About

I'm Jiří Klimeš, a software engineer by trade, and in my own time an ML researcher by stubbornness. I'm building a system that finds answers by search instead of guessing them, because guessing is all today's models do, and it stops at the boundary of the known.

What I work with

Rust the core, and the language I most enjoy thinking in. Runtime errors are mostly designed out, so the compiler holds the invariants and I get to think about the problem instead of the crashes.Python + JAX quick experiments that want composable transforms and arbitrary-order autodiff in both modes, which Rust's Burn lacks.PostgreSQL where I start. pgvector for embeddings, PostGIS for space. I'd rather push one database further than keep several in sync.NixOS on everything but my phone, so an experiment stays a closure I can rerun months later. When I don't have a laptop I sometimes mosh into the workstation from my phone with a Bluetooth keyboard and AR glasses. Works better than it sounds.rerun + MLflow + DuckDB rerun to watch a run as it happens, MLflow to keep track of them, DuckDB to dig through the results.Svelte / Flutter Svelte for web, Flutter for mobile. I can do it. Please don't ask me to.

How I got here

I taught myself to code out of boredom on a vacation: one eight-hour Python video, then a long stretch of tutorials. I climbed out by building things I actually wanted. A notification when my parents' phones hit the home wifi so I knew to look busy. An auto-ordering script for school lunches. A timetable app in Flutter for my class.

Then a friend asked whether neural nets would be hard, and we found out the slow way. We generated datasets of logic ops and addition and trained a single perceptron on them with random mutation. It got stuck on XOR, which is how we learned we needed layers. Then genetic algorithms evolving a population of Snake-playing agents, and finally we cracked backprop. Only after that did we touch Keras, then switch to JAX, writing our own CUDA kernels along the way. Digit classification, a diffusion model from scratch, anomaly detection built from our own ideas, transformers at small scale. When the ARC Prize was announced I spent three months on it. Between that and Chollet's paper On the Measure of Intelligence, it clicked: models trained to recognize don't learn to reason, and scale wasn't going to change that. So I went back to first principles, and that became light-search. Then I proved it couldn't do what I wanted. That proof is where avenance started: a purely symbolic system, one general algorithm that finds answers by search, in any domain.

Where this goes

The plan, stated plainly so I can be held to it: build an assistant that compounds how fast I learn and build. Use it to help build real intelligence, then aim it at aging: epigenetic reprogramming and the hallmarks of aging, pushed far enough to reach rejuvenation and buy us time. Use the time to make the mind substrate-independent, then go see the universe.

Reach me

Email is best.