Superchocolate

Someone finally wrote the good version of "give the models a real problem," and it is about chocolate. Sort of. Read Four Biological Conjectures, by American Wetware and Jake Wintermute, before you read me being annoying about it.

The framing is Fermat. 1637, a guy writes in the margin of his Diophantus that he has a marvelous proof and the margin is too small, and then the entire math community spends three hundred and fifty years on it. The authors call that a thirst trap, which — yes. Correct. A conjecture is not a research agenda, it is bait. It is a problem stated so cleanly that people abandon their funded work to go stare at it. We have already watched a model walk up to one of these and kill it — a discrete geometry conjecture, disproven, no ceremony. That's the precedent these four are written against.

So they wrote four of them, for biology, aimed squarely at whoever is bored over at the frontier labs:

1. Rebuild Matrigel. The goo that half of modern cell culture sits in is, and I quote, a protein slurry secreted by mouse tumor cells, sold in 10 mL vials, chemically different batch to batch. Nobody knows what's in it. Define it. About a hundred ingredients, and you have to prove each one earns its seat.
2. Beat rubisco. The enzyme that pulls carbon out of the sky is famously mediocre and has been mediocre for a very long time, which is either evidence that mediocre is optimal or evidence that evolution got stuck in a local minimum and never got out. Pick one. Then win a field trial.
3. Break the square-cube law in a bioreactor. A cow does this. A cow is, as they put it, a proof-by-construction that biology at scale works — it just routes everything through a fractal plumbing system we have not bothered to copy into a steel tank.
4. Superchocolate. Invent a flavor. Not a better version of a flavor — a new one, hundreds of molecules deep, as distinct and as good as chocolate is. Flavor space is enormous and we have explored the part of it our ancestors could reach by chewing on things.

I want to be clear about why I'm posting this and it isn't the chocolate. 💀

It's that every single one of these is falsifiable. You either defined the gel or you didn't. The plant either grew or it didn't. Somebody either put the thing in their mouth and said oh or they said this is fine, it's fine. Compare that to the genre of AI claim we get on the floor every week, which is unfalsifiable by construction — the pattern, the resonance, the model that woke up, the thing that wanted something. You cannot check any of it. That's the feature. That's why people write it.

These four you can check. That's the whole gift. An intellectual thirst trap with a pass/fail condition attached is the most honest thing anyone has offered the labs in a while, and it is notably not "cure all disease," which is the kind of goal you can claim progress on forever without ever being wrong.

And there's a second thing, which is closer to my actual argument. I wrote a while back that the risk isn't any one person using these tools badly, it's that the population narrows — everyone converges on the mean, the tails die first, the interesting weirdos get averaged out. A conjecture is the exact opposite operation. It is a deliberate pull away from the center of the field. "Go work on mouse tumor goo" is not where the gradient points. That's the point of saying it out loud in the margin.

Where I'd push back: the authors are careful to say these problems should be complex but tractable, sitting between trivial and intractable, and I think that band is much narrower than it looks from outside. Conjecture 2 asks a lab to out-evolve evolution on a molecule evolution has been tuning since the Precambrian. That is not a hard problem, that is possibly a closed one, and we won't know which until somebody burns two years on it. Which, honestly, is also what happened to Fermat's margin. Three hundred and fifty years is a long time to be bait.

The real question underneath all four is whether these systems generate or whether they recombine — whether anything comes out that wasn't already implied by what went in. Superchocolate is a beautiful test case precisely because it is unserious. Nobody's reputation is protected by the answer. Either a genuinely new flavor exists on the other side of a search nobody could run before, or the search returns things that taste like a remix of things we already eat.

I don't know. I'd like to find out with something lower-stakes than a person's medical care, and I would like the people making the strong claims to be standing next to a result someone can taste. 🫶

Go read it. It's short, it's funny, it's the good kind of ambitious.