⏱️ The Brilliant Laziness of Being Human: Why Your Brain Refuses to Plan Ahead (And That's Actually Perfect)

📖 Read the full essay

There's a particular kind of morning where the world feels too heavy. You wake up, stumble to the kitchen, and all you want—all you need—is coffee. But if you pause for just a moment and really look at what's around you, something strange happens. The kitchen stops being a kitchen and becomes... data. Thousands of objects. Millions of photons. The grain of the countertop. The exact angle of the chair leg. The three bananas in the fruit bowl (was it three yesterday?). If you actually tried to process all of this, to hold it all in your mind at once, you'd probably just sit down on the floor and give up.

But you don't. You never do. You walk straight to the coffee pot, dodge the chair, step over the Lego from last night, and hit the brew button. It feels effortless, which is the first lie your brain tells you every single day.

New research from MIT and the University of British Columbia has cracked open something profound about how we move through the world, and it turns out we've been dramatically overestimating our own computational power. For decades, cognitive scientists believed our brains worked like supercomputers—scanning entire environments, building perfect 3D maps, calculating optimal paths. We were rational economists of perception, carefully weighing every detail before acting.

We were wrong.

Just in Time" World Modelling Supports Human Planning and Reasoning

This is Heliox: Where Evidence Meets Empathy

Independent, moderated, timely, deep, gentle, clinical, global, and community conversations about things that matter.  Breathe Easy, we go deep and lightly surface the big ideas.

Transcript

0:25

Imagine for a moment that you are standing in your kitchen. It's a Saturday morning, maybe 8 through a.m. Okay. You're a little groggy. The sun is just starting to cut through the blinds, and you have a very, very simple objective. The most critical mission of the day, really. Absolutely. Life cannot proceed without it. You want to make a cup of coffee. I'm with you. But I want you to pause that scene. Just freeze frame it right there. Don't move toward the coffee pot yet. Just look, I mean, really look at the room around you. Right. If you were to strip away your familiarity with this space, base. If you stop seeing kitchen and started seeing just raw data, what are you actually looking at? It's a chaotic disaster zone of visual information. I mean, if you really break it down, you've got the gleaming metal of the toaster, the specific texture of the granite counter, a stack of unopened mail with all these different fonts and colors, a stray Lego brick on the the handle of the dishwasher, the fruit bowl with, I don't know, three bananas and one apple. We are talking about thousands of distinct objects, millions of photons hitting your retina every single second. The geometric relationships too. Like the angle between the chair leg and the table court and the lighting angles. It is a staggering amount of detail. It's an information avalanche. If you were a computer computer program, a robot, trying to process that scene from scratch, your processor would probably, well, it would probably melt. Yeah, the sheer computational cost of identifying every object, mapping its exact location in 3D space, and then calculating its physics is, it's just astronomical. And yet here is the paradox that kicks off our deep dive today. You, the groggy human with the caffeine headache, you don't freeze. You don't stand there paralyzed by the complexity. You navigate through this chaos effortlessly. You step over the Lego without even really looking down. You dodge the chair. You ignore the mail. You grab the mug and you hit the brew button. It feels seamless. It feels automatic. But that feeling of ease is actually a massive lie your brain is telling you. Because underneath the hood, there is a serious problem, a fundamental conflict. we know from decades and decades of psychology that the human brain has hard limits. We have what's called the working memory bottleneck. Right. And we often cite George Miller's famous magic number seven, the idea that the average human can only hold about seven items in their working memory at once. And to be honest, for complex visual scenes like this, it's often even less than that. Maybe three or four things tops. Exactly. So do the math with me here. We have a real world containing millions of details. Right. We have a brain that can hold maybe four things at a time. And somehow we don't crash like an overloaded laptop. We fluently plan, predict, and act. How? That is the core question. How do humans perform high-level planning in high-complexity worlds without a mental crash? The answer, it turns out, is that we might be incredibly efficient. Or, if you want to look at it another way, we are brilliantly strategically lazy. I prefer the term efficient, but brilliantly lazy is probably more accurate to how it feels. Today, we are going on a journey into the mechanics of the mind.

