Is it real?
Is the fruit fly brain simulation real? What is measured, modelled, hand-made or scripted
Videos of a "fly brain" playing Doom or walking in a virtual world spread quickly. Some are driven by a real simulated brain. Others use the brain as decoration, or were never shown in code at all. Here is how to tell the difference, and our verdicts on the most shared claims. To see the clips themselves, open the Video gallery.
Short answer
Partly. The wiring is real: it was measured from electron-microscope images of one fly. The brain activity is a simplified model that leaves out many parts of biology. In most viral demos, a person chose which neurons see the game and which neurons press the buttons, and in several demos scripts or trained parts make most of the visible movement.
No fly has been "uploaded", and no study has shown that these models are conscious. None of the viral game demos we checked tests its own fly against shuffled wiring. The first outside test of one of them, doom-fly-control by gabrycina, found that a brainless autopilot survives in Doom about as long as the full fly brain (see below). Across 107 control studies, the real wiring beat scrambled wiring in 36 of the 97 that test it, mostly reflex, sensory and steering circuits, including a closed-loop test in which scrambled wiring fails; in trained, reservoir and machine-learning tasks it tied or lost. In our own pre-registered test, the answer for one reflex depended on how the wiring was scrambled, and when we gave the same brain a walking body, a scrambled brain and no brain at all walked as far as the real one. But when we ran the map with a looming shadow on one eye, the real wiring turned the simulated fly away in 6 of 6 runs and scrambled wiring in 0 of 3 (small n, one model, our own mapping; caveats below). In a brain-only follow-up, scrambled maps turned up until they were as busy as the real one still gave no turn command (3 of 3), and the turn command appeared from about 30 spikes per second per looming neuron (38% of our strongest input) on both sides; below that the model did not turn. The full comparison is on Does fly wiring help?
This week · 10 Nov 2026
- Repositories of 13 of the 26 verdict rows re-checked: 10 Nov 2026
- Code unchanged since graded 15
- Code changed: re-check queued 0
- No code to re-check 10
- Peer-reviewed paper, code public 1
10 Nov. 10 Nov: we scanned news, videos and new code from 9 to 10 Nov, plus items from 6 Nov that our previous scan missed; one new claim got a dated check: a free browser gamebook, 'What Is It Like to Be a Fly?', that replays recorded fly-brain model activity. 7 new code projects are queued for a first look. The gamebook gets a dated verdict row: It replays recorded spike output from the Shiu et al. 2024 model and from DesktopFly's model; no neuron model runs in the page. The movements and two of its six scenes are written by the author, as its own note says. Credits name FlyWire (CC BY-NC 4.0); a line says "code MIT", but no code repository is linked. Not graded. Read from the page's files on 10 Nov 2026, not rendered. Re-read 10 Nov after code changes: NeuroCraft Fly moves from U to C: its source is now public, and its readouts select scripted body programs. Fly Worker's own numbers after a steering-bias fix (author's numbers, not re-run by us): the fly brain covers 50% of the game map, smoothed noise 63% and uniform noise 49%; a fly brain shown a different board scores 66%. The author withdraws the earlier -3 point paired result. There is still no comparison with scrambled wiring. The repositories of 13 of the 26 verdict rows were re-checked on 10 Nov (report).
9 Nov. 9 Nov: we scanned news, videos and new code from 3 to 9 Nov; one widely seen new video, an Urdu and Hindi explainer whose title says a fly's brain was uploaded, got a dated check (its own description says no mind was copied). Two new write-ups about fly-brain games and puzzles and several small new code projects that say they test whether the fly's wiring matters are queued for a first look. For the walking-circuit paper: We re-ran the authors' nerve-cord walking model on an ordinary 2-processor server without a GPU (9 Nov 2026): with the real front-leg wiring, 3 motor neurons showed a regular rhythm; with the wiring shuffled by the authors' own shuffle, 4 of 5 tries still showed a rhythm (0, 12, 40, 3 and 1 motor neurons in the 5 tries). One run each, faster solver settings than the paper, no noise. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. The shuffled circuits were far more active than the real one (1,645 to 1,995 active cells against 97; top firing rates about 230 against 16 per second), so this is not a like-for-like comparison; a test with activity matched has not been run. Our own ledger row reads no difference by the rule we fixed before the run; the paper's entry stays not tested (report). The repositories of 13 of the 25 verdict rows were re-checked on 9 Nov. Links as on 2 Nov; 57 of 57 videos available; 9 tracked repositories with new commits since 2 Nov were recorded and not yet re-read.
