So was it that LLMs used to be capable of making actual jokes, or were they always this bad and I was just more impressed by the talking computer back then?
Uncensored LLMs are funnier but most comedy just falls flat in text format. Once the uncensored multimodal models start rolling out we’ll get some real laughs.
Moshi is actually pretty funny just for having a 72 IQ
I would argue that Markov chains were a better tool for comedic purposes.
Notice that in all of the examples of using Markov chains, the person would see a potential, come up with a purpose, exercise the setup, and then fill the setup with generated text.
Likewise in a random generation examples, the person would asses the results, then find what parts were actually funny, and choose to spread this parts.
LLMs output has less potentially funny results, it has less potential for unexpectedly change realistic-sounding output to absurdist output, and so as a tool it is less fit for comedic purposes.
I chuckled a bit. They are OK, if you don't get exposed to them too often. And with an LLM you can get as much exposure as you want (and all of the jokes are naturally from roughly the same probability distribution).
I don't expect too much until AI self-play learning will be made possible, so I don't get disappointed by the expected shortcomings.
It's the "impressed by the spectacle" one. I tried jokes with LLMs many times, and they're always this. Riffing on a couple of themes loosely related to what was asked. Always unfunny and uncreative.
I found some of those jokes good, definitely better than I would've ever written them. If you watch shows about comedy like say Hacks you'll see human comedians riff on stuff and a lot of the off the top jokes get discarded or improved. So Claude did fine in my book
I wonder, though, whether jokes like these could be useful to professional humorists who have to come up with gags on a deadline. From what I’ve read about monologue writing teams for late-night talk shows and the like, the writers first propose many ideas, most of which are shot down quickly and the remainder of which get tweaked and polished before being used. Some of the above jokes by Claude look to me as though they might serve as good starting points for such brainstorming. At least, they’re better than anything I could create in a short amount of time.
LLMs were never very good at directly generating original jokes, for a simple reason: writing a good joke generally starts with finding a good punchline, and then setting it up. An LLM generating token after token will first write a set-up, and then try to shoehorn a punchline into it. Prompt engineering can fairly easily work around this, but just straight-up asking an LLM for a joke never really produced good results on average.