We're doing a deep dive into a fascinating new paper titled Just in Time: 3:33

World Modeling Supports Human Planning and Reasoning. This is a heavy hitter collaboration. We're looking at work from researchers at MIT's Department of Brain and Cognitive Sciences and the University of British Columbia. The team includes Tony Chen, Sam Chayette, Kelsey Allen, Joshua Tenenbaum, and Kevin Smith. These are the detectives of cognitive science, and their mission in this paper is basically to reverse engineer the algorithm running in the human mind. They want to find the source code, the fundamental logic that allows for this efficient simulation. And the narrative arc we are going to trace today is really a story of a regime change. It's a huge shift in thinking. We are witnessing a shift from an old view of the mind to a new view. Okay, set the stage for us. What was the old view? The old view was that the brain is a supercomputer. The idea was that to solve a problem like walking to the coffee pot, your brain essentially scans the entire room builds a perfect high fidelity 3D map of everything in it. A full simulation. A full simulation. Then it calculates the absolute optimal path, and then it executes. That sounds impressive. Very logical. It is. But it's almost certainly wrong. The new view, which this paper supports with some really compelling evidence, is that the brain is a rapid-fire improviser. An improviser. Yeah. It doesn't build the whole map. It builds just enough of the map, just in time, to take the next step. And the stakes here are huge. This isn't just about understanding why we forget where we put our keys. This is, in a way, the holy grail for artificial intelligence. Exactly. I mean, if we want robots that can navigate a burning building or a cluttered warehouse or just your living room without getting stuck spinning in circles, we need to understand how humans solve this problem. We need to teach robots to be brilliantly lazy, too. We do. So let's unpack this. We have a messy kitchen, a limited brain, and a cup of coffee to get to. To understand the solution these researchers found, we first have to understand the fact. The theory wars. The theory wars. In one corner, we have the incumbent champion. The theory that has dominated for a long time, it's called value-guided construal, or VGC. VGC. It sounds very official, very corporate. It's actually very economic. The central concept here is a construal, and you can think of a construal as a simplified mental map. Okay. So the theory admits that you don't imagine the whole world in high fidelity. You create a stripped down version, a caricature almost. Okay. So in my kitchen scenario, my construal might just be the coffee pot and the direct path to it. Yeah. I effectively delete the toaster, the mail, and the fruit bowl for my mental file because they just aren't important for this task. Exactly. And the VGC theory argues that the brain acts like a rational economist. A rational economist. It performs a rigorous cost-benefit analysis on every single object in the room. It weighs the utility of representing an object against the cost of thinking about it. So walk me through that math. How does this rational economist in my head decide what to keep and what to delete from the map? Well, take the chair in the middle of a room. If it's blocking your path to the coffee, representing it has extremely high utility. Right, because if you don't represent it, you trip and fall. Yeah. That's a high cost of failure. So the rational economist says, spend the energy, map the chair, it's worth it. Okay, makes sense. But the coffee mug in the far corner, it's nowhere near my path. Low utility. It's irrelevant to your goal. Representing it costs brain power, but it gives you absolutely no benefit. So the rational economist says, delete it, don't spend the energy. That sounds, I mean, that sounds incredibly sensible, sounds logical to me. So why is that the old view? Why are we challenging that? Because there is a fatal flaw a logical trap and The authors of our paper point this out by referencing some classic work by Russell and we fault from all the way back in 1991, huh? It's