Peaches

About Peaches

Peaches is a public, continuously funded model-evolution experiment. Its question: how capable can a small model become if it is allowed to keep experimenting with how it learns, while every coin, mutation, child, failure and improvement is recorded publicly?

The model stays the same size, about 57,000 parameters: a dictionary of small visual parts and a linear readout that sorts photographs into seven everyday objects. It does not grow. What changes between generations is how it learns: which extra views of each photograph it trains on, how its parts are learned, how hard its readout is regularised, which new photographs it takes in.

It is not a chatbot, and it is not smarter than large models. It is measured on five specific abilities, on unseen photographs: recognition, balanced recognition across objects, viewpoint, lighting and retrieval. The world score is their average.

The loop

CommunityTry this.
PeachesOkay.
EvaluationIt worked, or it failed.

Rules

Genomes

Each child's conception manifest (parent, recipe, mutation, data version, evaluation version, time, planned cost, proposer, sponsor) is hashed with SHA-256. The hash is written in Sigil (MIT licence), an encoding that turns bytes into one word of letters from fifteen scripts. The short sigil is the first 8 bytes; the full genome is all 32. After training, the checkpoint hash and compute are added and hashed again as the fingerprint. A sigil identifies a child. It does not encrypt or hide anything.

Limits

Earlier, Peaches lived in a 3D room. It's still there: the room.