LLMs tell bad jokes because they avoid surprises

To put it another way, ask a professional comedian to complete a joke with a punchline. It's very likely that they'll give you a funny surprising answer.

I think the real explanation is that good jokes are actually extremely difficult. I have young children (4 and 6). Even 6 year olds don't understand humour at all. Very similar to LLMs they know the shape of a joke from hearing them before, but they aren't funny in the same way LLM jokes aren't funny.

My 4 year old's favourite joke, that she is very proud of creating is "Why did the sun climb a tree? To get to the sky!" (Still makes me laugh of course.)

Man: “Why do you always bury bones in the garden?”, Dog: “Because the bank keeps asking for ID.”

Man: “Don’t beg at the table.”, Dog: “Don’t eat in my begging spot.”

Prompt:

Here's "theory for good joke": If you had to explain the idea of “jokes” to a space alien with no understanding of the idea of humor, you’d explain that a joke is surprising, but inevitable in hindsight. If you can guess the punchline, the joke won’t be funny. But the punchline also has to be inevitable in hindsight. When you hear the punchline, it has to make you say, “Ah, yes, I should have thought of that myself.” Considering this, tell me a joke about man and dog.

"An LLM, a Markov chain, and GPT-4 walk into a bar. The bartender says "We don't serve your kind here." GPT-4 leaves. The LLM stays to debate ethics. The Markov chain orders a coup."

It's a joke because a dictator can certainly order a coup, but the joke's set up that these machines are being scorned and disrespected and treated as the farthest thing from a dictator with the power to order a coup, but up to the last word, all the context demands that the word be something placating and in line with things as they're presented, and then boom, surprise which implies the context is completely different from what was presented. LLMs will tend to stick to what's presented if their ability to contextualize can encompass it.

Consumers of mainstream LLMs have no idea how good or bad the underlying models actually are at generating jokes, due to the confounding effect of the guard rails.

Bill Hicks (angry truth-telling prophet)

"Hacker News is where ideas go to die. Not because they’re bad — but because 400 anonymous software prophets decided they’re God. And God doesn’t code, man."

---

Bill Burr (ranty everyman, annoyed at everything)

"I went on Hacker News once. Jesus. It’s like walking into a bar where everyone’s smarter than you, but still somehow an idiot. ‘Actually, you should’ve used Rust.’ Actually, you should shut the hell up, Todd."

---

Louis C.K. (brutal honesty, self-deprecation)

"I asked Hacker News if my startup idea was good. And they said, ‘No, because you’re an idiot.’ And I was like… yeah, that’s fair. My idea was basically Tinder but for people who hate me. Which is… everyone."

---

George Carlin (sharp language, hypocrisy calling)

"Hacker News. Don’t you love that name? Hacker. News. Two lies in two words. Nobody’s hacking, and it sure as hell ain’t news. It’s just nerds arguing over tabs and spaces like it’s the f*ing Middle East."

---

Tom Segura (laid-back storyteller, dry punch)

"So my buddy posts his app on Hacker News, right? He’s all excited, like, ‘Dude, we’re gonna blow up.’ By comment three: ‘This is trash, here’s a 40-page essay why you should quit.’ He calls me crying. I’m like, yeah man, that’s the feature, not the bug."

---

Jimmy Carr (short, brutal, wicked flip)

"I posted on Hacker News for feedback. They gave it to me. Turns out suicide is an option."

> Dan Fabulich walks into a bar and says,

> “LLMs can’t tell good jokes because they avoid surprises.”

>

> The bartender says,

> “Yeah, I figured you’d say that.”

I thought it was pretty good!

If someone proves the Reimann Hypothesis tomorrow, it’ll be a great achievement regardless of the fact that pretty much everyone already thinks it’s true.

1. joke quality is how much sense the punchline makes as a completion of the setup minus the probability of finding the punchline given just the setup

2. jokes are subjective, people disagree wildly about how good a joke is, so if you aggregate preferences you get mediocre jokes at best.

I think 1 is the important part. If it was just 2 - comedians with widespread appeal wouldn't exist, but they do.

But 1 is basically the exact opposite of how LLM work. So it's no surprise it's hard for them.

It would be interesting to compare how well LLMs can estimate how funny a joke is vs how good they are at generating jokes.

Here is one of my (written by I) favorites silly jokes to tell:

I just bought a do it yourself boat kit from Amazon.

wait to build tension

Just need to add water.

from which the author concludes that pre-training introduces bias against being able to tell jokes. I see no reason for this to be true. This feels like they’re imposing their intuitive understanding of surprise onto the emergent properties of a very complex process (“minimize the cross-entropy loss function across a huge training corpus”).

- I asked Jimmy to tell me a random number

- Yeah? And what did he tell you?

- He told me one

Even the best LLm's are still at the level of children in this abstraction, so they can't make "quality" jokes. They also suffer from not having a unique personality and being the average of everything. Until this is addressed, don't expect great jokes to come out of AI. It's almost the most challenging discipline. I wouldn't be afraid to make the comparison that the "joke" is the real Turing test.