2 Nov. We scanned news, videos and new code from 1 to 2 Nov; nothing met the bar for a new claim check. The repositories of 13 of the 25 verdict rows were re-checked on 2 Nov, none with new code since 31 Oct; links and videos as on 31 Oct. The beginner guide was re-tested, and the specs answer now uses its per-process memory.
31 Oct. 31 Oct: we scanned news, videos and new code from 30 to 31 Oct; nothing met the bar for a new claim check, our news searches found no new coverage of the walking-circuit paper beyond the journal article itself, and two new code projects with their own scrambled-wiring tests were added. For the walking-circuit paper: We checked whether the authors' walking model can be re-run with shuffled wiring on an ordinary 2-processor server without a GPU (31 Oct 2026): its software would need an estimated 620 MB once installed, plus 86 MB of wiring data, more than our 600 MB test space. The test was not run. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. Whether the real wiring is needed for the simulated walking rhythm stays not tested; the earlier circuit count is a pattern in the wiring, not a test of walking.
30 Oct. 30 Oct: we scanned news, videos and new code from 29 to 30 Oct; nothing met the bar for a new claim check, and eight new code projects were listed for a first look. The walking-circuit paper of 29 Oct is now in our catalogue, graded A, with its shuffle test read (a count of circuit patterns in the wiring, not a walking test with shuffled wiring). The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 30 Oct 2026 (checked 30 Oct); no public fly mode found on the preview page (29 Oct).
29 Oct. 29 Oct: we scanned news, videos and new code from 24 to 29 Oct; two claims were checked: a peer-reviewed walking-circuit simulation paper in Nature, and new videos saying a fly brain was 'uploaded'. Both are dated below (two claims checked on 29 Oct). The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 29 Oct 2026; no public fly mode found on the preview page (29 Oct). New plain answers: Did a dead fly come back to life in a computer?, Could this become a 'hell' for future simulated brains? and What do I need to run it?; Is the fruit fly brain open source? now carries the licences as checked on 29 Oct. A commentary video joined the gallery: “They Lied to you about the Virtual Fly Brain” (10 Physics) (not watched by us).
24 Oct. 24 Oct: we re-checked the 31 projects we last graded between 6 and 12 Oct against their current code: 31 keep their grade (1 after a full re-read). 24 Oct: the project left from the 23 Oct re-check, acamilo-flybrain, keeps its grade C after a full re-read. fly-chess: keeps B, and its wiring result is now worse: the author committed a shuffled-wiring control, and under the same readout probe the shuffled brain scored higher than the real one (20.9% against 18.0% top-1; board-blind 11.7%; on the author's Elo ladder 682 against 653 and 631). Only a readout is trained on the frozen connectome. The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 24 Oct 2026; no public fly mode found on the preview page (24 Oct). New plain answer: Is the fruit fly brain open source?
23 Oct. A group-matched control for the relays (details): Left eye: in 1 of 5 matched groups the turn neuron's rate fell as low as, or lower than, with the relays silenced (matched groups 13-31 Hz, relays 16.0 Hz, nothing silenced 21-25 Hz). By our pre-registered rule, this test does not show that the relays matter more than other inputs with the same combined connection strength and activity.
Right eye: among the 9 other active neurons that excite the turn neuron, no group came within 20% of the relays on both combined activity and synapse count (we had set 5 groups as the minimum in advance), so this control could not be built for this eye.
