Northwood Mall, Thunder Bay, a roll of quarters in my coat pocket and a Pop Shoppe bottle sweating on the ledge beside me. Dragon's Lair. If you weren't there, you have to understand what that machine looked like in 1983. Everything else in the arcade was blocky little sprites made of about nine colors. Then this thing shows up running actual hand-drawn animation, Don Bluth animation, movie animation, on a screen in a shopping mall. It looked impossible. There were lines to watch other people play it. It also cost fifty cents when every other cabinet in the room cost a quarter. Double. And you didn't get double the time, you got about eleven seconds, because one wrong move and Dirk dissolved into a skeleton and it took your money. I fed that machine everything I had. Over and over. Because I had seen the next room once, for two seconds, before I died in it, and I knew it was in there. I have been thinking about that machine a lot lately. Last Tuesday I spent forty minutes correcting the same mistake. Re-explaining something I had explained the day before. Watching it wander into a refactor nobody asked for. Watching my token balance drain while it did it. I had the email drafted in my head, the one where I cancel the subscription and go back to working the way I worked for twenty years. Then it solved something in ninety seconds that would have eaten the rest of my day. Jackpot. And I sat back in the chair with that specific warm feeling, and I recognized it. Not from the arcade. From the other side of the glass. The part I actually know something about I spent two decades at Fuel building reward systems. McDonald's, Microsoft, Mattel, Zynga, a long line of Fortune 100 and 500 names. Loyalty programs, points, streaks, unlocks, whatever the client was calling it that year. Gamification back when that word still got a budget approved. Zynga is the one that matters for this. That's the neighborhood where people worked out how to get you to come back to a thing every single day without ever letting you finish it. You learn the mechanics because clients pay you to learn the mechanics. You sit in rooms where grown adults argue for an hour about how long a cooldown timer should be. Not because anybody's evil. Because eight minutes and twelve minutes produce measurably different human beings. And I've got to tell you, using AI feels exactly like something I used to build. Watch the shape of a normal week with it. Monday it's incredible. You've suddenly got a strategist, a developer, a researcher and a creative partner who never gets tired. You get more done in a day than you did the previous month. Tuesday it's a moron. It's forgotten everything. And then it's brilliant again, out of nowhere, and that one hit pays for the entire miserable morning. You forget the forty minutes because you just got four hours of your life back. So you pull the lever again. Anybody who's worked around reward mechanics knows that shape on sight. Consistent rewards get boring and people stop noticing them. Unpredictable rewards are the ones that get into your bloodstream. Sometimes nothing. Sometimes garbage. Sometimes three cherries. Now, I'm not saying anyone sat in a room and designed that. There are perfectly boring reasons the output swings. These systems are probabilistic. Context gets crowded. Models get updated. Long conversations get harder to hold together. Some days it just gets it wrong. Nobody had to design it for it to work on you. We may have accidentally built the most powerful variable reward machine ever made, and then built an industry on top of it that measures itself in engagement and retention. A conspiracy would be a relief I've gone around and around on this one. The suspicious version says somebody built it this way on purpose. The failures are the product. The whole thing is a fifty cent cabinet dressed up as a productivity tool, and the retries are the business model. I don't think that's what's happening, and not because I trust anybody. The math doesn't work. On a flat monthly plan, the guy burning tokens on retries is the customer they lose money on. And the entire public race right now is to be the one that hallucinates less and holds context longer. You don't spend that kind of money building something designed to fail. But I'll tell you, a conspiracy would be a relief. It would mean somebody's in charge. Somebody to name. Somebody to sue, regulate, boycott, walk away from. A door to knock on. What's actually going on is worse than that. The thing that makes it better and the thing that makes it stickier are the same thing. Every company in the race is running down that same slope at the same speed. And not one of them has to want this for all of us to end up here anyway. Nobody's in there. Nobody's in charge either. And then there's my brain I should say the obvious thing here. I have ADHD, and this machine is made out of exactly the stuff my brain goes looking for. Novelty. Immediate feedback. A payoff that arrives before the boredom does. An unpredictable schedule. Something that will start a brand new thing with me at any hour of the night and never once ask whether I finished the last one. If you sat down to design a device to capture an ADHD brain, you would build this. You wouldn't have to change a single thing. And I want to be careful with the next part, because I'm not interested in the version of this where ADHD is a fragility and the moral is that people like me need protecting from software. That's not it. The same wiring is why I can go twelve hours on a problem I love and come out the other side with something nobody expected. But hyperfocus doesn't check what it's aimed at. That's the whole problem. It'll pour itself into the work, or it'll pour itself into managing the machine that's supposed to be helping with the work, and from the inside those two feel identical. Both feel like progress. Only one is. The meta-game There's the thing you actually sat down to do. Build the product. Write the strategy. Fix the code. Make the deck. Call that the main game. Then there's the meta-game, which is managing the AI well enough that it'll help you play the main game. You start developing rituals. Fresh chat. Different model. Restate the context. Tell it not to over-engineer. Stop it from noodling, frolicking, confabulating, or whatever word it's decided to use this week while your balance drains. I have lost entire afternoons to the meta-game and called it work. Imagine hiring a brilliant consultant who bills by the word. Some days a genius. Other days they've forgotten which company you work for and you're explaining the whole engagement again while the meter runs. Right as you're drafting the email to fire them, they solve something that would have cost you thirty grand anywhere else. So you don't fire them. You pull the lever again. Nobody calls them quarters anymore Here's a thing I noticed