Humor , in general, relies on the presenter holding an accurate theory of mind of the audience, then subverting expectations in a way that mocks convention, is ironic, Or speaks a widely known but little acknowledged truth.

LLMs don’t have anything significant in terms of theory of mind, although that could probably be trained in with significant effect to the conversational capabilities of LLMs.

I'm not sure how you could train theory of mind into the capabilities of just LLMs. It seems they would need something more complex than attention based completion/prediction to model it efficiently.

Edit: By more complex I mean more deliberate / directed. Predicting the next word(s) is a concrete goal -- "training" a theory of mind sounds like "training" sentience. Sure... But how?

I’d say by training on a lot of contextual situation: what is person thinking pairs. It’s stuff that will be grossly underrepresented in content simply because it is “obvious “ and normally not mentioned. That would probably be enough to bring audience awareness into the worldstate of the token-space.

In theory, could you teach an LLM to simulate "humor" through pattern memorization?

I suppose to come up with an original joke you need to yourself be able to find something to be funny.

I don't know. Humor is a very high dimensional space in itself and different to each individual. I would be surprised if any LLM could be trained to be a consistently well reviewed comedian. "Humor" is just such a nebulous metric.

As far as I can tell, their way of generating humor is much more naive, rehashing comedic patterns from pre-training. You're probably right that genuine, original humor would require a deep enough understanding of human expectations to be able to subvert them.

While I agree that the comedian plays with the expectations of the audience and therefore has to have a rough idea of those expectations (or just conventions?), calling it an "accurate theory of mind" is a bit of a stretch.

Sufficiently accurate then.

While I kind of agree, I'd also like to point out that Markov chains can be very funny when fed the right material and tuned well.

Fun usually involves a bit of transgression, rebellion or surprise, all of which goes against the absurd subservience and most average character of common LLM:s. I can see why the attempt was fun to do, but the results in TFA aren't good enough that I feel like I want to share my Github details to see more.