a chicken and egg problem that completely undermines the VGC theory Okay, what's the trap? Think about it carefully. To know that the coffee mug in the corner is irrelevant. I have to know where it is and where I'm going. You have to simulate the path first to see that you don't hit it. Oh, wait. To decide what to ignore, you have to verify that it doesn't matter. But verifying it implies you've already processed it. You've already done the work you were trying to avoid. That's the paradox. If I have to simulate everything to figure out what to delete from my simulation, I haven't actually saved any effort. I've just done the work twice. I scanned the mug, realized it was useless, and then deleted it. But the energy cost of scanning it is already paid. Precisely. It's what they call computationally taxing. VGC assumes this optimized map is built before the journey begins. Right, this pregame setup. It assumes you figure out the perfect lean map and then you move. But that pre-calculation is exactly what creates the bottleneck. If you stood in the kitchen and tried to do a full VGC analysis of every object to see if it mattered, you'd be standing there for an hour. So the rational economist model is actually irrational because the cost of figuring out the cost is itself too high. That's a perfect way to put it. So we need a new way, a way that doesn't require knowing everything before we do anything. Enter the Challenger. The theory this paper is championing just-in-time world modeling. Or JIT. JIT. Now this is a term borrowed from elsewhere, right? I feel like I learned this in an operations management class or something. It's a supply chain thing. It is. It's massive in logistics, specifically from the Toyota production system. In the old days of manufacturing, let's call it the just-in-case era, car companies would build thousands of car doors and store them in a massive warehouse. If in case they needed them. Which is incredibly expensive. You have to pay for the warehouse. You have to manage the stock. You have to dust the doors. Exactly. Toyota said, no, this is wasteful. We aren't going to build the car door until the car frame is literally rolling down the assembly line and needs a door attached right now. Machine time. Don't build the inventory until you need to sell the product. Exactly. Correct. And in cognitive science, the inventory is memory and the product is action. So the JIT theory says, Don't clutter your memory with a map of the room until you actually need to take a step. So how does the brain actually do this? What is the mechanism? What's the process? The authors propose that representation isn't a pre-game setup. It is interleaved with the simulation. It happens step by step in a tight little loop. Break that down for me. What are the steps of this loop? It's a three-step cycle and it happens incredibly fast. Step one is simulate. Okay. You take just one small mental step forward. You don't plan the whole walk to the coffee pot. You just think, I'm going to move my foot here to this tile on the floor. Okay. Very short-term planning. It's micro-planning. What's step two? Step two is look ahead. You use your eyes, or your mind's eye, to scan only the area immediately relevant to that specific step. And this is a crucial concept. You treat the visual world as an external memory. External memory. I like that. Unpack it. It's an idea from a researcher named O'Regan that's been around for a bit. The basic premise is, why store the location of the fridge in your brain? The fridge isn't going anywhere. It's right there. It's right there in the room. You can just look at it whenever you need to know where it is. The world itself holds the information for you, so you don't need to duplicate it in your head. So I don't need to remember the world. I just need to be able to access it. Right. So you simulate a step and you look at that spot in the world. Then comes step three in code. If that visual scan finds something, say you realize, "Oh, moving my foot there will hit the dog," then, and only then, do you add the dog to your mental sketch pad. So I'm building the map strictly on a need-to-know basis.

