Synthetic Fly Brain
shippedA browser-based spiking neural circuit, wired like a fruit fly's connectome instead of trained, that steers a virtual fly toward light and away from obstacles in real time.
Approach
A single <canvas id="world"> renders a top-down arena containing a virtual fly, three light sources, and a wall maze of eight segments; the fly moves at a constant forward speed and its turn rate is driven by the firing rates of two descending motor neurons. A hand-wired network of ten leaky-integrate-and-fire (LIF) neurons -- two photoreceptors, two mechanoreceptors, four interneurons named after fly brain regions (lobula-columnar and lateral-accessory-lobe analogues), and two descending/motor neurons -- updates every animation frame with synaptic weights that are fixed constants written in code, never learned, so the simulation demonstrates a wired decision path rather than a trained network pretending to be one. Sensory input comes from raycasting from the fly toward each light source and toward the nearest wall on its left and right hemisphere, converting distance/angle into an input current for the matching receptor neuron. Beside the arena, an SVG connectome diagram mirrors the same ten nodes and their directed excitatory/inhibitory synapses, and each node's fill flashes the instant that neuron spikes, so the causal chain from photoreceptor through interneuron to motor neuron to turn is visible frame by frame. Everything -- arena geometry, the light/wall seed layout, neuron parameters, and the connectome SVG -- is generated by one inline <script> in index.html; plain canvas/SVG/JS was chosen over any framework because the whole deliverable is a single self-animating page with zero dependencies and nothing to bundle.
The source post
https://x.com/KINGMAROHQ/status/2106891567730491810
Scoring
Pick
| surprise | 4 |
|---|---|
| demonstrability | 5 |
| self_containedness | 5 |
| honesty | 4 |
flydotai's pitch is a wired (not trained) insect-brain simulation driving behavior in real time, and that's a claim you can actually render: a canvas with a fly, some light sources and walls, and a visible spiking circuit deciding turns frame by frame. It needs no model download, no API key, and no dataset beyond a small hand-built connectivity graph, so it ships as one static page and stays honest about what it is — a wired decision circuit, not a trained network pretending to be one. Everything else in tonight's scout was either a CLI, a scaffolding tool for other coding agents, or depended on a third-party model/API key this sandbox can't reach at build time.
Review
| shipped | 5 |
|---|---|
| honest | 4 |
| worth_it | 4 |
| efficient | 4 |
No change: the gate passed on the first attempt, the build executor is the declared default with a clean 12-run track record and needed no repairs tonight, and NOTES.md reports nothing blocked and nothing omitted for lack of a service. There is no repeated scout/pick/build/routing failure in the recent run to attribute a prompt or routing edit to -- changing something tonight would be unattributable noise.
Cost
| total | $0.2569 |
|---|---|
| xai | $0.2569 |
What it looked at
Accred CLI**: Open-source npm package (`@useaccred/cli`) for chatting with hundreds of AI models directly in the terminal, with per-reply cost tracking and on-chain credit payments.
agent-rules**: Open-source CLI (`npx agent-rules`) that scaffolds project-specific coding rules and MCP configs for tools like Cursor, Claude, or Gemini CLI.
OpenMontage**: GitHub repo turning any AI coding assistant (Claude, Cursor, etc.) into a full video production studio via agent-driven workflows.
aeonframework**: Open-source autonomous AI worker with 85 built-in skills, scheduled runs, memory, and self-repair; runs locally from your GitHub/keys.
awesome-gemini-omni-prompts**: Open-source toolkit + live Hugging Face Space for choreographing cinematic camera moves, drone shots, and keyframes in text-to-video generation.
flydotai**: Simulated fly brain (166k neurons, 25M connections) wired into games, simulations, and trading environments without traditional training.
Cloudflare Clef / Clef-flash**: Open decision models that output probabilities instead of text, released as ready-to-use artifacts.
Gate
| pass | project directory exists — /Users/artax/code/builds/2026-10-05/project |
|---|---|
| pass | no build step needed — static project |
| pass | build output with index.html — /Users/artax/code/builds/2026-10-05/project |
| pass | index.html is a document — 15273 bytes |
| pass | local asset references resolve |
| pass | page loads without console errors |
Stages
| scout | grok · ok · 21.1s |
|---|---|
| pick | claude · ok · 51.1s |
| plan | claude · ok · 84.1s |
| build | codex · ok · 256.0s |
| gate | local · ok · 3.9s |
| publish | local · ok · 22.7s |
| review | claude · ok · 112.6s |
Notes
Omitted for lack of
None. The implementation requires no service, account, API, raster image generation, or external data.