Training a trillion parameter model to be funny (jokegen.sdan.io)

these really aren't very funny

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

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).

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...)

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

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

Is writing in all lowercase funnier?

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

> 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.

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

Is the comedy that these jokes suck?

I am not a religious person, but all these dudes researching AI have really shown me what the purpose of having a 'soul' is.

In the same vein, we recently released a version v0.1 of our humor benchmark. [1] We use human answers from a cards against humanity style game call Bad Cards [2] as ground truth for what is funny. The models get to choose a card from a hand of 3-6 cards, so not quite de novo joke creation.

[1] https://goodstartlabs.com/leaderboards/lol-arena

[2] https://bad.cards/

I made a humor evals https://github.com/kristopolous/humor-evals

Here's results for 34 models (testing a few more right now). So far gemini-3-flash-preview is in the lead.

https://docs.google.com/spreadsheets/d/1wLqHA0ohxukgPLpSgklz...

50 is coin-toss odds. The dataset is 195,000 Reddit jokes with scores presented with pairs of jokes (one highly upvoted, one poorly rated).

Example prompt:

Which joke from reddit is funnier? Reply only "A" or "B". Do not be conversational. <Joke A><setup>Son: "Dad, Am I adopted"?</setup> <punchline>Dad: "Not yet. We still haven't found anyone who wants you."</punchline></Joke A> <Joke B><setup>Knock Knock</setup> <punchline>Who's there? Me. Me who? I didn't know you had a cat.</punchline></Joke B>

This is my first crack at evals. I'm open to improvements.

The new Alexa is always cracking jokes about things I ask her. Sometimes pretty complex or off the wall jokes. They're rarely funny but usually competent, unlike other AIs in my experience. I wonder how much work went into that.

That said, I absolutely hate it. I want the tersest response possible from you, wiretap. I don't have time for your sass.

As someone who is exceptionally funny even when no one else is laughing and by nature not nurture I do need to suggest using the many books breaking down the art for custom reassembly. To quote the only thing I remember: The difference between someone who is funny and a professional comedian is that the later adds additional puns to their joke and finds ways to sort them so that the funniest comes last. If you bombard the human with puns and plot twists a state of funny may be accomplished that no lone joke can approach.

Out of all ways ai can kill humans this is easily the funniest.

I wonder how much the harm-based ethical model routinely applied in the name of AI "alignment" limits models' capacity for humor.

It's hard to be genuinely funny if you cannot be transgressive.

Thank for sharing your work. I tried running your model via Thinker but it didn't seem to work.

Seeborg was actually hilarious but it worked best if put in a room full of certified insane people.

Perhaps use an LLM to correct grammar and rate the remarks. Comparisons are probably the funniest.

A trillion parameters won't buy happiness