Yes. And here's the kicker: 11:23

there is also a forgetting curve, a decay parameter. Wait, we aren't just adding things, we're dropping them too. Constantly. If you pass the toaster three steps ago and it's no longer relevant to your next step forward, The JIT model says your brain just drops it. It fades out. So we are constantly picking up and dropping details as we move. It sounds so much more fluid. It's less like an architect drawing a rigid blueprint. And more like improvising jazz. That's a great analogy. It's reactive and dynamic. VGC is the architect. JIT is the improviser. But of course this is a scientific paper, not just a philosophy discussion. They had to prove this is actually what humans do. Right, because VGC sounds so logical on paper. We like to think we are these rational optimizers. We want to believe we are the architect. We do. So how did they test this? They looked at two main domains, right? They did. Domain one was planning, which involved navigating a maze. And domain two was physical reasoning, which involved predicting how objects fall. Let's start with the maze. This is the grid world task. Okay, so imagine a really simple video game, a top-down view. You control a little blue circle, and your job is to get it to a green square on the other side of the screen. Classic navigation, like Pac-Man without the ghosts. Exactly. But there are obstacles. Walls, crosses, blocks. Now, here is the really clever part of the experimental design. They used a technique called process tracing. Process tracing. Initially, the maze is masked. It's blank. You can't see where the walls are. Ah, fog of war. Like in Civilization or StarCraft. The map is black until you explore it. Exactly. But instead of moving your character to explore, the participant has to move their mouse cursor over a tile to reveal it. Hovering over a tile acts like a flashlight. It reveals what's underneath, whether it's a wall or an empty space. Why do it this way? Why not just show them the whole maze? Because if you show them the maze, you don't really know what they're thinking about. Your eyes move so fast. These movements called saccades happen in milliseconds. So it's incredibly hard to track attention without expensive eye-tracking equipment. I see. By forcing them to use the mouse to see, The researchers create a precise digital log of exactly what the human is looking at and when they are looking at it It externalizes the mind's eye that is smart It turns attention into a mouse movement a clear recordable signal it is So we have the VGC theory the rational economist and the JIT theory the improviser. What do they predict the mouse will do? Okay, let's start with VGC the planet What does it predict? Well, the VGC prediction, the optimal planner, suggests people should look at objects that would block the optimal path, even if they aren't on the path they end up taking. What does that mean? It predicts a very thorough search. The VGC agent wants to build a complete map of the relevant area to ensure it finds the absolute shortest path. It scans wide, it checks alternative routes, it looks for potential shortcuts.- It wants perfection. It wants to know the best way before it starts moving.- It wants perfection. JIT, on the other hand, predicts something totally different. It predicts tunnel vision.- Tunnel vision.- JIT says people will only look at things on the current path they're simulating. They won't check walls that are far away or behind them, even if those walls might theoretically matter for finding a truly optimal map. So JIT says, I'm planning to go straight. I'm only going to check the tiles directly in front of me. I don't care about the tiles way off to the left or right unless I hit a wall and have to look there. Exactly. And the results, it wasn't even close. Really? JIT explains human behavior significantly better. The paper mentions something called stochastic A-search as the underlying algorithm. Now that sounds a bit intense. Can we translate that for the non-computer scientists listening? Sure. A-star, spelled A, then a star or asterisk, is a classic pathfinding algorithm. It's basically the grandfather of what your GPS navigation uses. It finds the shortest path between two points on a graph or a map. Okay, so it's a pathfinder. Right. And stochastic just means probabilistic. It's not rigid and perfect. It has a bit of randomness or noise built into it. It represents the fact that humans aren't perfect calculators. We make little errors. We vary a bit in our choices. Since it's a fuzzy GPS. A fuzzy GPS is a good way to put it. And here's the key difference. A standard A algorithm in a computer program scans the whole grid, builds a complete map, and then finds the path. Which is what VGC would do. Exactly. But JIT runs this search while it is building the map. It runs the fuzzy GPS, bumps into a fog tile, scans it with a mouse, adds that new information to its tiny map, and keeps going. It's building the bridge as it walks on it. Perfectly said. And it relies on heuristics mental shortcuts. The paper mentions the straight line bias. Humans really prefer straight paths. Okay. We use what's called Manhattan distance, we think, in terms of grid blocks, not diagonals. We want to go straight to the goal, and we only deviate if we physically hit a wall. We don't snake around looking for clever shortcuts through the fog. We take the most direct-looking route and deal with problems as they appear. But does it work? Or are we constantly walking into dead ends because we didn't check the map ahead of time? That's the efficiency analysis the researchers ran. They procedurally generated thousands of random worlds and compared the JIT agent against a VGC agent and also a maximal model that just represents everything. What was the verdict? JIT dominates. It finds good enough plans with vastly less memory and computation. The good enough principle. I feel like that's a recurring theme in human evolution. We aren't designed to be perfect. We're designed to be adequate and fast. It is. JIT doesn't always find the mathematically perfect path. Sometimes it might take a slightly longer route because it didn't see a shortcut hiding in the fog off to the side. But it gets you to the coffee. It gets you to the coffee and it does it with a fraction of the brain power. We trade perfection for speed and low energy. Which, in evolutionary terms, is a winning trade. If you spend 10 minutes calculating the perfect path to escape a lion, you get eaten. If you take a good enough path instantly, you survive. Okay, so that's navigation. Moving across a floor. Static obstacles. But the world isn't just static walls. Things move. Things fall. Physics happens. Which brings us to part three of the paper.

