Simplexions

Population-based online sequence prediction.

Build log
Simplexions screenshot

Simplexions is an experimental algorithm for online sequence prediction. Probability distributions live on the categorical simplex, a population of “simplexions” learns and predicts as the data streams past, and a single wake-sleep mechanism decides which members learn from each token and which take part in each prediction. The population evolves as it goes: new members are born from pairs of parents, and the least useful ones die off.

The reference implementation is a byte-level engine in C with optional OpenMP and Rust bindings, plus an SDL3 and Dear ImGui visualizer. You can watch a population track a small Markov chain on the probability triangle, or feed it a text file, see each byte coloured by how surprised the model was, and have it write from a prompt.

It’s the first research project in my Lean Machine portfolio. There’s no business model, just a technical hypothesis: that the algorithm has useful reach on real text. It’s early days, and each open design question is being settled by experiment.

Learning children's stories byte by byte, coloured by surprise, with samples written along the way
Learning children's stories byte by byte, coloured by surprise, with samples written along the way
The same text re-encoded as trits, so it can be learned on the probability triangle
The same text re-encoded as trits, so it can be learned on the probability triangle