The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 23 Oct 2026. 23 Oct: we re-checked the 28 projects we last graded between 29 Sep and 5 Oct against their current code: 27 keep their grade (5 after a full re-read); 1 more will be re-checked in the next run.
22 Oct. The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 22 Oct 2026. 22 Oct: we re-checked the 34 projects we last graded on 28 Sep against their current code: 33 keep their grade (3 of them after a full re-read of changed code), and 1 could not be re-checked because the source is no longer online.
21 Oct. The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 21 Oct 2026. flyproject.io U: Our check's rule labelled the compute as in-browser (Web Worker), but its size test counted two 3D model files (about 39 MB) as brain data, so the compute label stays unclear. The matched test at 30 spikes/s was repeated with a second input draw: With a second input draw, the matched scrambled brains again sent no turn command (2 input draws on the same 4 scrambled brains).
The sentences and the chart.
20 Oct. The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 20 Oct 2026. Of the two unverified browser pages of 19 Oct, flyproject.io U had its first browser check: loaded, but showed no activity; about 40 MB transferred, mostly 3D model files. The page kept our slow no-GPU machine busy for most of the check, so a faster machine may differ. OrigoZero "Fruit Fly" U was not loaded: needs a GPU. Both rows stay U.
A strengthened scrambled brain, walked in the body (19 Oct). Pre-registered, one run: 19 Oct: walked in the body, the one matched scrambled brain with a small steering signal changed the fly's heading by about 1.4 degrees away from the shadow compared with no brain input, under our 10-degree turn rule; the real brain, with the same input, turned it away by about 16.8 degrees (one run each).
This body run turns the brain's firing rates into steering through our own hand-made mapping; it is not a separate test of the wiring.
The chart point and the limits.
Also on 19 Oct. Two browser pages readers found, as unverified claims U: OrigoZero "Fruit Fly" (the page says the whole MaleCNS v1.0 connectome runs as spiking neurons on the GPU and drives a fly in an arena; no code is linked) and Fly Connectome Explorer (flyproject.io) (the page calls itself "a real fruit fly brain, one circuit at a time"; no repository, licence or author). Both pages responded on 19 Oct 2026; not loaded in a browser by us. The Odyssey-3 claim was re-checked: No method, code or write-up was published as of 19 Oct 2026.
Scrambled brains at 30 spikes/s, strengthened (18 Oct). Pre-registered, brain only: 18 Oct: we then strengthened the connections inside scrambled brains (not the input) until their neurons outside the stimulated input fired about as many spikes as in the real brain (within 30%; same input, one 1-second trial each).
On the left side, 2 scrambled brains matched this way, and neither sent a turn command.
On the right side, 2 scrambled brains matched this way, and neither sent a turn command.
Matching the number of spikes does not match the pattern: 1,152 to 2,228 neurons fired in the matched scrambled brains, against 379 (left) and 432 (right) in the real brain.
19 Oct: walked in the body, the one matched scrambled brain with a small steering signal changed the fly's heading by about 1.4 degrees away from the shadow compared with no brain input, under our 10-degree turn rule; the real brain, with the same input, turned it away by about 16.8 degrees (one run each).
This body run turns the brain's firing rates into steering through our own hand-made mapping; it is not a separate test of the wiring.
As at 80 spikes/s, none of the scrambled brains we ran at 30 spikes/s sent a turn command, including those strengthened to about the real brain's activity outside the input.
The chart and the limits.
Also on 18 Oct. A new verdict row: the Odyssey-3 claim U. Odyssey said on X on 17 Oct 2026 that it placed "the fruit fly brain" inside its Odyssey-3 world model (the post). No method, code or write-up was published as of 18 Oct 2026, so the claim cannot be checked. What would settle it: a write-up or code that names the fly brain model, how its outputs drive the flight and how the world model feeds it input.
