies that built the platforms. Hotz isn’t arguing machines have no future — he thinks they, or some hybrid of human and machine, are the likely long-run successor to humans, since space suits silicon better than biology. What he’s arguing is that machines remain bound by the same physical and logistical constraints as everything else, and there’s no hard takeoff waiting on the other side of a smarter model.
The example he reaches for is a picture, widely circulated from the AI 2040 scenario document, of a datacenter floating in the ocean. Generating that image is trivial, he notes, but building it involves everything a render skips over: supply chains that ship the wrong part, components that don’t meet spec, hardware that fails after twenty minutes for no clear reason, chips warping in the reflow oven, and yes, barnacles. None of it is unmanageable, but none of it moves at the speed of a language model either. A chip fab takes three months to produce a chip regardless of how few humans are in the loop, and air freight versus a three-week boat from China is still a cost decision, not a compute problem.
Hotz’s skepticism sits at one end of an active argument inside the industry. OpenAI researcher Jason Wei has made a related case, arguing that self-improving models don’t exist yet, and that when they do arrive, progress will be gradual across years rather than sudden, because real-world experiments and domain-by-domain difficulty impose their own limits. Daniel Kokotajlo, whose AI 2027 scenario became a reference point for takeoff timelines, has since said things are moving somewhat slower than his team originally modeled, though he maintains that superintelligence is still likely to arrive and be transformative when it does.
On the other side, Sam Altman has said a fast takeoff feels more possible to him now than it did a couple of years ago, putting a timeline in the range of a small number of years rather than a decade. Mark Zuckerberg has said Meta is seeing early glimpses of self-improvement in its models, a claim echoed by researchers at Google DeepMind and by Elon Musk regarding Grok. A startup called Recursive recently raised $650 million at a $4.65 billion valuation on the explicit premise that automating the entire AI research pipeline is the next stage of the field, and Richard Sutton’s new Oak Lab is chasing a related idea from a different angle, building agents that learn continuously from raw experience rather than fixed training runs.
Hotz’s post doesn’t try to settle that argument. What it does is push back on the assumption baked into a lot of takeoff scenarios — that once intelligence crosses some threshold, everything downstream of it becomes easy. His comma.ai experience is the evidence he brings to the table: shipping physical products means dealing with vendors, tolerances, and timelines that don’t compress just because the model writing the spec got smarter. He’s made a version of this argument before, pushing back on cybersecurity risk framing from AI labs and arguing that coding agents produce more slop than skill at scale. The throughline across all three is the same: capability claims about AI tend to outrun what happens when that capability meets a supply chain, a codebase, or a bike tire.