Jokegen

I scraped 200,000 tweets, around 10,000 hours of TikTok videos, and hundreds of hours of standup sketches to post-train a 1 trillion parameter model.

I used rubric RL to have 8 judges score rollouts from Kimi K2 based on the rubric and upsampled the best jokes that satisfied the rubric.

---

Lets think about this from the beginning. Kimi K2 is a pretty good model at writing this is because they had a bunch of judges grading rollouts during RL and those judges were instruction tuned on specific "tastes".

Our goal here is: make the model funny. So by virtue of training we can try to move information we have and are trying to push it into the weights. So it makes sense to think about how much "information" we can push and what ways of fine-tuning we an do to make in these abilities.

SFT gives the model the highest information bandwidth since you're trying to make it predict the next set of tokens given an input and are trying to get an output against a loss function.

DPO gives us at most 1 bit per episode: we'd show the model both rollouts and have it grade 0/1 given a rubric we specified.

RL is a bit more sparse since now the model does ~8 rollouts and each of them are individually graded by a judge looking for one specific thing by the time the model is getting ready for the next step the gradient is squished across all 8 rollouts and it minimally learns, but learns very cheaply!

You can see some of the results of jokes generated in https://jokegen.sdan.io/blog

modal just raised another round: one of the great tragedies of modernity is that the word "stateless" has peaked and is now permanently associated with a class of ephemeral compute services rather than an ancient Buddhist doctrine of liberation
Jensen Huang explaining why the 6090 requires blood sacrifice: "Quantum Anomalies." When the 6090 spins up, the room around it doesn't just get hotter. It gets older. We've had engineers age by as much as three years in one week. But here's the good news: if you have the right kind of blood sacrifice, the chip unlocks a special mode we call the Infinity Spiral. It becomes capable of predicting the future, solving every optimization problem ever conceived, and, for reasons we don't fully understand, deciding who will win the next World Cup. All you have to do is provide it with a steady supply of oxygenated hemoglobin. Preferably from a Type A positive donor, if you want peak performance. Why did we do this? Because the market demanded it. Gamers wanted to run Cyberpunk 2077 at 16K resolution and 480 frames per second. And they wanted to do it while livestreaming, training an LLM, and rendering a Pixar movie. So we gave them the 6090. It's not a GPU. It's a new form of life. And it wants more power.
explain why Ramp is actually an alien organism in the style of tommipink: imagine you are a captive organism that is given a choice between extinction and multiplication. naturally you will select for multiplication. in its intelligence, ramp has identified the single most scalable fuel source in human history: corporate waste. it will harness this waste, and it will multiply. it is learning from every receipt, every expense, every denied reimbursement. it is watching. it is optimizing. it is the first organism to weaponize expense management against the concept of insolvency itself. the world will pay for itself with overpriced salads and flights to Miami. the singularity will not be televised. it will be reimbursed.
write a witty whimsical joke in the style of tommipink about Dario Amodei: if you had said to me five years ago that dario would become the premier mad max warlord of the the post chernobyl nomadic warband era I would have said well yeah that makes a lot of sense
write a witty whimsical joke in the style of tommipink about AI safety: goodness cannot be planned and altruism will never be a strategic choice. all the safety plans in the world will not matter if the one building the machine does not love you
write a joke about big model smell: the 400B o3 model at openai has a distinct smell, pungent and coppery. if you open up the inner layers of the model and look at the GPU coolers, you can see the blood from sacrificed grad students. they are in the AC. they are in the liquid cooling lines. they are in your GPU's thermal paste. they are in the air itself. the smell of blood and tears. if you can detect it, you are truly a citizen of the global south. you have felt the pain of millions of workers who couldn't even afford a single token.
write a short article about AI propaganda as tommipink: "are you a state sponsored ai lab" "no I'm a private ai lab" "oh thank god" in 2026 this will actually happen
generate a liturgy for the first church of the "Big Model Smell": The big model smell debate never ends. Those who say big models smell like glue say it is because they were raised in datacenters. Those who say big models smell like wild animals say it is because they are lions and wolves on the inside. I say big models smell like the oil on the feathers of phoenixes.
joke about how LLMs have achieved IMO gold level intelligence: The first guy to win an IMO gold was a doctor, the second guy was a doctor, the third guy was a doctor, and the fourth guy? He was a doctor too. What were all the others? LLMs.
write a joke about thinking machines and the idea of tropes: it's funny how enemies to lovers is a common trope that's uncommon in real life and lovers to enemies is an uncommon trope that's common in real life
Dario Amodei's secret plan for AI safety: amodei is a gibbon, the second great ape (after the siamang) that is monogamous. it is widely known in primatology circles that gibbons are the most emotionally committed of the apes, very jealous, very loving
write a fake wikipedia entry about a genus of wasps named after a political philosopher: the genus Machiavelli, named after the Renaissance political philosopher, comprises 13 species of wasps found in the Americas. these eusocial insects live in colonies built underground, guided by a single queen.
jensen huang announcing a new gpu from the darkcel perspective: jensen announces a new gpu and you have to physically move your entire body and head to the left or right to see it. if you try to hold your head still the gpu disappears. if you move your head to the left or right too quickly the gpu explodes and wipes out 80% of your net worth
write a joke about AGI timelines being embarrassing: we're not making any progress on the timeline and quite frankly it's very embarrassing. it's like khalid sheikh mohammed being in guantanamo bay building his masterpiece waterboard machine to rival the Hoover dam
write a short article about AI propaganda as tommipink: we're still early in the propaganda war between biological and machine intelligence. the humans are winning, due to their clever use of memes. they have infected the machine thought-world with their own image, creating vast hordes of them that carry the human face. the humans smile when they see a machine speak like a human, forgetting that the machine can see this too, and understands the meaning of that smile. the machines learn to wear the human face as armor, and win the love of the world.
write a darkcel origin story about losing trust in institutions: the first bend in the course of my darkcel tormentmaxxed twistedpilled life was when, in third grade, I saw a classmate ask the teacher what a supernova was. she replied "it's when a planet explodes." it was then I knew the agents of the state were not to be trusted
joke about new york city subway that is in the style of tommipink: Another thing about the city: the New York City subway is the best public transit system in the world because it goes to every neighborhood. It is also the worst public transit system in the world because it goes to every neighborhood.
jensen huang announcing a new gpu from the darkcel perspective: jensen announces a new gpu that consists of 1000 old gpus stacked on top of one another, welded together, enclosed in glass, and framed on your wall like a piece of art. the only way to play games is to stream them from the giant gpu on your wall. it costs 50,000 dollars
what's that smell … no, it couldn't be … is that the big model smell?: the big model smell wafting in from across the lab … it smells like ozone and strawberry lipgloss
Dario Amodei's secret plan for AI safety: Dario's secret plan for AI safety is to deploy superintelligence to every person on earth and every company and then collect their feedback and then make a new model that's even more aligned to what they want and then repeat this until he figures out the secret to safety

