The Exponential Time Hypothesis Falls: Reading the AI Claim in Theoretical Computer Science Through Erik Larson

Peter Jonsson, a professor at Linköping University, has work scheduled for presentation in two weeks at one of the most prestigious conferences in theoretical computer science. As of Monday, he says, much of it is "dead." OpenAI and Anthropic both claimed this week that their AI models had disproved the exponential time hypothesis, a 27-year-old assumption that underpins large stretches of complexity theory and algorithm design. Jonsson's results rested on it, and a theorem that begins "assuming ETH" loses its footing when ETH falls. The coverage has settled on a word for this: decades of research made "irrelevant." That word is strong enough to send any tenured mathematician back for a second coffee. It also hides the question that decides what the story means. When an AI system knocks out a long-standing assumption, which parts of the work did the machine do, and which parts did people do around it?

Most reports so far give you the verdict and skip the method. You learn that something large happened and that a respected researcher is counting his losses. You learn very little about how to weigh a claim like this one. Which steps did the models perform, and which did people perform before and after? Who picked ETH as a target, and who has checked the disproof line by line? Without those answers, "irrelevant" floats free. The likelier meaning is narrower than the headline: work that assumed ETH has lost its premise, while the techniques built along the way probably hold up as well as they did last week. To judge claims like this one, you need a working account of what machines do when they appear to reason.

Erik J. Larson's The Myth of Artificial Intelligence supplies that account, and it comes from inside the trade. Larson is a computer scientist who has worked at the forefront of natural language processing, so his skepticism was formed at the workbench. His argument starts with inference. Current AI, he says, runs on induction: it crunches data sets and predicts outcomes from patterns it has already seen. People who think well make conjectures, informed by context and experience, and they jump to an explanation before the evidence is complete.

The plain-language version is easy to feel. A system trained on a million engine recordings can flag an unusual noise with impressive accuracy. A good mechanic hears the same rattle, notices the new roof rack, and guesses the cause before running a single test. The guess might be wrong, and it still opens a useful line of inquiry. Larson's claim is that nobody knows how to program that move, and that it sits at the heart of common sense. On his account, piling on more data leaves the distance where it is, because the jump never depended on volume.

Larson traces the confusion to Alan Turing, after whom enthusiasts began treating artificial and human intelligence as the same kind of thing measured on one scale. Theoretical computer science descends from Turing too, and its whole purpose is to state precisely what mechanical procedures can do and how fast they can do it. ETH is exactly that sort of statement, a wager that certain problems demand exponential time. Larson brings a similar appetite for precision to his own field's boosterism, measuring how far today's systems sit from the superintelligence that futurists keep putting on the calendar.

I part ways with him on how cleanly he draws the line between induction and conjecture, and this week's news presses hard on that line. A disproof that strong theorists missed for 27 years looks a lot like a conjecture from the outside, whatever is happening inside the model. Larson's framing makes it easy to file such results under pattern-matching, and that reflex can harden into a faith of its own. The blurbs make things worse. Peter Thiel's endorsement, praising human intelligence as "our most irreplaceable resource," pushes a careful argument about inference toward manifesto.

The book earns its place anyway, through the questions it trains you to ask. Who framed the problem? Who recognized that the output mattered, and who proved it? Larson would call those the conjecture-shaped steps. My bet is that the published accounts of the ETH result will show people doing most of them, with the models running a fast, wide search in the middle. If the models turn out to have done more, Larson's book becomes the clearest statement of exactly which belief got overturned, and that is still worth having on the shelf.

If you pick up Larson, give most of your attention to his treatment of inference, since it supplies the vocabulary this week's coverage lacks. Then read the write-ups of the ETH disproof as they appear, even if the proofs are beyond you. Introductions and acknowledgments usually say who did what, and acknowledgments are where academics hide their most honest sentences. You may finish Larson convinced he is too sure about where machines must stop. Keep that objection, because it hands you a precise claim to hold against a real result. His book has placed its bet in plain sight, and the next few months of checking in theoretical computer science will show how well it was placed.