A World Model is an AI's internal representation of cause and effect. A robot with a World Model can predict "if I push this cup, it falls" without destroying the cup. That is, simulate outcomes before acting, plan ahead and reason about consequences. This predict-observe-update cycle mirrors how we operate. We build expectations about how things work, and when reality violates those expectations, our mental model adjusts. This is Bayesian reasoning in action: prior beliefs updated by new evidence, weighted by how surprising that evidence is.
The brain is a World Model
Drop a ball, it falls. Touch fire, it burns. Say something rude, people get upset. Your brain has learned the dynamics, it has built the internal physics, your very own Matrix.
You're constantly running simulations. Right now, as you read this, your brain is predicting the next word, the next sentence, where this is going. When I write something unexpected - like mentioning that your tongue is resting on the roof of your mouth - your brain registers a tiny prediction error. It thought we were talking about AI, not your mouth feel. But it's a tiny error, maybe there's a broader point to all this, file it away for now.
Your brain is perpetually predicting what happens next; and correcting when the reality differs. The prediction errors are your training signal. Big surprise means big update. The biggest minds in AI research are trying to get this biological behavior encoded into algorithms.
Conversations are a two-player game
Conversations.
When you talk to people, you're predicting what they'll say next. Most of the time subconsciously. Sometimes, very consciously, when you get into an overthinking spiral and simulate an entire branching conversation with probability weights and prepare for all potentially significant outcomes so you're not caught unawares. No? Not everyone? OK. But subconsciously, yes. Everyone does this. Your brain is constantly running simulations.
A conversation is a tight feedback loop. I predict how you'll react to what I say, what you know and believe and want, whether you're being sarcastic or earnest or confrontational. You predict my meaning and interpret beyond literal words. Both of us continuously update models of each other. This is why good communicators actively test their predictions ("Does that make sense?" "Are you familiar with...?") rather than just broadcasting.
But a good conversation isn't just predictability - it's a dance between expectation and surprise. The sweet spot is when someone violates your expectations in a generative way. "I didn't see that coming, but now that you said it, that's fascinating."
Improv, a formalization
Improv comedy is formalized prediction manipulation. The fundamental rule - "Yes, and" - is about accepting established reality then subverting it productively.
Scene partner says: "I can't believe you brought a lion to the library."
You don't reject it. You accept the model ("Yes, and he's here to return his overdue books").
The laugh comes from the prediction violation. The audience expects confrontation about the lion. They get bureaucratic mundanity instead. Their model updates: Oh, in this world, lions have library cards.
Good improv is structured surprise. You establish patterns (the who, what, when, where), then violate them at precise moments. Too predictable and the audience tunes out. Too chaotic and they can't build a model to violate. The skill is sensing which thread to pull and when.
Surprises Stick
Prediction errors force your brain to encode new information. When reality violates expectations you have to update the model. That update process is memory formation.
Add emotional arousal (surprise, delight, confusion, social bonding), and the memory gets priority encoding. Your brain tags it as significant, worth remembering. This explains why perfectly predictable interactions vanish from memory. "Happy birthday, here's a nice pen." Model unchanged, no update needed, nothing to encode. Gone by next week, even if the pen was expensive. But an unexpected gift? That carves neural pathways.
Expect Subvertations
Being memorable isn't just about being random or shocking. It's about understanding which expectations exist and how to violate them productively.
Productive surprise means the violation makes the model better, not just different. Random chaos forces an update but doesn't improve understanding. A well-timed subversion recontextualizes what came before, adds new connections, reveals hidden structure. The update strengthens the model.
In conversation: Know when to confirm predictions (maintain coherence, build trust) and when to break them (offer insight, humor, novelty).
In improv: Accept what's been established, then add the unexpected element that changes things.
In storytelling: Set up patterns, then subvert them at the right moment. Not before the audience has absorbed the pattern, not after they've grown bored of it.
In relationships: Surprise people in ways that show you understand them, not just in ways that get attention.
The goal isn't constant disruption, that's chaos. It's giving world models something valuable to integrate - something that makes people see things differently, update their understanding, encode a memory worth keeping.
An elephant, as promised
Last week at a party, in between a conversation with a friend of a friend, I said: "Hey, our mutual friend told me it's your birthday. He also mentioned you like elephants."
Our mutual friend, overhearing this, immediately backed me up: "Oh yeah, we did have that conversation."
I say, "So, I brought you an elephant."
I hand him a tiny elephant.
Of course, we didn't have a conversation about his appreciation of our massive mammal brethren. This is an absurd fiction.
But consider the processing running in my friend's head: I love elephants? They're talking about my love of elephants? Wait, he brought an elephant? Oh, it's a tiny elephant!
Each step forced his brain to update its model. The improv principle: "Yes, and." I established fiction, our mutual friend accepted and built on it, my friend got pulled into the shared reality. Collaborative world-building in 30 seconds. He becomes part of an inside joke that literally just started, lore created in real-time.
The origami itself was like ten minutes, some paper, and YouTube. But I bet he's going to remember this gift for a while, because of the experience. The hope is that the origami is a physical anchor for the memory. Every time he sees it, his brain reconstructs the full scene.
Events. Memories. Origami ... gifts are not the things you give but the experiences you share. I've learned that people will forget what you said, people will forget what you did, but people will never forget how you made them feel.
That last bit was by Maya Angelou, not me.