FlyWire Fly Lab · 果蝇全脑实验室

flywire-fly-lab

English · 中文

**Start with [REPRODUCTION.en.md](REPRODUCTION.en.md)** (中文): every claim of each paper we checked, what we got, and what we could not do.
The full lab notebook is [docs/log/report.md](docs/log/report.md) (Chinese).

**Running the whole Drosophila brain connectome on a 16 GB laptop, wiring it to a body, shipping it as a browser game — and doing 24,250 virtual surgeries on it.** It documents what works, what does not, and every mistake the author caught in their own work.

▶ Play online (15,055 real neurons firing in your browser, living in a 3D landscape) · ▶ Ecobox · Full report · Data license ⚠

The web pages follow your browser language (English or Chinese; there is a switch in the corner). The code comments and the lab notebook are in Chinese. This README is generated by scripts/render_readme.py from the same result files as the Chinese README, so the numbers match. Issues in English are welcome.

🎬 Videos

What this is

In 2024 FlyWire published the complete connectome of the adult fruit-fly brain; Shiu, Sterne et al. built a whole-brain leaky integrate-and-fire (LIF) model on top of it; Eon Systems open-sourced that model.

This repository is a complete reproduction plus extensions, all done on one 16 GB M1 Pro MacBook:

Key findings

1. "Is the model accurate?" is the wrong question — its reliability is local. Same parameters, same decision threshold, three pathways; agreement with the paper (Cohen's κ, chance-corrected, all three normalized to the null median) differs a lot:

2. Redundancy cannot be predicted from the wiring diagram. The pre-registered hypothesis was "backup = disjoint paths to the output". Measured correlation: -0.008 (sugar) / +0.013 (water) — not weak, zero, with metric degeneracy ruled out. To know whether two neurons back each other up, you have to knock both out and run it.

3. One behaviour, two circuits. Sugar and water both trigger proboscis extension through the same motor neuron MN9, but their active neurons overlap by only a third and their hubs are private: CB0883 never fires in the water pathway, and knocking out CB0051 does nothing in the sugar pathway. Knocking out the inhibitory neuron Phantom multiplies water-pathway output by 3.0 (whole brain, null-median normalized).

4. Our own pre-registered "concentration → predictability" hypothesis was refuted by our data. The JON pathway is the most concentrated (Gini 0.922) yet the least predictable from wiring alone (G3 0.288 open four > four, open three > …) is entirely correct, and its first choice agrees with a depth-8 search 49.5% of the time (random 1.5%, no brain 34.8%). Given to a rules-only search, it beats the hand-written teacher 31–9 (6-ply vs 4-ply) and 13–7 (10-ply vs 8-ply). But a shuffled connectome (48.8%) and a random network of the same size (51.1%) do no worse; head to head 23–16–1 — the real connectome contributes nothing special to Gomoku; the look-ahead is done by the search, not the fly. See docs/log/report.md §46.

Gomoku: playing against a fly brain

The game page (Digital Fly Lab) has a Gomoku tab: a 15×15 board, and the live spike / descending-neuron / sensory panels show the fly that is playing.

• Three match types: you vs the fly (you play black, with renju forbidden moves), fly vs fly (each side has its own 5,563-neuron brain), fly vs teacher.
• Reasoning depth (1–6, 8) can be switched at any time. Each depth is a trained readout whose labels were "the best move found by an N-ply search"; no search runs while playing, only model inference.
• Its thinking is drawn on the board: orange circles = the model's probability for each candidate; numbered ghost stones = the continuation it expects.
• Real-brain check: each move's line patterns are re-run through the real fly brain and compared with the trained value table.
• External search (off by default): hands the fly's value table to an alpha-beta search. That is an algorithm, not the fly, and the page says so.

How it plays: for a candidate point, take one line in each of 4 directions (4 cells on each side; each cell empty / mine / theirs / edge). There are only 14,641 such patterns, few enough to run every one through the brain: 8 cells × 3 states = 24 channels, each wired to about 61 real sensory neurons, Poisson-driven for 300 ms; downstream spikes are counted (4 input assignments × 3 random seeds, averaged). A move's score = 4 attacking lines + λ × 4 defending lines.

Pre-registered criterion "model trained on depth-8 labels beats model trained on depth-1 labels, ≥ 55% over 200 games" → does not hold (36.0%).

Equally important: (1) the real connectome contributes nothing special (shuffled 48.8%, random network 51.1%); (2) self-play reinforcement ("dopamine"-style three-factor rule) is a negative result: 6–194 against the pre-reinforcement table, and the reward-prediction error is computed outside the connectome; (3) every number of Gomoku v1 is void: stimuli were not cleared between inputs (pattern A alone = 252 spikes, A after B = 1,901). Details: §46.

Life mode: the five senses go to real receptors; the connectome does the rest

A single fly (sub-circuit v4: 6,296 neurons, adding hearing, two olfactory glomeruli and 4 interoceptive neurons) lives in an open world where day and night, fruit, puddles, odours, buzzing, heat, wind, dust and incoming balls just happen. Each one reaches the fly only as a physical quantity at its real receptors — there is no "smell it, go there" or "hear it, run" rule.

Details: §47.

Ecobox: you design the world; it learns and evolves by itself

Open the Ecobox → (docs/ecobox.html, a single offline-capable file)

An unattended round terrarium: food that rots, puddles that dry up, day and night, wind, rain, beetles. You give the fly no task and cannot change its reward — the reward is only that its body's six internal states (hunger, thirst, stamina, health, scent, body temperature) got closer to comfortable. All you can change is the world.

• Single-fly learning: after 30 consecutive lives, learning is frozen and the fly is tested in unseen worlds — no learning 425 s, learned ≥ 1800 s (× 4.231, censored); a time-shuffled-reward control does not improve; 9 of 10 other individuals learn; a long-term training task with 10 fresh individuals: 10/10 learn, 10/10 on a re-test in new worlds (node eco/train_long.js, 210 lives).

What it learns: "stop and drink when you touch water" (in the connectome, water taste drives the proboscis motor neuron MN9 to only 13 Hz in the real brain — not enough innately) and "turn less while humidity is rising". It does not learn to find water from a distance: 422 s vs 417 s in a drought world.

• Evolution with 32 flies: no fitness function — only well-fed, well-watered flies accumulate eggs. Beetle swarm: lifespan 408 → 926 s, flight tendency falls; Calm: lifespan 451 → 930 s, flight tendency flat; Sparse: lifespan 406 → 481 s, flight tendency flat. Pre-registered criterion (different worlds push flight tendency in different directions) passes. Flight is a decision, not a controlled skill: whether to take off is up to the fly, the flight itself is a fixed trajectory — flight control is not wired into the connectome, and we do not pretend it can learn to fly.
• Discovery cards: lying flat, hugging the glass, spinning in place, camping on food… reward hacking and failures are dete