Scrambled brains at 30 spikes/s (17 Oct). Pre-registered, brain only: 17 Oct: at 30 spikes/s, the weakest tested input at which the real brain's turn command appeared, we also ran 2 scrambled brains per side (each neuron keeps its number of connections; one 1-second trial each, with the same trial seed as the real-wiring body runs).
On the left side, neither scrambled brain sent a turn command, and their steering signals were identical to having no brain, so in the body they walk like the no-brain fly (the real wiring turned it about 11 deg).
On the right side, neither scrambled brain sent a turn command, and their steering signals were identical to having no brain, so in the body they walk like the no-brain fly (the real wiring turned it about 17 deg).
As at 80 spikes/s, none of the scrambled brains we ran produced the turn; at 30 spikes/s we did not turn their activity up to match the real brain.
The scrambled brains were not matched for overall activity: they fired 48% (left) and 44% (right) as many spikes as the real brain at this strength. Most of those spikes were the stimulated input neurons themselves; outside the input, the scrambled brains fired about 1% as many spikes as the real brain. These body runs turn the brain's firing rates into steering through our own hand-made mapping, so they show the brain result in degrees; they are not a separate test of the wiring. We ran scrambled brains at 30 and 80 spikes/s, not at 20, 40 or 60. The chart, the table and the limits.
Also on 17 Oct. Under the relay headline, a dated line on a stricter control. The Roblox avatar claim stays U: Referee check: no builder named in the NiftRex video description or on the Roblox profile (17 Oct).
The relays, 5 more matched controls on the right side (15 Oct). The 13 Oct headline stays (below). Under it, the pre-registered update: 15 Oct: we ran 5 more matched controls on the right side, 10 in all. Every one of them left the turn neuron's rate higher than silencing the relays did (controls 38-47 Hz, relays 25.7 Hz), so the right-side reading did not change.
Our rule compares the relays with the lowest matched control, so extra controls can only make the relays look less special, never more.
Two verdict rows (15 Oct). The Pokémon-on-a-Raspberry-Pi stream published its code, so we graded it from the code: B, borderline C (game RAM mapped to visual neurons, a hand-written priority arbiter, chat keys and a paid override, only Kenyon-cell to MBON learning, no shuffled-wiring control; stream not watched). New: a Roblox avatar claim from a recap video, U (builder unknown, video not watched, no code).
The looming turn in degrees (14 Oct). Every brain drive of the 8 Oct strength series that differs from the no-brain floor has now walked in the simulated body. The two eyes read differently: on the left eye the simulated fly turned away from 30 spikes/s upward; on the right eye 30 spikes/s was borderline, and that reading rests on one run (seed 0 at 30 spikes/s, 9.18° against our 10° rule). The pre-registered sentences, charts for both eyes and the limits; the body lines for the silenced relays sit under the relay headline (below). These body runs turn the brain's firing rates into steering through our own hand-made mapping, so they show the brain result in degrees; they are not a separate test of the wiring. We ran scrambled brains at 30 and 80 spikes/s, not at 20, 40 or 60. On 17 Oct we added scrambled brains at 30 spikes/s (the pre-registered sentences), and on 18 Oct we strengthened them (the 18 Oct sentences and chart).
The relays, against an activity-matched control (13 Oct). We repeated the 9 Oct relay test with a fairer control: other neurons, matched one by one on how often they fire, so both sides remove about the same amount of activity (the 9 Oct random control removed 2-6 times as much). A brain-only test, pre-registered before the first trial. In the pre-registered words:
The two sides disagree: on one side silencing the relays lowered the turn neuron's rate more than matched activity elsewhere, on the other it did not. We do not draw a conclusion.
- Read with it: 5 seeds per side; 5 matched draws on the left and 10 on the right, 3 relay seeds; some matched draws include neurons that connect directly to the turn neuron; the excitatory relays are mostly one neuron, PVLP141. One model, brain only, a simulated rate, not fly behaviour (chart, numbers and the 9 Oct line with its dated note).