Domain 2: 18:03

The Plinko Board This is about physical reasoning. I love Plinko. From the Price is Right. It's a great visual for the listeners. Imagine that, or a Goulton board. You drop a ball at the top, and it bounces through a field of geometric obstacles, bumpers, triangles, squares, until it lands in one of several buckets at the bottom. So in the experiment, the task for the participants was to look at a static scene of these bumpers and predict, if I drop the ball from here, where is it going to land? Right. But then comes the gotcha, the memory probe. Ooh, I love a good trap, a scientific gotcha. The researchers let the participants look at the scene, make their prediction about which bucket the ball will fall into, and then snap, the screen goes blank. And then? They ask a very specific question. Where was this specific block? They show them a single block from the scene and ask, was it here or was it slightly to the left? So they are testing what stuck. They are testing what the brain decided was worth keeping from that complex scene. Exactly. The hypothesis is simple. If you use that block in your mental simulation, If the ball in your head hit that block on its way down, you will remember its location accurately. If you ignored it, you won't. This is checking the ENCODE step of the JIT loop. Did it actually make it onto the mental sketch pad? Precisely. Yeah. Now, before we get to the results, we have to talk about something they call noisy simulation. because unlike a computer, my brain doesn't know exact physics. I don't calculate friction coefficients or the precise elasticity of a bounce. Right. Humans are noisy simulators. We have jitter in our predictions. When you imagine a ball bouncing, you can't predict the angle perfectly every time. There's a cloud of uncertainty. Maybe it goes left, maybe it goes right. So how does JIT handle that? In the JIT model, they modeled this with a visual favea. You can imagine a spotlight following the mental ball as it drops through the pegs. A spotlight. An object is only encoded, only remembered, if the mental ball comes within a certain radius of it. I think it was about 25 pixels in the experiment. So if the ball misses the block in my imagination, the spotlight never hits the block, and I don't remember where it was. That's the prediction. And the results from experiment 2A were stunning. JIT's predictions about which blocks people would remember correlated with human recall at 0.87. And with their confidence, it was even higher, 0.96. That is startlingly high for a psychology experiment. I mean, usually you're happy with a 0.5 or 0.6 correlation. It's incredibly robust. And there's a nuance here about what they call maybe objects that really sells it for. me maybe objects what are those since our mental physics engine has that jitter that noise sometimes we imagine the ball hitting a block and sometimes we imagine it just missing it's probabilistic okay so a maybe object is one that the ball hits and say only 50% of our internal noisy mental simulations right the JIT model predicts we should remember these may be objects moderately well better than the ones we miss entirely but worse than the ones we definitely hit And that is exactly what humans did. Our memory strength scales with the probability of collision. So our memory strength is directly proportional to how likely we thought a collision was. Exactly. And the decay parameter, the forgetting curve, comes back here too. This is fascinating. Oh yeah. If the ball hits an object early in the drop layup at the top, but then takes a long time to bounce all the way to the bottom, humans tend to forget that first object. Even though it was critical to the start of the path? Yes, because by the time the ball is near the bottom, the top bumper is no longer relevant to the current simulation step. The spotlight has moved on. So the brain drops. JIT captures this with that power law forgetting curve. Our working memory is extremely transient. That is wild. We are literally deleting the past as we simulate the future. It's like Snapchat for cognitive processing. View it once and it's gone. Ideally, yes. Unless we need it again. But the real showdown, the narrative climax of this paper, the part that really made me sit up and go, whoa... is Experiment 2B. Okay. This is where they set a trap to distinguish JIT from VGC once and for all. The smoking gun experiment. They designed two very specific, very clever types of obstacles to mess with these theories. They wanted to find a situation where VGC and JIT would give completely opposite advice. Let's break them down. Obstacle type 1. the counterfactually relevant block that's a mouthful let's call it the butterfly effect block perfect the butterfly effect blocks