Abstract visualization of a recursive self-improvement loop

RECURSIVE SELF-IMPROVEMENT · LIVE DEMO

A website that improves itself, a little more every day.

Loading the latest generation…

–generations
–quality vs original
–repos tracked
–visitors

Watch: how the loop improves itself (2 min)

Narrated in English or Mandarin; it follows the language switch. Images, motion and voice are AI-generated; the numbers are the site's real results.

Is it actually getting better?

Each run, the loop rewrites its own research strategy, the instructions it gives the AI. The current strategy, a new challenger, and the frozen original all study the same repositories, and a fixed judge that the loop cannot edit scores what each one learned. A challenger only takes over if it wins. If the current strategy keeps beating the original on identical data, the improvement is real.

–quality vs original
–challengers promoted
–strategy version

How a learning is scored (0–1): 25% grounding (are the cited repositories really in what it studied? checked by code) · 25% novelty (how different from everything already learned, measured with embeddings) · 50% specificity and insight, rated by a different AI model that doesn't know which strategy wrote it.

The strategy in charge now

–

Show the original strategy (the frozen control)

Recent challengers

What is Recursive Self-Improvement?

Recursive Self-Improvement (RSI) describes a system that improves its own capabilities, where each improvement increases its ability to make further improvements — a compounding feedback loop. The strong, theoretical form is an AI that rewrites its own cognition. This site is an honest, bounded demonstration of the loop's mechanics: it cannot rewrite its own mind, but it does study the world, learn, and rewrite its own content and rendering code — measurably, and reversibly.

RSI loop ①OBSERVE ②LEARN ③IMPROVE ④RECORD ⑤MEASURE ⑥META-TUNE

The recursive part is step ③: the loop rewrites the very instructions it researches with, and keeps a rewrite only if a fixed judge, one the loop cannot edit, scores it higher on the same data. Step ⑥ then tunes how boldly it rewrites, based on how often challengers win. The process that does the improving is improved by its own output, and checked against a frozen copy of where it started.

The methods, step by step

StepFunctionWhat it does
① Observeobserve()Queries GitHub for real trending projects and the recursive-self-improvement corpus, sorted by stars.
② Learnlearn()Upserts each repo, computes star deltas since last run, and gathers topic and language diversity.
③ SynthesizeevolveAndSynthesize()Champion vs challenger: the AI proposes a revised research strategy. The current strategy, the challenger and the frozen original each distil lessons from the same repositories, and the best lessons are saved.
④ Improveimprove()Rewrites the headline, insight, and a declarative render recipe into the database — informed by the newest learning.
⑤ Judgeevaluate()A fixed judge scores every lesson: grounding (checked by code), novelty (embeddings), and specificity and insight (rated by a different AI model). The challenger wins only by a clear margin.
⑥ Meta-tunemetaTune()Tunes how boldly the strategy is rewritten using the 1/5 success rule: if challengers win often, take bigger steps; if they rarely win, smaller ones.

The knowledge network

A live 3D map of what the loop has connected: learnings, repositories it studied, and topics. Edges link each learning to the projects it cited as evidence. Drag to rotate · scroll to zoom.

Observation breadth across generations

How much the loop looked at each run: stars, new repositories, topic diversity. Until generation 47 this was called “fitness” and steered the loop, but the loop could raise it just by fetching more, so it is not a measure of improvement. Real improvement is tracked above.

What the loop is learning from

Live from the database — the trending and recursive-self-improvement projects the loop currently tracks.

RepositoryStarsΔTopics
Loading…

What the loop has learned

Each run, the loop feeds the projects it studied to an LLM through the Cloudflare AI Gateway and distils concrete lessons. They accumulate here as a growing knowledge base — the site learning from the outside world, one generation at a time.

Self-improvement log

What the site changed about itself on each run, newest first.

  1. Loading…