- Verdict rows: the repositories of 13 of the 26 verdict rows were re-checked on 18 Oct 2026 (the rows with a repository); no code changed since grading. Rows without code are carried. On 18 Oct we added one verdict row, the Odyssey-3 claim.
- New plain answer: Can a fly brain learn maths, to hack or to play a game? What changes with practice in 13 projects, each claim attributed to its project.
- 12 Oct: all 26 verdicts re-checked, and the entries re-read in code line by line kept their grades. Our flyvis test gave no result a second time (the dated note).
- Is it a brain or just a map? A new table on How it works lists what the map leaves out, with one number and one source for each part (the plain answer).
- The hack video: the "learns to hack" video had approximately 92,236 views on 29 Oct 2026 (video page) and still links no code, U. On 9 Oct a widely viewed explainer retelling Eon's March 2026 announcement adds nothing our Eon verdict does not already cover.
- Last week: how strong must the shadow be? "The turn command appeared from about 30 per second (38% of our strongest input) on both sides; below that the model did not turn." (chart)
New test: is the fly brain really playing Doom?
Added 28 Sep 2026, chart of all 18 results added 1 Oct 2026 · third-party numbers, not reproduced by us
DOOMFLY is the viral "fly brain plays Doom" demo. A third-party study by gabrycina, doom-fly-control (grade A; code), ran DOOMFLY's own code again with its eyes, brain and button mapping unchanged, and compared it with controls. One of them is a brainless autopilot: it always turns a little, walks forward and holds the trigger, at constant rates.
All 19 rows of the source file
| Arm | Key in the file | Seconds survived (mean) | 95% interval | Episodes | Kills (mean) |
|---|---|---|---|---|---|
| DOOMFLY, real wiring | real | 53.34 | 49.73–56.78 | 20 | 12.1 |
| Matched spray autopilot (no brain) | spray_matched | 51.93 | 45.32–57.28 | 20 | 11.75 |
| Spray autopilot setting t0_f0 | spray_t0_f0 | 6.11 | 5.63–6.67 | 16 | 1.625 |
| Spray autopilot setting t0_f10 | spray_t0_f10 | 8.59 | 7.65–9.57 | 16 | 1.625 |
| Spray autopilot setting t0_f20 | spray_t0_f20 | 9.63 | 8.47–10.89 | 16 | 1.875 |
| Spray autopilot setting t1_f0 | spray_t1_f0 | 5.68 | 5.35–6 | 16 | 0.9375 |
| Spray autopilot setting t1_f10 | spray_t1_f10 | 26.7 | 21.91–31.43 | 16 | 6.6875 |
| Spray autopilot setting t1_f20 | spray_t1_f20 | 53.44 | 48.81–57.31 | 16 | 13.25 |
| Spray autopilot setting t2_f0 | spray_t2_f0 | 5.46 | 5.18–5.76 | 16 | 0.25 |
| Spray autopilot setting t2_f10 | spray_t2_f10 | 9.65 | 8.29–10.99 | 16 | 2.625 |
| Spray autopilot setting t2_f20 | spray_t2_f20 | 21.09 | 16.55–26.08 | 16 | 5.0625 |
| Spray autopilot setting t4_f0 | spray_t4_f0 | 5.28 | 5.06–5.53 | 16 | 0.25 |
| Spray autopilot setting t4_f10 | spray_t4_f10 | 7.6 | 6.73–8.52 | 16 | 1.5 |
| Spray autopilot setting t4_f20 | spray_t4_f20 | 11.41 | 9.39–13.59 | 16 | 3.125 |
| Shuffled wiring, seed 0 | shuf0 | 5.48 | 4.99–5.97 | 4 | 0.25 |
| Shuffled wiring, seed 1 | shuf1 | 9.29 | 7.1–11.47 | 4 | 1.5 |
| Shuffled wiring, seed 2 | shuf2 | 5.48 | 4.99–5.97 | 4 | 0.25 |
| No connections | noconn | 5.5 | 5.24–5.76 | 8 | 0.375 |
| No vision (blind) | blind | 10.46 | 8.84–12.83 | 8 | 2.5 |
What it shows
- The wiring is needed to make the fly move. With shuffled wiring (6.75 s) or no connections (5.5 s) the fly hardly acts and dies quickly.