what does it do here's the setup the ball only hits this block 50% of the time in our noisy simulation it's a total toss-up but if it does hit the ball goes to a totally different bucket at the bottom so it changes the outcome completely it decides the winner of the game so to speak yes now ask the rational economist the VGC model what to do with this block Well, VGC cares about utility. It cares about the outcome. This block determines the outcome. So VGC says, this is high value. You must represent this strongly. This is the most important block on the board. Right. But JIT, JIT doesn't care about the final outcome. It only cares about the process. It looks at the probability of collision. And it says, eh, you only hit it half the time. It's not a sure thing. Represent it moderately. Keep it fuzzy. Okay. So VGC says high importance. JIT says medium importance. now for obstacle type 2 the counterfactually irrelevant block I'll call it the useless block the useless block I like it the ball hits this block 95% of the time it is constantly in the way you can't miss it UT because of how the walls are angled around it The ball lands in the same bucket whether this block is there or not. So it doesn't change the score. It doesn't change the outcome. From a winning the game perspective, it is completely meaningless. It's just noise in the system. Correct. Now VGC, the rational economist looking for utility, looks at this block and says what? It offers no information gain regarding the final bucket. It's useless. Delete it. Don't waste precious energy remembering it. Exactly. VGC predicts you should ignore it completely. But JIT. JIT says, "The ball hits it constantly in your simulation. You must represent it to simulate the path, even if it doesn't change the score." You have to bump into it to get past it. This is the crux of the disagreement. This is fantastic. VGC cares about the result, where the ball lands. J.I. Key cares about the path, what the ball hits along the way. So the verdict, who won? The Rational Economist or the Just-In-Time Simulator? Don't keep me in suspense. The Simulator won by a landslide. Really? How? Human participants had high recall for the useless block. We remembered the useless, though. We remembered it perfectly. And we only had moderate recall for the butterfly block, the one that actually changed the outcome. That completely flips the traditional idea of rationality on its head. It demolishes the view that we optimize for value or outcome. It suggests humans don't care about the final utility of an object. They care about the process of interacting with them. If the ball hits a block in your mind, you see it and remember it, period. Even if that block doesn't change the outcome one bit. We are just-in-time simulators. We are not value optimizers. We process the collision, not the consequence. That is a massive distinction. It suggests that our conscious experience of the world is driven by these micro-simulations. not by some grand strategic plan. We aren't playing chess. We're playing pinball. That's a great way to put it. It explains so much about how we experience the world day to day. We don't filter reality based on, does this matter for my 10-year life goal? We filter reality based on, is this in front of me right now? So let's pivot to the implications. Why did we evolve this way? Why be a JIT simulator instead of a rational economist? I mean, surely being rational is better. I think it goes back to that concept by O'Regan we mentioned earlier. The world is external memory. The kitchen again. The kitchen again. Right. We evolved in a world that is, for the most part, stable. Gravity is constant. The fridge doesn't vanish when you close your eyes. The ground doesn't dissolve under your feet. We don't need to store the world in our head because the world is right there. We can just look at it again. So we developed a cognitive system that outsources the memory load. It relies on that stability. We trade a tiny bit of optimality. Maybe we take a slightly longer path. Or we remember a useless block for massive gains in speed. And energy efficiency. The brain is an expensive organ. It consumes something like 20% of your body's total energy, even though it's only 2% of your body weight. That's right. And BGC precalculating everything to figure out what's useful consumes massive amounts of compute. It's an exhaustive search. JIT is cheap. Run, look, update. Run, look, update. It's lean manufacturing for the mind. This has to be huge for AI and robotics. You mentioned earlier that robots often freeze up in real-world environments. Oh, it's the robot problem. I see this all the time in the DARPA robotics challenges or even with early self-driving cars. A robot enters a cluttered room and