AGI is not coming
2026
Any honest version of this argument must begin with a concession, because the dishonest versions begin by refusing one. Machines now do things that were, until recently, canonical examples of what they could never do. In 2024, DeepMind’s AlphaProof reached silver-medal standard on International Mathematical Olympiad problems; a year later, systems from both Google and OpenAI performed at gold-medal level. Research mathematics has not been spared: Peter Scholze’s Liquid Tensor Experiment — a proof he described as too intricate to fully trust his own checking of — was formalised in Lean by 2022. Benchmarks built to be unreachable are reached with monotonous regularity. Anyone whose scepticism about artificial general intelligence depends on machines failing at mathematics has already lost the argument.
Mine does not. The claim is not that capabilities will plateau. It is that what accumulates, however fast, is the wrong kind of thing.
Consider what has actually been automated. Search, verification, recombination at scale — the algorithmic pole of thought. This should surprise nobody: these were always the mechanisable parts, which is precisely why mathematicians have long regarded them as the last step rather than the essential one. Poincaré’s account of discovering the Fuchsian functions is instructive not because it is charming but because it is structural. The insight arrived as he stepped onto an omnibus in Coutances, mid-conversation, thinking about nothing relevant — and it arrived complete, with certainty attached, verified only later at his desk. Hadamard, surveying his colleagues half a century on, found the pattern general: preparation, incubation, illumination, verification. Machines have industrialised the first and last stages. The middle two are not slower or weaker in machines; they are not located anywhere in them, because they are not stages of an algorithm at all. They are episodes in the life of a subject.
This is where the argument becomes phenomenological, and where I am obliged to be careful, because phenomenological arguments about AI have a poor reputation — usually deserved, when they are deployed as proofs. Penrose held that mathematical insight is non-algorithmic; his Gödelian formalisation of that intuition is widely and, I think, rightly contested. But the intuition itself does not stand or fall with the formalisation. The phenomenological observation — descended from Merleau-Ponty, sharpened against early AI by Hubert Dreyfus — is that intelligence, as actually lived, is not the execution of procedures over representations. It is the having of a world: a horizon of relevance against which some things show up as mattering, as anomalous, as interesting, before any criterion for interest has been articulated. A large language model has, in the strict sense, no such horizon. It has a corpus. What it returns to us is a reconstruction of the recorded surface of human judgment — statistically magnificent, phenomenologically hollow. Baudrillard would have found the whole situation unsurprising.
The objection writes itself: every generation of this argument has ended with the sceptic explaining why the newest system doesn’t really understand, while the systems went on to do whatever understanding was supposed to be required for. If the argument keeps retreating, isn’t the retreat the refutation? But notice what the milestones that keep falling have in common: each one is a task — bounded, scoreboarded, specifiable in advance. The pattern is not that machines keep acquiring pieces of mind; it is that we keep discovering how much of what we called mind was, all along, task. That discovery is genuinely humbling, and it is the true intellectual content of the past decade of AI. It does not touch the remainder. The remainder is not a harder task; it is not a task at all. Deciding what is worth proving, noticing that a definition is wrong before one can say why, feeling that a proof is correct but ugly and that the ugliness is a clue — Gadamer’s observation that discovery is a choice governed by a sense of beauty names a capacity that has no loss function, because it is what the setting of loss functions is for.
Hence a prediction, so that this position risks something. Machines will keep winning wherever there is a scoreboard, and they will remain dependent wherever the work is deciding what the scoreboard should be. The dependence will be easy to overlook, because a human will always quietly supply the judgment — choosing the problem, curating the corpus, deciding which of ten thousand generated conjectures deserves attention — and the machine will supply the spectacle. General intelligence, if the word “general” is doing any work, is the former capacity. It is not late. It is not en route. It is not the kind of thing that arrives by accumulation.
None of this is pessimism about the machines; it is a division of labour. When the solving of well-posed problems becomes cheap, the binding constraint shifts from solving to selection — and selection is ours. The interesting future is not the one where mathematicians are replaced, but the one where mathematics reorganises itself around the question machines cannot ask: what should we want to know? I am building my corner of that future now, and I have never felt less obsolete.