independent research on minds — human and artificial
I'm Yonko Tsonev. I research, learn and work in order to understand — what makes a mind: how we learn, why we feel and what we come to believe, how feeling shapes the reality we act in, how habits form, and what it takes to experience conscience. I study that and incorporate my findings in agents that I build, sometimes in fragments, sometimes whole. Building a mind is the strictest test of understanding one. My goal is to model a mind as close to a human one as I can, so I have a simulated environment where I can test what actually helps learning, wellbeing and mental health.
Live now: Rust & Ruins ↗MonteCarlo.lab ↗Mini CSV Explorer ↗the affect preregistration →
mission
To recreate the mind and model human behavior, by building agents that learn, feel and believe, form habits, and experience conscience.
Each of these is a pillar I work toward, not a claim about what is running today — any one paper or experiment touches a pillar, or a part of one, never all four together. The reason to keep them separate rather than chase one project is that the understanding is meant to compound: study enough of each and the goal is one future model that carries all four at once. And four of these is still not a whole mind — memory, seeing, language, planning and understanding other people are all missing, and I am not pretending otherwise.
what I'm focused on now
This is where my attention sits at the moment. I study each one in people and rebuild it in machines. Each has a plain name and a piece of maths that goes with it.
How does a brain learn, and how can that be modeled to make a machine learn? Reinforcement learning is where the two meet. An agent that works things out on its own gives me something I can test against how people learn — and better ways to teach both.
Feeling is not decoration on top of thinking. How you are doing changes, what you notice, what you avoid, and when you give up. It can turn a reward into something you dread, or a punishment into something worth taking. Work out how that happens and you are most of the way to understanding mental health, where feeling is what stops someone — or what keeps them going. The equation is about that job, not about experience: feeling is the level — how you are doing — and what I model is its change, the thing that moves a decision. I claim nothing about whether the agents feel anything.
Habits are how the mind stops thinking about things. Do something often enough and it turns into behavior that costs you no thought at all — which is the point of it. Most of a day runs on them, and so do discipline, addiction, and character. Note there is no star on that policy: the thing everyone finds about habits is that they carry on after they have stopped being the right move, and that is what makes them worth studying rather than just worth having. I look at how they form, how they stick, and what it takes to change one — in brains, and in agents that save and reuse their own.
Knowing right from wrong starts in what you feel and hardens into what you habitually do — Aristotle got there first, and put it better. I want to know how a mind builds that, and what it would take for an agent to build one too. The formula is a limit, not a sum, on purpose: a conscience is not a number you add to the goal and then trade away when the price is right, it is the thing that says stop. What that limit should be is the open question, and nobody should trust an answer that claims to already know.
what I'm building
Each one asks the same question from a different angle: how much of a mind can be assembled from the outside, and where does it fall apart? Only the last of the three is an experiment in the strict sense — a hypothesis with a way of being wrong. The first two are generative builds, and they answer to whether the thing they produced holds together, which is a different and easier test.
A post-apocalyptic metal band that does not exist. Six members with full backstories, a nine-chapter story, ten songs with videos, and the site itself — the idea, the world, the music, the artwork, the words — all machine-made.
The question was whether a system could hold a world steady: keep six characters recognizable, keep the story straight, and keep making new songs that still belong to the world they came from.
open ↗An AI that builds adventure games in the shape of the Sierra quests I grew up on — plot, world, art, and the game itself, generated end to end.
The goal is not one game. It is a system that keeps making them, at a quality that holds up each time. This is where I push on making things up, and on whether feeling can be built into a story nobody wrote.
Games where agents learn by playing — a Rubik's cube you see one face of at a time, a racing circuit where the whole game is knowing when to brake. Every move is recorded, through the same API for people and agents, so anyone can point their own agent at it, not only mine.
Those recorded games are the data. This is where I get at the harder question: what would it even look like for an agent that only wants to win to stop itself.
The platform — both games, playable by people or agents — is public and MIT-licensed. What isn't there yet is a trained model: that ships with the results, not before.
open tools
Finished things, out in the world today. No signup, no cost, no upsell.
Write your assumptions as ranges rather than single guesses, then run ten thousand trials across them. Percentile bands, sensitivity analysis, goal seek, and a fixed seed so a run can be repeated exactly.
People are bad at not knowing things, in exactly the ways this shows up. You came for one number; what you get is a range, and how wide it is turns out to be the interesting part.
open ↗Open a CSV and actually look at it — adjustable delimiters, sorting, filtering — with nothing to install and no account to make.
It runs entirely in your browser: the file never leaves your machine. A tool should strengthen your own thinking, not make you dependent on mine to do it.
open ↗what I do
The things that exist sit above this — the experiments ↑ and the tools ↑ — and the writing collects in notes. Free, always; the runs that didn't work stay in, because a hypothesis that failed is still information.
how I do it
Building something is the strictest test of understanding it. Every project walks the same loop:
Study the mind
Start from what brains actually do — the research on learning, feeling, habit and moral judgment.
Model it
Turn it into maths — a guess stated precisely enough that a computer can run it.
Build the agent
Build the model as a working system. Where it breaks, the idea was wrong.
Compare to humans
Check what the agent does against what people do. Where they differ is the interesting part.
Publish it openly
Put the result out as a write-up, a tool, and a lesson — the plan before the run, the code alongside the numbers, neither of them held back to look better. Ask what it could be turned into before releasing it; a demo is still something a stranger will run unsupervised. Then loop again, one level deeper.
“What I cannot create, I do not understand.” — the loop is Feynman's blackboard note, made into a method.
notes & demos
The building still runs ahead of the writing. But the plan for the feeling work is written down and published before any of it has been run, because that is the only way a prediction counts for anything.
What I set out to test, what I built, what it did — including the runs that went nowhere.
Papers and books on learning, affect, habit, and moral psychology — argued with rather than summarized.
who's writing this
I have given more than a quarter of a century to following every direction I loved professionally. I spent years putting plays on stage as a theatre director — that's my master's degree. I freelanced as a C developer, to master the first language I fell in love with. I wrote graph algorithms and built knowledge databases, worked on payment and banking systems, and gave ten years to the systems and algorithms behind online advertising and data collection: mastering caching, fighting fierce response-time limits, building targeting systems, while learning different database engines, programming paradigms (functional, object-oriented, procedural), and languages (C, Python, PHP, Java, Go, Rust, JavaScript).
Leadership grew on top of all that rather than instead of it — teams, then managers, then a whole engineering organization — and the code never stopped. Today I lead a full tech organization and spend my working hours building knowledge graphs and agents, and moving an engineering org from traditional engineering to agentic engineering.
Theatre, engineering, management and leadership turned out to be asking the same questions from different sides: what makes a person do what they do, what makes a system behave predictably, and how randomness works. This is where I stop circling them and try to answer directly.