Companion repository for https://cmpr.ai/hutter/
Each experiment is one commit with self-contained source code. Between experiments the repo is cleaned to a blank slate.
git log --oneline # find the experiment
git checkout <commit> # check out that commit
bash reproduce.sh # run it
Pre-trained models and papers are published at https://cmpr.ai/hutter/archive/ — each experiment's commit message includes the relevant archive URL.
Most experiments use enwik9: http://mattmahoney.net/dc/enwik9.zip
| Commit | Experiment | Archive |
|---|---|---|
f36b42a |
KN scaling: byte KN to 1.784 bpc on full enwik9, zero structure | 20260212 |
efbaac3 |
Weight construction + ES isomorphism: all 82k params from data, 1.89 bpc | 20260211 |
86fbef8 |
Total interpretation: 128-bit Boolean automaton (Q1-Q7), 11 tools | 20260211 |
74ddb51 |
Pattern chains: factor map, reverse isomorphism, viewer | 20260208 |
f808449 |
Pattern priors, skip-patterns, weight construction | 20260207 |
8ce5e5f |
SN visibility (full W_hh + UM comparison + calibration) | 20260207 |
23a9427 |
Saturation + pattern chains | 20260206 |
160190b |
Bayes from table | 20260131_6 |
11a7231 |
Bayes patterns | 20260131_6 |
60acfd9 |
Spectral radius | 20260131_6 |
db2b4ca |
SVD components | 20260131_4 |
51ef90b |
Memory depth | 20260131_4 |
21e719f |
Memory trace (PCA) | 20260131_6 |
c9a60c0 |
AC–RNN trace | 20260131_6 |