Workspace

Use the infotheory library from your browser. This does not fully encapsulate all functionality of the library

Uploads/URLs are capped at 10MB and cleaned after each run. We do not track user activity, nor view, analyze, or otherwise snoop on provided information. Please use resources responsibly, I am only a student!

Overview

InfoTheory provides fast estimators for entropy, mutual information, compression distances, and predictive rate modeling (ROSAPlus, CTW). This workspace allows you to experiment without downloading anything, as it runs on my server, so we can gain understanding!

Welcome! Get to experimenting, or read more if you prefer via the sidebar.

Workspace

Batch mode runs one JSON object per line and returns JSON per line.

Examples
Code references
Search is disabled on the hosted UI.

The cascade is compute-heavy and tuned for local repos. Run it on your own machine to explore the full search stack and tweak priors, granularity, and ranking settings.

Stage 0 · Prefilter

Unigram likelihood-gain proxy removes the weakest candidates quickly.

  • stage0_keep_frac (default 0.2)
  • top_k (cap for later stages)
Stage 1 · Filter

Entropy-rate filter, optionally conditioned on a universal prior corpus.

  • --prior (local CLI only)
  • --max-order <n>
  • --rate-backend rosaplus | zpaq
Stage 2 · Rerank

KMI/NCD rerank on the top-k shortlist with optional prior summaries.

  • --stage2-prior-mode use | none | summarize
  • --level snippet | file
  • --method <name>
Try it locally
infotheory search "encryption" ./src --level snippet --top-k 50 --max-order 8 --stage2-prior-mode use

Add a prior: --prior "./notes/prior.txt"

Code references

AIXI runs are precomputed offline and summarized here to avoid heavy compute in the UI. The snapshot shows the end of a run: actions, observations, and rewards.

Last steps
Evaluation

Running arbitrary environments is supported by the CLI, but not hosted here due to compute constraints. Use the Library page for more environments and configs.

Code references

Inputs are taken from text, uploads, or URLs only. max_order sets the Markov order (0 = marginal, >0 = sequential, -1 = infinite/auto); CTW backends use method as depth; ZPAQ rate uses method as a streaming level (1-3); mixture is configured below (builder); RWKV7 NCD uses method as coder (ac or rans).

Open documentation
Input A
Input B
Command preview
Code references
Examples

Use __FILE1__ and __FILE2__ to reference Input A/B from the CLI runner. Local file paths are not available in the hosted UI.

Output

stdout

            
stderr

            

            
exit