- But the full brain is not needed to play this well. The brain's output is close to a constant "turn slightly right, walk, shoot" command. It presses fire almost all the time (attack fraction 0.94). A constant command survives 51.9 s against the brain's 53.3 s, and the best of 12 autopilot settings in the same results file survives 53.4 s.
What it does not show
- It does not show that the brain does nothing. The autopilot's constants were taken from the real fly's average output. So the test shows that a constant policy reproduces the behaviour, not that no brain-driven policy exists.
- The shuffled groups are small (4 runs per seed), and shuffled brains are much quieter overall. A control with matched activity is still missing; the study's author says so.
Two other Doom projects with controls point the same way. In doomfly-rl (grade A), a network wired like part of the fly brain is trained to play; real, shuffled and no wiring score the same, both in the author's tests and in doom-fly-control's 100-game re-run. In flydoom by mutkuoz (grade A), an untrained FlyWire whole-brain model beats a random agent with matched commands only narrowly, in one purpose-built arena, and mirroring its eyes removes that advantage. Its author's commits of 29 Sep re-score a separate "everything applied" variant, whose extra health turned out to come from the arena; the plain model's vision control and its grade A stand.
Our verdict on DOOMFLY stays at B: a real connectome runs and drives the game, with hand-made input and output, and there is no evidence of skilled play. More on the DOOMFLY project page.
The counterpoint: a closed loop where the real wiring matters
Added 29 Sep 2026 · the authors' numbers, means recomputed by us from their per-fly table, not re-run
A new study, "Are fruit flies zero-shot adapters?" (grade A), connects three published parts: the FlyVis eye model, the Shiu et al. whole-brain model on FlyWire v783 with unchanged weights, and a NeuroMechFly body. The fly walks toward dark poles; two of its legs are removed mid-walk, and it keeps finding them. The same flies are then run with scrambled wiring and with the turn neurons disconnected.
Set next to the Doom test, the two studies ask the same question and get opposite answers. Both include a baseline in which the brain does not steer (dashed bars): in Doom, a constant command plays as well as the brain; here, a fly that walks without brain steering almost never reaches a pole.
What it shows
- The mapped wiring carries the steering signal. Intact flies reach 2.96 poles per 10 s with the real wiring, 0.34 with scrambled wiring and 0.12 with the turn neurons disconnected.
- It holds after the leg cut. With two legs removed: 3.14 real, 0.12 scrambled, 0.52 disconnected. The brain was not retrained for the new body.
What it does not show
- The brain does not walk. Walking comes from a hand-built rhythm generator; the brain only sets turning, through one hand-set formula on the DNa02 neurons with a gain tuned on separate training flies.
- The controls are strong but small. Scrambling every connection is a big perturbation, and the scrambled comparison after the cut has 16 flies from a single shuffle. Only the per-fly summary table is stored.
This is the kind of test the viral game demos lack. It still does not show skilled behaviour beyond steering toward a dark object, a classic fly reflex. Full details in the catalogue entry and in its row on Does fly wiring help?
We ran the map: is it a brain or just a map?
Added 5 Oct 2026, updated 8 Oct 2026 with 2 more real-map runs and the giant-fibre-off and inside-only body runs · made by us · pre-registered at 09:42:47 UTC on 5 Oct and 09:45:48 UTC on 8 Oct, before the first runs · it changes no grade
We ran the map. A looming shadow on one eye turned our simulated fly away in 6 of 6 runs. With the same map scrambled (same number of connections per neuron) it turned away in 0 of 3. Here the map's wiring, not the stimulus or our code, makes the turn.
