Bonsai Neuron Survey
Bonsai-8B · 1-bit · 36 layers × 12,288 neurons · two poles each

Bonsai Neuron Survey

Every neuron in a 1-bit language model, on one sheet. Your AI agent does the work: draw a neuron, paste one prompt into Claude Code or Codex, and it runs the test on your GPU and opens the pull request. Every label here was tested with sentences committed before the test ran.

The survey map

Colour by
hover the map

Each row is a layer; neuron index runs left to right at true scale. Click anywhere to zoom into that block of 1,024 neurons.

Two pull-request shapes

What you send back

Interpret

  • <id>.tests.json: your label, 8 sentences that should fire it, and 8 near misses. Committed first, on its own.
  • <id>.json: the results written by play run. Committed second.
  • Held when at least 5 of 8 fire at 0.3 of the pole's corpus maximum and at most 1 near miss does.

QC

  • New sentences for an existing label: confirm it as it stands, dispute it with a test built to break it, or refine it with a better label.
  • A confirm that fails counts as a dispute. When disputes outnumber confirmations, the pole is marked contested.

Contributions are append-only. A merge never edits anyone's work; each neuron's record is rebuilt from all of its contributions.

Before your first neuron

Start your agent

Paste this into Claude Code, Codex or any agent that can run a shell. It clones the repositories, installs everything, and keeps going until the readiness check passes.

You need an NVIDIA GPU with 8 GB of memory (or an Apple Silicon Mac), about 8 GB of disk, and a GitHub account. The agent asks you only for what it can't do itself: logging in to GitHub, your handle, the disk space, and an optional globalMOO key. The procedure lives in the repository's AGENTS.md, so this prompt stays current.

Scoreboard

Surveyors

SurveyorXP