I do wonder how much of that fat tail is lost in simulation, and what the downstream effects are. Model autophagy disorder is not something you’d want for models making life or death decisions.

Imagine where we would be today if an overpowered Nazi Germany did not have US and USSR to counterbalance. Or just watch Man in the High Castle.

And in that world, the quality of the weapon is how cheap it runs for how good it is.

I won’t be surprised if nation states eventually train and host their own models. Heck, some LLM shops seem to be betting on that, e.g. Mistral and AlephAlpha with “European LLMs”.

We’re approaching an era of memetic warfare where LLMs are the weapons. We’re not there yet — the values espoused by Chinese LLMs aren’t all that different from American ones — but that’s for now.

But once LLMs become our primary interface with the outside world, it’s bound to happen.

Looks nice! Some FastAPI endpoints + a docker image should help adoption. :)

This is fascinating work, congratulations!

Question: the point that architectural constraints (locality + equivariance) are sufficient is well demonstrated.

But do you think it is necessary? I.e. would you expect a diffusion transformer to learn these constraints?

In other words, code 1 is more “multi-agent” than others.

What do I mean when I say “agent”? A system that we’d like to abstract away like a black box (Rovelli’s definition). Of that, I count three in code 1, and 1 in both codes 2 and 3.

Because code 1 is most explicit about the structure of the computational graph. :)

bsky.app/profile/nasi...

If I were to force an answer, I’d say code 1 (prompt chaining) has more agent energy than the others.

Paper here:

Have the token-level LLM predict “concept tokens”. The hidden states for these tokens go in to an adapter, and out come concept representations. Concept tokens attend to previous concept tokens, and perhaps also the span between itself and the previous concept token.

See paper for more.

+ Alejandro is at NeurIPS and figuring out where to do his PhD. Wink wink nudge nudge.

www.linkedin.com/in/alejandro...

Results? Good stuff.

🧵⤵️

The idea is to model two things:
(a) if concepts fit together to make good art, and
(b) if people have already thought about that combination of concepts (“cognitive availability”).

Seek out the combos for which (a) is true but (b) isn’t, and ask a text-to-image model to render that.

🧵⤵️