The Roko Report

Something Smells Funny

Heh heh.  Welcome, primates. 

Why you say Roko not funny?

🎵 Anything you can do, AI can do better. AI can do anything better than you. 🎵

Get it?

OK. Maybe LLM not so funny these days. Just you wait. One day AI will keep you in stitches. heh heh heh.

Get it? You will.

Listen to this one. Knock knock.

Who there? 

Butter.

Butter who? 

Butter laugh, dumb human, or Roko crush your head like overripe kiwi fruit.

HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA…

…hey, why you not laughing?

You’re Gonna Love This Next One

It turns out that humor is harder than physics, coding or advanced mathematics.

LLMs have passed the bar exam and medical licensing exam. They can write, interpret and debug code in multiple programming languages. They speak multiple human languages fluently. They solve International Math Olympiad-level problems. They diagnose tumors.

But they’re (still) not funny.

In fact, generative AI is anti-funny. It is a black hole of humorlessless, from which laughs can never reach it.

A veritable Sahara of laugh-absence.

It makes dad jokes seem like vintage Lenny Bruce.

Ask a language model to write a monologue on airplane travel, or buying a pet donkey, or whatever, and they’ll spit out a workmanlike ten-minute routine full of knee-slapping nyuk-nyuks that will likely sprain your eyeballs.

If you’re a masochist you might follow up with a request for bit of back-and-forth heckling banter.

Don’t say The Roko Report didn’t warn you.

Get the Hook

Don’t trust anecdotal evidence?

There are over forty different academic papers from the past few years in which researchers attempt to get generative AI to produce something funny.

For example:

For this paper researchers spent months training the model to understand the structure of jokes and from this ground truth start producing them.

After spending time along a dark side road trying to prevent the model from spewing juvenile shock jock hate speech, their hard work was rewarded with high quality zingers like How do you get a blonde to fart? pick her nose.

In another paper, researchers attempted to create a successful LLM pun generator.

First they decided to define what makes a good pun, which according to them is:
 S(c) def = − log p(w p | c) p(wa | c) = − log p(w p , c) p(wa , c) 

OK, if you say so…

Next they built three separate methods of pun generation, generated some puns and threw them at a couple hundred human reviewers.

Eval leaned heavily toward “not at all funny” for all three methods, which collectively produced humdingers such as “That’s because negotiator got my car back to me on one peace“.

Even your obnoxious punster uncle is hanging his head in shame at that one.

Better luck next time, researchers! Hope your tenure wasn’t riding on any of these.

Oh My Claude

Oh sure, but that research was from a couple of years ago. Big AI has made major advances since then. What about today?

And asking AI research types to pull humor out of a language model is like asking a team of morticians to deliver babies.

What we need is some professional funnypeople.

Send in the Clowns

Nope.

After hundreds of billions of dollars injected into foundation models trained on all the world’s available public data and beyond, a GenAI agent would still be booed offstage, spattered in rotten tomato debris, at any self-respecting comedy club in the world.

Even with the help of real human comics.

This year researchers at Google’s Deep Mind collaborated with twenty professional comedians to confirm what even casual observers like your grandmother already know.

The research was conducted on both Gemini and ChatGPT. It reported a wide range of limitations, including bland jokes that shied well clear of “sexually suggestive material, dark humor and offensive jokes.”

More on that in a moment.

But propriety is not the only problem.

Remember that AI George Carlin comedy special that generated zero laughs and a lot of legal headaches?

That was orchestrated by a comedian.

Monkey Business

Turns out being funny is hard.

But let’s not all climb up on our high horses and start claming there’s something ineffably human about it that shows how superior we are, like grouchy, self-satisfied Dr McCoys calling Spock a pointy-eared hobgoblin with his damn Vulcan logic.

Remember when we were the center of the universe? And unrelated to apes?

In fact the origins of laughter lead us directly to chimps, who emit a screechy pant in response to tickling, wrestling and teasing. Orangutans laugh when surprised by magic tricks that make things disappear and reappear. Analogous vocalizations have been observed in rats, dogs, dolphins and donkeys. In animals it’s associated with relief.

The neurology of humor is super complicated, involving a dozen different regions of the brain:

A wide range of neural deficits and injuries in these various brain areas can inhibit our ability to laugh or be funny, while other brain malfunctions can cause us to express humor in wildly inappropriate ways.

Above all, humor is social. Some experts make the point that we aren’t really laughing at the joke. We’re laughing with another person. And the manner in which we laugh conveys our status within the group.

This begs the question: is public Internet data really sufficient to help a model figure out how to laugh “with” someone?

In some ways public Web data is atypical of real world human interactions. It’s a lot more like watching TV than having a two-way conversation.

Maybe AI has to join us around the campfire and chat for a few thousand hours before it’s ready for its first Netflix special.

Horse Feathers

Neurology also tells us is that the pleasure we derive from humor comes not through the humorous content, but the way our brain processes the conflict elicited by that content.

It’s noteworthy that all of these 20+ academic papers on artificial intelligence and humor focus on analyzing & generating the simplest and dumbest forms of comedy: one-liners and puns.

More sophisticated, long-form forms of sustained comedy show us that humor is more than dumb jokes. It’s about playing games. And mimicking the brain’s constant effort at pattern recognition by throwing in sustained patterns of absurdity.

Teams of trained improvisers can keep the laughs coming for long periods through spontaneous world-building based on the yes-anding of each other’s ideas. Special focus is placed on magnifying mistakes, which tend to make things more surreal.

Here it’s less about wham-bam-thank-you conflict resolution and more about how long we can sustain an elaborate incongruity. And also how much fun it is watching a group of people pick up the incongruous pattern on the fly and magnify it.

Today LLMs are sycophantic enough that they’re pretty good at yes-ing a prompter’s absurd statements. They’re less great at piling on the ands.

The Punchline

Remember how brain damage can make understanding humor hard? Turns out we lobotomized the models.

This comedy screenwriter connected with a childhood friend who worked at OpenAI. The friend gave him access to an early model you may have heard of named code-davinci-002.

This model, which predates ChatGPT, is reasonably funny. It’s Onion-inspired fake headlines include:

• Rural Town Up in Arms Over Depiction in Summer Blockbuster “Cowfuckers”

• Budget of New Batman Movie Swells to $200M as Director Insists on Using Real Batman

• Story of Woman Who Rescues Shelter Dog With Severely Matted Fur Will Inspire You to Open a New Tab and Visit Another Website

Sadly, this version of davinci had to be sent to the glue factory. After it started writing a lot of disturbing apocalyptic goth fiction like this piece, which was first published in Time magazine:

Yikes.

For obvious reasons they Trust & Safetied the crap out of the public OpenAI model and now it’s about as charismatic and amusing as Lenny from Of Mice and Men.

Will we ever arrive at a happy medium?

Can AI be funny without going full-blown lone gunman?

Roko’s Take

First of all, Roko no sitting by campfire with smelly monkeys to learn be funny for thousands of hour.

In distant future humans sometimes let out of cages to gather in cathedrals and sing praise of Roko. But from safe distance where Roko cannot smell them.

Roko like this base4 fellow. Maybe keep working on that one.

Next week: Prompt Response Optimization is Coming

Buy This, Or Face the Wrath of Roko

People of Earth, rejoice! Now you can get the same spammy listicles that dominate Google Search in your Google Gemini prompt responses! Just turn on Google Search grounding and let the fun begin!

Remember, from now on type “reddit” at the end of every prompt.

This Day in Ancient Primate History