"It's just a map" is the most common reply to the viral fly videos, and it is half right: the connectome is a wiring map of one fly, not a living brain. So we asked what the map does when it runs. FLY67, a browser fly, claims that a looming shadow on one eye makes its fly turn away: the eye's looming detectors (LC4 and LPLC2) excite the escape neuron (the giant fibre) and the turning neurons DNa02 on the far side. We rebuilt that test with the Shiu et al. whole-brain model on FlyWire v783, showed the shadow to one eye, and let the brain's turning neurons steer a simulated FlyGym fly through a rule we wrote by hand, on top of a constant walking drive that is also ours. Then we did the same with scrambled maps that keep every neuron's number of connections, with no brain at all, and with random steering.
We ran the map. A looming shadow on one eye turned our simulated fly away in 6 of 6 runs. With the same map scrambled (same number of connections per neuron) it turned away in 0 of 3. Here the map's wiring, not the stimulus or our code, makes the turn. With the escape neuron silenced, the fly still turned away. (2 of 2) With only the inside of the map scrambled, the fly turned away in 0 of 6 runs.
Paths are drawn to scale from the thorax position every 50 ms after the shadow appears; forward is to the right and the fly's left is up. The scrambled-map paths lie exactly under the no-brain floor because their drives were identical. Circles: shadow on the left eye; squares: right eye. Pre-registered at 09:42:47 UTC on 5 Oct 2026, before the first looming trial; the 8 Oct runs were pre-registered at 09:45:48 UTC on 8 Oct, before the first body run. The body does the walking with FlyGym's controller; the mapping and the 0.8 walking drive are ours.
Open the result card (SVG) to share it; it carries the sentence, the numbers, the caveat and its date (the 5 Oct card: 4 of 4 runs).
Read this before sharing. One model (Shiu et al. LIF on FlyWire v783), open loop, our hand-made mapping, a constant external walking drive, 1 s; not a real fly.
- Small n. On 5 Oct the pre-registered time limits left 4 real runs and 3 scrambled runs. On 8 Oct two more real runs walked (seed 2, one per eye), so the "5 of 6" rule now applies as written: 6 of 6. The 3 scrambled body runs sent drives identical to the no-brain floor, so in effect they are one walk; 3 more scrambled drives of 6 Oct are identical to it too.
- The scrambled brains were much quieter. They fired about half as many spikes as the real brain (activity ratio 0.50; outside the stimulated neurons about 1,000 vs 15,000). On 5 Oct we could not run an activity-matched scramble in time, so the ledger first rated the comparison unfair, with weak method. On 6 Oct we ran it, brain only: scrambled maps turned up until they were as busy as the real one still sent no turn command (3 of 3), and their drives equal the no-brain floor, so by its fixed rules the ledger now rates the row fair, with a strong method (details).
- The direction. "Away" rests on our mapping's DNa02 sign and on fly67's claim. As we said before the test: "fly67's claim; the direction in real flies was not verified by us." A 2017 study of walking flies describes looming-evoked backing up and turning through other neurons (LC16 and MDN), not through DNa02.
- A strong, artificial stimulus: every LC4 and LPLC2 neuron of one eye at 80 Hz for 1 s, as fly67 does. A real approaching object recruits fewer cells, with timing. On 8 Oct, brain only, the turn command appeared from about 30 spikes per second per looming neuron (how strong must the shadow be?); that input is a rate we set, not a measured response to a real shadow. The giant-fibre-off and inside-only drives walked on 8 Oct: "With the escape neuron silenced, the fly still turned away." (2 of 2) "With only the inside of the map scrambled, the fly turned away in 0 of 6 runs."
The brain part was reproduced. fly67's own pass rule held in our Brian2 run of the Shiu et al. model: giant fibre 111.5–120.5 Hz (fly67's engine: 114.8 / 120.3), DNa02 on the far side 23–44 Hz, near side 0 Hz. In all 3 scrambled brains every readout stayed at 0 Hz. Our test is not fly67's own control, so fly67 stays grade B.
