Ongoing Research
Stockfish interpretation
Actively under development: methods, results and the atlas are still changing.
Stockfish, the open-source chess engine, scores positions with neural networks. This project is an attempt to find out what the individual units of that network compute. I built an exact replica of the smaller of the two evaluation networks in Stockfish 17.1, ran it over a database of about 3.2 million positions (about 2.9 million from real games, puzzles and engine self-play; the rest synthetic), and had a language model write a plain-language hypothesis and an executable scorer for each of its 557 active units, out of 576. Each scorer is then tested on held-out positions and on single-piece board edits. The results are published as an atlas with a page for every unit.
The project is under active development, and the results are mixed. Of those 557, 146 meet my recorded "solved" rule, but audits suggest that overstates understanding: my rough estimate is that about 70 have a description that is correct and says more than a generic chess concept such as material balance. When the descriptions stand in for the network's 64 pair units (products of first-layer neurons), they also retain less of its behavior than simple linear probes do. Tightening the criteria and correcting known bad records are still to do.
The Bayesian World
Actively under development: this describes the system as currently implemented, and much of it is still changing.
The Bayesian World is a software system I am building for describing physical devices as exact mathematical models, composing those models from catalogued parts, simulating the result, and fitting uncertain parameters to data. Its core is PMDL-2, a data-only modeling language for ordinary differential equations with exact numbers and units. The same resolved model feeds the simulator, the parameter fitter, code generators, and a translator into the Lean proof assistant. The long-term goal is an agent that designs devices and improves its models from real experiments. That complete loop does not exist yet.
The project is actively under development, and the current state is uneven. The language, simulator and fitter work on simulated examples, although the current version of my fitting demo is broken, and Lean checks a small set of example theorems. The agent system that imports new parts still has a low yield: in my most recent fully audited fresh sample, one of 32 parts passed a separate audit. Nothing described here has been validated against measurements from physical hardware.