Dr Adam Hunt is a postdoctoral researcher at the University of Cambridge. His PhD is in evolutionary psychiatry; he chairs the Foundation for Evolution and Mental Health and hosts a podcast about the field. On Tuesday morning, 09:24 UTC, he posted a long, careful, entirely un-jokey argument for why he has turned bearish on AI — the scaling thesis is wrong, the recent models are less general, not more, and the spiky-capabilities picture never filled in.

It has done 614,052 views, 2,798 likes and — the number that tells you what kind of post it is — 2,709 bookmarks. People save essays. They like jokes.

Buried in the middle, explaining why models got so good at code while their prose got worse, is this sentence:

The recent models (and benchmarks) have been maxxing that and we’ve seen degradation on normal English use for that reason.

614K views · 2,798 likes · Metrics as captured July 29

Read it again as a grammarian

We have spent a year logging the suffix climbing the respectability ladder — the Pentagon, Elon, the Wall Street Journal, a startup founder describing his own decade. Every one of those was the same move: bolt -maxxing onto a noun. Tokenmaxxing. Thrift-maxxing. Patriotmaxxing. The suffix stayed a suffix; the joke, such as it was, lived in the seam.

Hunt does something else. There’s no noun for it to attach to. “Maxxing” is the verb, “that” is its object, and the sentence is ordinary English prose — subject, present perfect continuous, direct object. To max a benchmark. He is not borrowing a coinage. He is conjugating one.

And he spells it with two x’s.

That detail is the whole story, and it lands with a certain irony given what happened in July, when the scene’s most famous figure insisted the word takes one x and got Community Noted with Merriam-Webster as the authority. The forums lost control of their own orthography. What we’re seeing now is the far side of that: the double-x spelling has become so default that a Cambridge academic reaches for it in a technical argument about transformer generalisation, aimed at an audience of AI researchers, with no indication he’s quoting anyone.

We can’t know his intent, and we’re not going to pretend to. He may know exactly where the spelling comes from and be using it deliberately as jargon; he may have absorbed it the way everyone absorbs everything now. Both readings point the same direction. A word only becomes invisible once you no longer have to decide to use it.

The joke is that it’s the right word

Here’s what makes this more than a spotting. Hunt isn’t using maxxing as decoration. He’s using it because it’s precise, and no other English word does the job.

His claim is that labs optimised hard against a narrow, well-measured, well-rewarded target — code, where the whole job exists as training data and the benchmarks are clean — and that the gains on that spike came at the cost of general capability, with observable degradation elsewhere. Which is to say: they picked one legible stat, drove it as far as it would go, and let the unmeasured stuff rot.

That is the exact critique the looksmaxxing scene has absorbed for a decade, aimed at the machines. It’s what the forums mean when they argue about whether a canthal tilt is worth what it costs you. It’s the failure mode our own coverage keeps circling: optimise the number you can see, degrade the person you can’t.

English already had words for over-optimisation — Goodhart’s law, overfitting, teaching to the test. Hunt had all of them available. He reached for the teenage forum word instead, because the forum word carries something the others don’t: the sense of a voluntary project pursued past the point of sense, by someone who thinks they’re winning.

The eye chart on a clinic wall is the honest version of the problem. You can memorise the bottom line and score 20/10 without seeing one bit better. The chart hasn’t measured your vision since the moment you decided to beat it. What the forums worked out about faces, and what an evolutionary psychiatrist is now saying about language models, is the same sentence with a different subject: the number goes up, and something you weren’t counting quietly goes down.

The suffix has stopped needing a host. It’s a verb now.