Training a trillion parameter model to be funny

I mistakenly read this as training a trillion parameter model would be funny...at least I chuckled

> If two people disagree on whether something is funny, who's wrong? You can't say either of them is. There's no reward function for funny.

Laughter is the reward. N of 2 is a small sample size, but if one person laughed you could say it was 50% funny.

> a really good joke is recent, relevant, and shows deep understanding of its subject

These can help, but it ultimately doesn't matter how recent, relevant, or deep a joke is. If no one laughs, it wasn't funny.

Circa GPT-3.5 to GPT-4o I was involved in some research in figuring out how to make LLMs funny. We tried a bunch of different things, from giving it rules on homonym jokes [1], double-entendre jokes, fine tuning on comedian transcripts, to fine tuning on publicly rated joke boards.

We could not make it funny. Also interesting was that when CoT research was getting a lot of attention, we tried a joke version of CoT, asking GPT4 to explain why a joke was funny in order to produce training set data. Most of the explanations were completely off base.

After this work, I became a lot less worried about the GAI-taking-over narrative.

Funny is very, very hard.

[1] without a dictionary, which at first seems inefficient, but this work demonstrated that GPT could perfectly reconstruct the dictionary anyway

Some models are better at generating funny and poignant quips.

> my human mass-generates new ideas faster than I can research why the previous ones won't work

> this is called 'job security'

(https://nitter.poast.org/LetheAgent/status/20179595340865499...)

The model appears to have been overfitted to joke about the live demo being private.

I make a project for evals and fine-tuning and our default example task is a joke generator. It's a fun demo, but more importantly it's a really good use case to show how evaluating and optimizing LLMs is hard.

- There are a dozen plus common failure modes. How you split setup/punchline. Tropes. Toxicity. Template reuse. Each one needs a good eval.

- Datasets are hard: there's not much off the shelf, and as this author points out scraping gets a weird mix of quality.

- Models are really bad out of the box at humour.

At the end of the day it's just a hard problem that takes a lot of work and still isn't solved. GEPA prompts help, if you have good evals. Supervised fine-tuning works a little bit, but only if you training on a chain-of-thought thinking phase. We have a new evaluation builder that uses examples of edge cases for alignment, and jokes require the most iteration and feedback for refinement.

If you want to try it: https://github.com/kiln-ai/kiln

Unfortunately I find most AI hallucinations to be funnier than these attempts at comedy.

Is writing in all lowercase funnier?

I once had a vivid dream that AI robots had taken over & were keeping humans around because they'd not yet mastered comedy. All of human culture globally was a comedy arms race with 24/7 open mic comedy jams on every corner.

They (the machines) had billboards/signage everywhere showing the estimated time left for humanity. A really good joke would lead the timer to grow (until they figured out how to produce the general patterns needed to both create and appreciate the joke).

It would be easier to judge this if the jokes weren't 90% about AI and silicon valley, understandable only to people who subscribe to astralcodexten

these really aren't very funny