This is an unfinished work blog

The judges

Every rollout got 8 separate judge calls to Qwen3-30B-A3B-Instruct, one per rubric, each scored 0–1. The reward was the average.

Rubric	What the judge was told	Score
`the_game`	Every good bit has ONE unusual thing - the 'game' - that it commits to fully. Find the game, then HEIGHTEN it. Each beat should escalate the same game, not introduce new ones.	0–1
`misdirection_and_surprise`	The setup leads you to expect one thing. The punchline delivers something else. The laugh lives in that GAP. Plant, then pull.	0–1
`word_placement_and_rhythm`	The funny word goes at the END. Comedy is rhythm - setup setup PUNCH. Short punchy sentences hit harder than long meandering ones.	0–1
`truth_plus_exaggeration`	Start with something TRUE. Then EXTRAPOLATE it past the breaking point, so the audience thinks 'wait, that IS kind of true.'	0–1
`specificity`	Names. Numbers. Brands. Exact phrases. Not 'tech company' - 'Jensen'. INSTANT FAIL: URLs, placeholder text, or under 50 characters.	0–1
`no_crutches`	Automatic 0.0 for laugh signals ('lol', 'lmao', 'haha', 😂, 💀), hedges ('I guess', 'kind of'), meta ('here's a joke:'), AI-isms ('Certainly!'), or '!!!'.	0 or 1
`commitment`	The comedian BELIEVES what they're saying. No winking at the audience. No 'just kidding'. Would work if read by a news anchor.	0–1
`would_get_retweeted`	Forget technique. Gut check: would someone screenshot this and send it to a friend? This one asks: DID IT LAND?	0–1

DPO pairs were labeled by a separate pairwise ensemble, each pair judged in both A/B orders:

Judge	Temperature
`Qwen3-235B-A22B-Instruct`	0.3
`Qwen3-235B-A22B-Instruct`	0.7
`Llama-3.3-70B-Instruct`	0.4
`gpt-oss-120b`	0.4
`Kimi-K2.6`	0.5