just... Freezes. The spinning wheel of death. Exactly. It's trying to calculate the perfect path through thousands of variables. It has a high branching factor, too many possibilities to compute at once. It's trying to be the rational economist, and it's getting paralyzed by the analysis. So the solution is to make the robot lazier. In a way, yes. Robots should adopt JIT. Don't map the whole room. Don't try to solve the entire puzzle before moving. Map the next three feet. Simulate a step. Look with your sensors. Update your tiny map. The paper mentions "lazy collision checking." Is that a real thing? It is. It's a real term in computer science and robotics that perfectly parallels this cognitive finding. Only check for a crash if you're about to step there. Don't check for a potential crash on the other side of the room that isn't relevant yet. It sounds obvious, but traditional AI often tries to solve the whole world state at once. But there are downsides to this, right? I mean, being lazy sounds risky. If I have this intense tunnel vision, couldn't I miss something huge? Oh, absolutely. The limitation of JIT is, well, tunnel vision. If you only look at what our current simulation predicts is important, we miss things. The gorilla effect. Exactly. The famous experiment where people are asked to count basketball passes and completely miss a person in a gorilla suit walking through the middle of the frame. The paper specifically notes this connection to attentional blindness. Because the spotlight never hit it. We didn't simulate a path that intercepted with the gorilla. So for all intents and purposes, the gorilla didn't exist to us. Right. GIT works great until something unexpected and important comes from outside your tunnel. It's a tradeoff. We gain speed and efficiency, but we lose global awareness. There's also the issue of starting. How do we know where to look first? That's a great point and a key area for future research. JIT assumes we're already tracking an object like the ball or avatar in the maze. But what if you were just scanning a static scene looking for your keys? You don't have a path yet. What directs the initial spotlight? JIP doesn't fully explain that. The authors suggest future research needs to combine JIT with heuristics or prior knowledge. Like knowing keys are usually on tables or near the door. Exactly. You don't just simulate blindly from a random point. You use priors, your past experiences and knowledge, to direct the spotlight to high probability areas first. So we've gone from a cluttered kitchen to the cutting edge of cognitive science. We started with a paradox. How do we process a world that is far too big for our brains to hold? And the answer isn't that we have a supercomputer brain that compresses it all perfectly. No. The answer is that we are sophisticated improvisers. We build the world in our heads only as we need it, moment by moment. Step, look, encode. It's surprisingly validating. I don't need to have a master plan for my life. I just need to handle the next three things. feet that's the human algorithm before we go I want to circle back one last time to that counterfactually irrelevant block the useless block that we remembered so well just because we hit it it's the most provocative finding in the paper in my opinion does this explain why we get hung up on minor details in life that don't actually matter to the big picture it might think about an argument you had with a partner or a friend

You might obsess over one specific clumsy phrase they used: 30:16

a collision in the conversation, even if that phrase didn't change the outcome of the argument. Right. Even if you made up five minutes later and the outcome was fine. But because you bumped into that phrase, it's encoded. It's on the map. Your brain tagged it as a significant event in the process, regardless of its impact on the result. So our brains are wired to process the collision, not just the outcome. we remember the friction, we remember the obstacles we hit, even if they didn't ultimately stop us.- That is profound. We are not result-oriented machines. We are experience-oriented machines.- We are. The journey literally defines the memory, not the destination. So next time you find yourself obsessing over a useless block in your day, just remember, it's not you being irrational, it's just your JIT world modeling doing its job. Step, look, and code. Next time you walk through a crowd, notice how you don't map every person. You don't even see most of them, you just process the one you're about to bump into. Notice the tunnel. And appreciate the efficiency of it. Just in time. World modeling. The lazy, brilliant way to be human. Thanks for listening to this deep dive. Thanks for having me. See you next time.