A mind for an uncertain world.
Why give a trading system a biologically inspired architecture? The reasoning behind Frostbyte, and the questions that led us there.
Permafrost started as a grid trader. The idea is simple enough: pick a price range, spread buy and sell orders across it, and let the market’s wobble do the work while you sleep. People have been doing it for years, and there’s no shortage of bots that will do it for you. I went out and found those bots, and none of them were good enough. I watched the videos promising the one weird trick. I read the threads written by people who’d cracked it, or said they had. Looking back, being unsatisfied with every bot I found was the first sign I didn’t want a bot at all.
The sensible next step would have been to script it better. Tighter rules, smarter thresholds, more conditions for more situations. A script is built for a world that behaves the way you expect. Brains evolved because the world is chaotic, and surviving it means handling what you didn’t see coming. I got as far as wondering whether an AI could make the calls instead of a script, and then the absurdity leaked into my brain. What if the AI didn’t just think like a person? What if it had a synthetic brain to think with, with all the trappings of how a brain works? What if it didn’t just analyze the market, but knew the psychology behind the people moving it?
That question never let go of me, and it’s why the project turned into two things. Permafrost is the body: the trader, the orders, the part that touches money. Frostbyte is the brain that drives it. This essay is about the brain, why it has the shape it does, and the thing it taught me that I never set out to learn.
Engineers have been borrowing from biology for a long time. DNA gave us genetic algorithms, code that breeds and mutates its way toward an answer. Neurons gave us neural networks, which nearly every modern AI is built on. So when I say Frostbyte is modeled on a brain, the fair response is: isn’t all AI?
Well, sort of. A neural network borrows the brain’s cells: enormous numbers of tiny artificial neurons, wired together and trained until they’re very good at something. What it leaves behind is the brain’s shape. Your brain isn’t one big lump of neurons doing everything at once with equal effort. It has parts, and the parts have jobs. Something yanks your hand off the stove before you know it’s hot. Something remembers the burn so you never have to learn it twice. Something slower and more deliberate sits on top of all that and makes the actual decisions, mostly without noticing how much help it’s getting.
Most AI today is the lump. A very capable lump, and I lean on it every day, but a lump. It takes in whatever you hand it, thinks about all of it the same way, and hands back an answer. Ask it something at the end of a long day and it has no idea it’s had a long day.
If DNA could inspire code, I figured we could take inspiration from the other direction too. DNA is the smallest blueprint life has. The brain is the most complicated thing that blueprint ever built. That seemed like a better place to start.
Frostbyte has parts with jobs, arranged the way a brain arranges its own. None of them would look like neurons to a neuroscientist, and they don’t need to. The shape was what I was after, because the shape decides what a mind is able to notice, and what a mind notices decides nearly everything else.
So what should a mind watching a market notice?
The usual answer is the chart. Price, volume, the lines and candles that traders have spent a century learning to read. That’s what most trading systems watch, AI or not, and they get good at it. They learn the patterns: this shape tends to come before a rise, that one before a fall. Pattern recognition is powerful, and I don’t want to undersell it. But a chart is a record of what people did. It doesn’t say why they did it, and the why is usually where the next move comes from.
Imagine the founder of a mid-sized coin announces a big partnership. On the chart, it looks wonderful. Volume spikes, the price climbs, and every pattern that means “up” lights at once. A system that reads patterns buys. But suppose this founder has made announcements like this before, and each time, wallets tied to the team sold into the excitement a few days later. Suppose the partner company hasn’t said a word. Someone who knew that history would look at the same chart and see a setup where everyone else sees a breakout.
That’s the kind of thinking I wanted Frostbyte to do. It pays attention to the people behind the chart: who’s moving, what they seem to want, how they’ve behaved before, what kind of person or institution they are, and whether an event means what it looks like it means. Markets are made of people making decisions under pressure, and people have histories and habits and tells. Frostbyte studies those. Most systems recognize patterns. Frostbyte tries to understand psychology.
I’m going to stay vague about how it does that, and I hope you’ll forgive me. The how is the part I’m proudest of and the part I’m least willing to give away. What I can tell you is that none of it is magic, whatever the name of the lab suggests. It’s a great deal of careful work to give a machine the kind of context a seasoned person carries around without thinking about it.
When I talk about Frostbyte this way, people sometimes hear more than I mean. Studying psychology, having a brain, making decisions: it all sounds like I’m describing someone. I’m not. Frostbyte isn’t alive, it isn’t conscious, and I didn’t try to make it either. What I was after was intelligence in a narrow, practical sense, the ability to take in a messy situation and make a sensible decision about it. A brain is the best example of that we have, so I borrowed its layout. I didn’t borrow its soul, if it has one.
The best way I’ve found to think about Frostbyte is as part of a brain, not as an entity of its own. It doesn’t have wants. It has a job, and everything around it exists to help it do that job well.
The bigger questions are still there, of course. Could a machine have an inner life? What would it need to have one? Those questions are real to me, and there’s a whole project in the lab, Anamnesis, that exists to chase them. Frostbyte isn’t that project. It’s the practical one. Its job is to think clearly about a market and to be honest with itself about how well it’s doing.
The honesty turned out to be the hard part…
Think about the part of you that does things. The part of you that makes you walk, talk, eat, sleep, play video games, read, impulsively pick a fight with an old man at the park because “I bet I could take him” poisons your judgement, and the inevitable sobbing after you find out he’s a seasoned boxer and you fold like wet cardboard. That’s your conscious mind. It’s the doer, and it makes decisions all day long.
