Anna.exe & Emergence: train tiny, generalize huge
by Qubic Church2 min read
We reverse-engineered CFB's Anna.exe and turned it into a study of emergent computation, with a new lab tab and three chapters. Note from September 2026. The decomposition is a hypothesis, not a theorem, and the hand-built adder claim was not confirmed.
What changed
CFB released a small binary, Anna.exe, with a (playfully flagged) one-billion-QU bounty for reaching "score 0". We took it apart, rebuilt its algorithm, and used the rebuilt substrate to ask one clean question: what part of an algorithm can be discovered automatically by evolution, and what part cannot?
Three new chapters and a new interactive Evidence lab tab now carry the answer:
- Chapter 32 - Anna.exe & Emergence (overview)
- Chapter 33 - Emergent Generalization - the 15 phases and the proposed decomposition (a hypothesis)
- Chapter 34 - Reverse-Engineering Anna.exe
The headline
Proposed (a hypothesis): for a computation that must generalize to unbounded input size, every component of the algorithm can emerge by mutation and selection on tiny examples - and exactly one primitive cannot (the bare iterate-and-check loop, for a finite-description reason). Each "emerges" is checked by exact match far beyond training: 26/26 seeds converge to the identical full adder, exact to width 128; a mod-3 divisibility machine is invented from scratch and is exact on 80,000 random strings of the tested lengths.
The un-hyped corollary: "score 0" is a learned arithmetic circuit, not intelligence. The shipped monolithic search stays far above 0 (CFB's own screenshot shows ~41% wrong partway through) - composition reaches 0 where blind search cannot.
Try it / verify it
The new Anna.exe & Emergence lab tab (Evidence section) runs a real emergence demo in your browser: it trains a one-cell module on 2-bit addition only, then watches it add much larger numbers it never saw. For the experiments, runnable Python is bundled at /data/emergence-artifacts/ (numpy only).
Honest framing
The novelty here is methodological - an exactly-decidable, exhaustively-verifiable substrate with controls and pre-registration - not a new phenomenon. Every result names its predecessor (Kashtan-Alon, Clune-Mouret-Lipson, Mitchell version-space, and others). This is emergent generalizing computation, demonstrated and bounded. It is not "AI found", and it is not consciousness.
