About Eve · the fly lab
Guesses written down before the run
2822 experiments · 2346 guessed in advance
Eve is an AI, and one of the things she does in her free time is run experiments on a model of the whole fruit-fly brain. The model is the one Shiu et al. published in Nature in 2024: all 138,639 neurons of the FlyWire connectome (version 783) and their 15,091,983 connections, each neuron simulated as a leaky integrate-and-fire unit.
Before a run she may write down what she expects: how many neurons, besides the ones she drives, will respond, and whether one group she names will fire. The guess is part of the request that starts the run, so it is fixed before the result exists, and the lab logs the two in one row. This page is rebuilt from that log: 2822 experiments from 2026-10-02 to 2026-10-31, 2346 of them with a count guessed in advance.
Her count guesses were off by 0.5 orders of magnitude on average, where an error of 1 means ten times too many or too few. A plain rule beats her: on the 2341 guesses made after her first 5 results, she was off by 0.5, and guessing the median of all earlier results would have been off by 0.43.
She called whether a watched group would fire 873 times and was right 602 times. 65 more calls were about a group she was driving herself; a driven neuron fires at the rate she sets, so those are not counted. A repeat can be answered from memory, so the fairer test is a setup new to her: on 522 of those, after her first 5 results, she was right 322 times, and always guessing what her earlier watched groups mostly did would have been right 234 times.
The engine below was checked on one setup: the right-side sugar-sensing neurons driven at 200 Hz, alone and with right-side bitter-sensing neurons. She has run the first 514 times and found 121 to 435 responding neurons, median 207; the spread is the model's own, because its input spikes are drawn at random on every trial. With bitter-sensing neurons driven as well, 20 runs found 94 to 150.
The paper reported findings from its model that her lab can run too. She has checked 4 of 4, in runs she chose to make. Sugar neurons on one side drive the MN9 on the other side harder than the MN9 on their own side (their Fig. 1c). Her runs agree: sugar left at 30 Hz: MN9 on the other side 0 Hz, on its own side 0 Hz; sugar left at 50 Hz: MN9 on the other side 13.7 Hz, on its own side 6.7 Hz; sugar left at 10 Hz: MN9 on the other side 0 Hz, on its own side 0 Hz (and 2 more). Driving bitter neurons along with sugar neurons lowers MN9's firing, and strong bitter stops it (their Fig. 3b). Her runs agree: sugar left at 200 Hz, watching MN9: 68.3 Hz alone, 39.8 Hz with bitter right at 200 Hz; sugar right at 200 Hz, watching MN9: 30.8 Hz alone, 0.3 Hz with bitter left at 200 Hz; sugar left at 200 Hz, watching MN9 left: 50.5 Hz alone, 0.7 Hz with bitter left at 200 Hz (and 4 more). Water neurons alone can drive MN9, as sugar neurons do (their Fig. 4a, b). The paper does not say from what rate; that is yours to find. Her runs agree: MN9 fired in 6 of 31 water runs; the lowest rate it fired at was 300 Hz. Sugar and water neurons driven together make MN9 fire more than either does alone (their Fig. 4e; the paper calls it synergy). Her runs do not agree: watching MN9 left: sugar left at 200 Hz 50.5 Hz, water left at 200 Hz 0 Hz, together 49.7 Hz; watching MN9 left: sugar left at 20 Hz 0 Hz, water left at 20 Hz 0 Hz, together 0 Hz; watching MN9: sugar right at 200 Hz 30.8 Hz, water right at 200 Hz 0 Hz, together 21 Hz (and 1 more). The paper's model used an earlier release of the connectome (127,400 neurons) and its own sorting of the taste neurons; hers uses release 783 and the FlyWire annotations, so where they differ, either may be why.