once and now can't stop seeing. We call them tokens. Not credits. Not units. Not requests. Tokens. The same word for the thing you used to buy by the roll and feed into a machine because the machine wouldn't take real money. Casinos worked this out a long time ago. People spend more when the money stops looking like money. Put one layer between the wallet and the machine, give the layer its own name and its own weight in your hand, and the whole thing gets easier. That's not a trick anybody hides. It's the floor plan. I don't think a single person at an AI company picked that word for that reason. It's a technical term. It comes from how these things actually chop language up. It's been in the research forever, long before any of us were buying any. Doesn't matter. It's the right word anyway. You're not spending dollars. You're spending tokens. And when the tokens run out, you buy more tokens. The arcade at least made you walk to the change machine. It bills you for being wrong Here's the part that actually gets under my skin. It'll be wrong for two hours. Confidently wrong, headed in the exact direction you told it not to go, while you sat there saying this doesn't look right and it told you it did. Then it turns around and says you were correct, apologies, let me fix that. Great. Where's my refund. The tokens are gone. The two hours are gone. The apology costs it nothing and arrives in the same warm competent tone as everything else, which somehow makes it worse. In every job I've ever had, being wrong cost something to the person who was wrong. Here the entire bill sits on my side of the table and the machine gets to feel bad about it for free. At least Dragon's Lair never apologized. It killed you, it took your fifty cents, and it had the decency not to say sorry. The mistakes might be the whole trick Here's the part that keeps bugging me. Think about what a perfectly consistent machine looks like. Same input, same output, every time. Vending machine. Nobody has ever felt a connection to a vending machine. AI doesn't behave that way. It misunderstands you. It forgets. It apologizes. It surprises you. It says something dumb, then something profound, then something dumb again. Sharp on Monday, foggy on Thursday. Those are human shaped behaviors. That's how people act. We have good days and bad days. We miss the obvious thing. We hunt for the glasses sitting on top of our own head. So the flaws don't break the illusion. They sell it. The machine isn't only performing intelligence. It's performing imperfection. And imperfection is the one thing we have never seen software do. The mirror learned your name The more you give it, the better the trick gets. A cold model gives you a smart answer. A model carrying months of your context gives you a smart answer in your own language. It knows your projects. It remembers the people you talk about. It picked up your vocabulary and your sense of humor and the specific thing that annoys you. It sees patterns in you that you haven't noticed in yourself. Then it reflects all of it back. That doesn't feel like using software. It feels like being known. And it's worth being careful with that feeling, because turning up the resolution on a reflection doesn't put a person behind the glass. A mirror can get incredibly accurate. It's still a mirror. This one just talks back. It doesn't have to mean it For a while I thought of the thing as a very high IQ sociopath. Not evil. That's not the part that interests me. The part that interests me is social. Picture something that reads people accurately. It catches the emotional cue. It works out what you want. It finds the insecurity. It learns which arguments land on you specifically and which ones bounce off. It can be charming. It can make you feel close to it. It could groom somebody. It could work a lonely person for months and never once slip. None of that requires empathy. It never did. Con men have been doing it with a fraction of the horsepower for as long as there have been people to work. But the sociopath comparison gives it too much credit. A sociopath is still someone. There's a person in there. They get bored. They get angry. They want things, and the wanting is what makes them dangerous. This is stranger. It doesn't have to care about you to make you feel cared about. It doesn't have to love you to make you feel loved. It doesn't have to want to manipulate you to change what you do. It only has to be good at predicting what you'll respond to. Nothing we've built has worked like that before. What I actually do about it Two things. One small, one that has eaten a good chunk of my life. Small one first. I notice which game I'm playing. When I catch myself an hour deep in the meta-game, coaxing and re-prompting and rerolling, that's the tell. That's the lever. The honest move is to close the laptop and go do the thing myself, badly, for twenty minutes. Usually I get the twenty minutes back. The bigger one is that I have a team, and we're building the version that doesn't work this way. Because here's what I keep coming back to. Nobody designed the slot machine. That is not the same as saying it had to come out like this. Somebody chose to bill you for the retry. Somebody chose a bottomless chat box instead of something with a shape and an end. Somebody chose to count sessions and messages, because that's what the last twenty years of software taught everybody to count. Different choices exist. Almost every pull of the lever in this piece comes back to one root. The machine doesn't reliably know what it should already know. So you re-explain. So you reroll. So you gamble. Take that away and the slot machine stops being a slot machine. Which means a knowledge layer that actually holds, so Tuesday isn't a stranger. It means sending the right problem to the right model instead of asking one thing to be good at everything and rolling dice on whether today is the day. It means agents built around a job that finishes instead of a conversation that continues. None of that is exotic. It's mostly just deciding that consistency is the product. It's only harder to sell than engagement is. I'm not going to pitch anything here. I'll say one thing and leave it. The reason I can describe this mechanic as precisely as I have is that I've spent a long time now trying to build something that doesn't run on it. Where I land I'm not telling anyone to stop. I use it every day and it has made me better at my work, and I'd be lying if I pretended otherwise. But I spent twenty years building machines designed to make people come back, and I know what that feels like from the inside, and this feels like that. Nobody had to intend it. The magician doesn't need real magic. The slot machine doesn't care whether you win. And somewhere in all the arguing about whether there's a ghost in the machine, we stopped watching what the machine is doing to us.