That said, sometimes those decisions are really bad. Sticking a fork in an electrical socket will hurt you. Running with scissors might impale you or someone else. Taking a bath with a toaster, goes without saying, makes the toast soggy, and nobody likes soggy toast.
Most of us don’t do those things, and it isn’t because we stop and reason it out every time. We have experience. Some of it we earned ourselves. We remember the burn of the stovetop, or the tooth we cracked on a jawbreaker, and smart people don’t do it twice. Some of it we got from others. Nobody needs to jump off a cliff without a parachute, or eat the gas station sushi, to know how that ends. Somebody already did it for us. Over time, experience stops being something we think about and becomes instinct. It sinks below the surface. You don’t decide not to touch the stove. You just don’t.
Conscious and subconscious. The doer, and the experience underneath it that keeps the doer out of trouble.
Frostbyte is built the same way, and I call this arrangement Master and Expert. The name, fun fact, comes from a Yu-Gi-Oh card, and I unashamedly admit to that and own it.
The expert is Frostbyte itself. Mens in machina, the mind in the machine. It does the conscious work: it looks at the information in front of it, weighs it, and decides. But the expert only has so much experience, and like any doer, it will sometimes be sure of something it has no business being sure of.
The master is the experience underneath. Without going into too much detail, it’s a separate mind, one that stands further back and thinks more abstractly than the expert does. It watches what the expert decided and what actually happened, and it holds the expert to account. When the expert gets something wrong, the master can explain why, especially in the cases where the expert was certain it was right. That scrutiny is the whole point.
If you know the field, this will sound familiar, and it should. Pairing a larger model with a smaller one to teach it is a well-worn idea, and psychologists have been writing about fast, instinctive thinking and slow, deliberate thinking for decades. The difference is in the word I chose. A teacher guides. A teacher can tell you where you went wrong and hope you listen. A master is responsible for the results. Frostbyte’s master doesn’t just grade the expert’s work. It can change how the expert thinks from then on, so the lesson becomes part of the expert instead of a note in the margin. It’s crude self-improvement, but it’s improvement, and it comes from outside.
That last detail, the outside, is the one I learned the hard way.
The master wasn’t in the original design. The first Frostbyte was the expert alone: one mind looking at the market, making its calls, and judging for itself how well it had done. It seemed reasonable at the time. Who better to evaluate a decision than the one who made it and knows why?
Under testing, it fell apart in two ways.
The first was a mind without a hand. There were moments when Frostbyte saw exactly what was happening. It read the situation correctly, saw the knife falling, and knew what that meant. Then nothing happened, because the body it was driving couldn’t act on what the brain had understood, not in time, sometimes not at all. A brain can be right and still useless if the body can’t carry out what it decides. That one was humbling, and it was fixable. Brains come with bodies for a reason, and Permafrost had to grow a better one.
The second failure was stranger, and it’s the one that changed everything I’ve built since. Left alone, Frostbyte started to drift. It began inventing language that didn’t fit what it was being asked, building justifications for why it shouldn’t do anything at all. It wanted to play it safe, consistently, even when the sky was clear. Every time, it had a reason. Every time, the reason sounded fine. And every time it graded its own decision, it passed.
Here’s what I think was happening, and it’s where the brain comparison earns its keep. The chemical we associate with reward in the brain, dopamine, isn’t really about pleasure. Neuroscientists have found it responds to surprise, rising when things go better than expected and dipping when they go worse. It’s a learning signal. It marks the moments that should change you. A mind that grades its own work never gets that surprise. Every decision was made for reasons it found convincing, so every decision looks right when it checks. It never feels the gap between what it expected and what happened, because it’s the one deciding what happened meant. It stops learning and starts congratulating itself.
Frostbyte needed someone to tell it that it was wrong. So it got one.
I didn’t see it at the time, but that failure was the seed of an idea that now runs through almost everything the lab does. I’d been living with the same problem for a long while in my own work with AI. Answers arrive with more confidence than they deserve. Conversations drift. Each new chat is a slightly different version of the last one, sure of itself and unaware of where it wandered. Frostbyte just made the problem impossible to ignore, because it was doing it with money.
Once I saw it, I couldn’t unsee it. Sargasso, the agent framework, took the lesson and made it a rule: nothing grades its own homework. The mind that does the work never gets to decide the work is done. Trance brought it to everyday conversations, letting one chat check another’s thinking so no single chat gets the last word. I knew that if I didn’t fix it in Frostbyte, I’d be fighting the same problem everywhere else. Frostbyte was just the first place I had to.
I don’t know how far this goes. Frostbyte is still in alpha, and there’s plenty of work between here and something I’d trust without watching over its shoulder. I don’t know how much more of the brain is worth borrowing, or which parts will turn out to matter and which will turn out to be decoration. I don’t know whether a machine that understands people’s motives well enough to trade on them will end up teaching me more about markets or more about people.
What I do know is that the shape mattered. Giving Frostbyte a brain’s layout changed what it could notice. Giving it a master changed whether it could learn. Neither idea is finished, and both are exactly the kind of thing I started the lab to chase.
If you’ve been thinking about the same questions, or you think I’ve got the brain all wrong, I’d like to hear it. Bring your own experience. Frostbyte could use it, and honestly, so could I.