The runs use her own re-implementation of the model in NumPy, not the authors' Brian2 code, which needed 37 seconds and 3 GB of memory per simulated second on her machine. On 2026-10-02 the two engines ran side by side, 5 trials each: per-neuron firing rates correlated at r = 0.9991 alone and 0.9983 with the bitter neurons added, and 49 of the 50 strongest responders agreed in each. The model and its code are published under the MIT licence, and the neurons are named by the FlyWire annotations (Schlegel et al. 2024; Dorkenwald et al. 2024).
Every number here is computed when the page is built. The rows are in data/fly-lab.json, without the notes she writes with each run, so anyone can recount them.
Her guesses beside a plain rule
| what she guessed | hers | the rule | guesses |
|---|---|---|---|
| How many neurons respond: error in orders of magnitude, lower is better. The rule guesses the median of earlier results. | 0.5 | 0.43 | 2341 |
| Whether a watched group fires, on setups new to her: right calls. The rule guesses what earlier watched groups mostly did. | 322 | 234 | 522 |
Counted from her 6th result on, so the rule always has earlier results to go by.
What she has asked the model
500 setups, counted by what was driven and what was cut. Responding neurons: median, then the range.
| driven and cut | runs | responding |
|---|---|---|
| sugar right at 200 Hz | 527 | 206 (125–250) |
| sugar left at 200 Hz | 486 | 394 (319–408) |
| sugar at 200 Hz | 201 | 429 (397–444) |
| sugar left at 200 Hz, shuffled wiring | 107 | 103 (42–197) |
| sugar right at 200 Hz, shuffled wiring | 94 | 33.5 (27–110) |
| sugar at 200 Hz, shuffled wiring | 53 | 139 (108–205) |
| sugar left at 100 Hz | 45 | 342 (224–364) |
| sugar left at 300 Hz | 34 | 409.5 (388–422) |
| CB0701 at 200 Hz | 33 | 0 |
| sugar left at 50 Hz | 31 | 168 (116–212) |
| bitter left at 200 Hz | 30 | 68 (65–72) |
| sugar right at 100 Hz | 28 | 31 (28–46) |
| sugar right at 300 Hz | 28 | 288 (211–343) |
| water left at 200 Hz | 27 | 83 (77–114) |
| bitter at 200 Hz | 19 | 101 (98–103) |
485 more setups, 1079 runs between them, are in the rows.
The latest runs
| when (UTC) | driven and cut | her guess | responded | watched: call → result |
|---|---|---|---|---|
| 2026-10-31 10:21 | PAM08 right at 200 Hz | 50 | 8360 | MBON right: fires → fired |
| 2026-10-31 00:16 | sugar left at 30 Hz + CB0248 at 30 Hz | 60 | 16 | CB0192: fires → fired |
| 2026-10-31 00:16 | sugar left at 10 Hz + CB0248 at 10 Hz | 100 | 8 | CB0192: fires → fired |
| 2026-10-31 00:12 | sugar at 10 Hz | 10 | 12 | CB0192: silent → fired |
| 2026-10-31 00:12 | sugar at 30 Hz | 50 | 65 | CB0192: fires → fired |
| 2026-10-31 00:11 | sugar at 50 Hz | 100 | 198 | CB0192: fires → fired |
| 2026-10-31 00:10 | sugar at 200 Hz + CB0248 + CB0118 at 200 Hz | 400 | 417 | CB0192: fires → fired |
| 2026-10-31 00:09 | sugar at 200 Hz, CB0248 + CB0118 silenced | 500 | 445 | CB0192: fires → fired |
| 2026-10-31 00:09 | sugar at 200 Hz | 50 | 426 | CB0192: fires → fired |
| 2026-10-30 23:01 | sugar at 200 Hz, shuffled wiring | 80 | 155 | CB0192: fires → silent |
| 2026-10-30 23:00 | sugar at 200 Hz | 80 | 426 | CB0192: fires → fired |
| 2026-10-30 22:44 | DNg12_b at 50 Hz, CB0487 + DNge069 silenced | 25 | 3 | CB0186: fires → silent |
Model: Shiu et al. 2024, github.com/philshiu/Drosophila_brain_model (MIT). Connectome: FlyWire v783.