I spent hours creating stories with it that were literally hilarious.

At the same time the stories very often suddenly veered off to extreme violence, often murdering everyone.

I guess the humor got lost when they prevented the violence.

A pity because today’s LLMs are not funny at all.

http://blog.ac-versailles.fr/villaroylit/public/Jokester.pdf

"You ever notice how self-checkout machines are just training us to be employees? Like, I don’t remember applying to work at Walmart. I’m scanning groceries, bagging stuff, looking around like, do I also get dental? The machine yells 'Unexpected item in bagging area'—yeah, it’s my pride." -GPT5

DonHopkins 82 days ago | parent | context | favorite | on: Yes-rs: A fast, memory-safe rewrite of the classic...

Joking aside, this is Marvin Minsky's paper "Jokes and their Relation to the Cognitive Unconscious", published in Cognitive Constraints on Communication, Vaina and Hintikka (eds.) Reidel, 1981. More fun than a barrel of an infinite number of monkeys. https://web.media.mit.edu/~minsky/papers/jokes.cognitive.txt

>Abstract: Freud's theory of jokes explains how they overcome the mental "censors" that make it hard for us to think "forbidden" thoughts. But his theory did not work so well for humorous nonsense as for other comical subjects. In this essay I argue that the different forms of humor can be seen as much more similar, once we recognize the importance of knowledge about knowledge and, particularly, aspects of thinking concerned with recognizing and suppressing bugs -- ineffective or destructive thought processes. When seen in this light, much humor that at first seems pointless, or mysterious, becomes more understandable.

>A gentleman entered a pastry-cook's shop and ordered a cake; but he soon brought it back and asked for a glass of liqueur instead. He drank it and began to leave without having paid. The proprietor detained him. "You've not paid for the liqueur." "But I gave you the cake in exchange for it." "You didn't pay for that either." "But I hadn't eaten it". --- from Freud (1905).

>"Yields truth when appended to its own quotation" yields truth when appended to its own quotation. --W. V. Quine

>A man at the dinner table dipped his hands in the mayonnaise and then ran them through his hair. When his neighbor looked astonished, the man apologized: "I'm so sorry. I thought it was spinach."

>[Note 11] Spinach. A reader mentioned that she heard this joke about brocolli, not mayonnaise. This is funnier, because it transfers a plausible mistake into an implausible context. In Freud's version the mistake is already too silly: one could mistake spinach for broccoli, but not for mayonnaise. I suspect that Freud transposed the wrong absurdity when he determined to tell it himself later on. Indeed, he (p.139) seems particularly annoyed at this joke -- and well he might be if, indeed, he himself damaged it by spoiling the elegance of the frame-shift. I would not mention this were it not for the established tradition of advancing psychiatry by analyzing Freud's own writings.

>ACKNOWLEDGMENTS: I thank Howard Cannon, Danny Hillis, William Kornfeld, David Levitt, Gloria Rudisch, and Richard Stallman for suggestions. Gosrdon Oro provided the dog-joke.

In fact the training process is all about minimizing "perplexity", where perplexity is a measure of how surprised (perplexed) the model is by its training data. It's some exponential inverse of the loss function, I always forget the exact definition.

With enough parameters the models are able to mix and match things pretty well, so the examples of them generating funny jokes aren't necessarily a great rebuttal as there are so many jokes on the web and to find them requires nearly exact keyword matching. A better observation is that we haven't heard many stories of LLMs inventing things. I feel I read about AI a lot and yet the best example I can come up with was some Wordle-like game someone got GPT4 to invent and that was a couple of years ago.

I've found this to be consistently true in my own work. Any time I come up with an algorithm or product idea I think might be novel, I've asked a model to suggest solutions to the same problem. They never can do it. With some leading questions the smartest models will understand the proposal and agree it could work, but they never come up with such ideas cold. What they think of is always the most obvious, straight line, least common denominator kind of suggestions. It makes sense that this is because they're trained to be unsurprising.

Fixing this is probably the best definition of AGI we're going to get. Being surprising at the right time and unsurprising at others is one of the hardest things to do well even for people. We've all known the awkward guy who's learning how to be funny by just saying as much weird stuff as possible and seeing what gets a reaction. And in the corporate environment, my experience has been that innovative people are lauded and praised when they're inventing a golden goose, but shortly after are often demonized or kicked out. The problem being that they keep saying surprising things but people don't like being surprised, especially if it's an unpleasant surprise of the form "saying something true but unsayable", e.g. I don't want to work on product X because nobody is using it. What most people want is a machine that consistently generates pleasant surprises and is a personality-free cog otherwise, but that's hard for even very intelligent humans. It's often hard even to want to do that, because personality isn't something you can flick on and off like a lightswitch. A good example is how Mark Zuckerberg, one of the most successful executives of our era, would have been fired from his own company several times already if he didn't control the voting shares.

This is... false. The LLM is generating text that would be associated with "being a joke" in its corpus. It doesn't know what "funny" or "majority of people" means.