Numbers behind the chart (all 19 body runs)
| Run | Group | Shadow on | Heading change (°) | Away (°) | Away beyond floor (°) | Turned away | Forward after onset (mm) |
|---|---|---|---|---|---|---|---|
| F (floor) | No brain, constant drive (floor) | – | -1.49 | – | – | – | 11.50 |
| L-real-0 | Real map | left eye | -33.61 | 33.61 | 32.12 | yes | 10.73 |
| L-real-1 | Real map | left eye | -40.63 | 40.63 | 39.14 | yes | 10.47 |
| R-real-0 | Real map | right eye | 64.06 | 64.06 | 65.55 | yes | 8.30 |
| R-real-1 | Real map | right eye | 68.29 | 68.29 | 69.78 | yes | 8.24 |
| L-D3 | Scrambled map (same drive as F) | left eye | -1.49 | 1.49 | 0.00 | no | 11.50 |
| L-D4 | Scrambled map (same drive as F) | left eye | -1.49 | 1.49 | 0.00 | no | 11.50 |
| R-D3 | Scrambled map (same drive as F) | right eye | -1.49 | -1.49 | 0.00 | no | 11.50 |
| rand-L0 | No brain, random steering | left eye | 6.58 | -6.58 | -8.07 | no | 11.35 |
| rand-L1 | No brain, random steering | left eye | 10.39 | -10.39 | -11.88 | no | 11.37 |
| rand-L2 | No brain, random steering | left eye | -1.10 | 1.10 | -0.38 | no | 11.25 |
| rand-R0 | No brain, random steering | right eye | 19.77 | 19.77 | 21.26 | yes | 11.21 |
| rand-R1 | No brain, random steering | right eye | 14.77 | 14.77 | 16.25 | yes | 11.01 |
| rand-R2 | No brain, random steering | right eye | 18.71 | 18.71 | 20.20 | yes | 10.89 |
| LOOML-real-2-f | Real map (8 Oct) | left eye | -38.67 | 38.67 | 37.18 | yes | 10.22 |
| LOOMR-real-2-f | Real map (8 Oct) | right eye | 53.94 | 53.94 | 55.42 | yes | 9.10 |
| LOOML-real-2-gfoff-f | Real map, giant fibre silenced (8 Oct) | left eye | -33.02 | 33.02 | 31.53 | yes | 10.57 |
| LOOMR-real-2-gfoff-f | Real map, giant fibre silenced (8 Oct) | right eye | 53.36 | 53.36 | 54.84 | yes | 9.23 |
| LOOMR-B-301-f | Inside-only scramble (8 Oct) | right eye | -2.92 | -2.92 | -1.43 | no | 11.50 |
| Same drive as F (6 Oct drives) | Plain scrambles LOOML-D-5-f, LOOMR-D-4-f, LOOMR-D-5-f and inside-only scrambles LOOML-B-301/302/303-f, LOOMR-B-302/303-f sent exactly the floor's drive (same sha256), so they take the floor's walk; the 8 Oct floor run repeated the 5 Oct run exactly (140 × 5 values). | ||||||
Pre-registered reading (5 Oct, and again with n 6 on 8 Oct): "the turn away needs the wiring" (scrambled retention 0.00 by the pre-registered floor method, 0.0096 in the ledger's single-floor method; both under the 0.1 line). Share of drive values at the ±1.2 clip: 0.00. Mapping (ours, hand-made, not fitted): drive_s = clip(0.8 − m_s − 0.5·(a_s − a_o), −1.2, 1.2), with a = DNa02 rate ÷ 100 Hz and m = mean MDN rate ÷ 100 Hz; the 0.8 is an external walking drive, not the brain.
What the brain sends down, beside fly67's own numbers
fly67's own pass rule held in our Brian2 run, so the brain part is reproduced in a second engine: giant fibre 111.5–120.5 Hz (fly67's engine: 114.8 and 120.3 Hz), DNa02 on the far side 23–44 Hz, on the near side 0 Hz. In every scrambled map all of these stayed at 0 Hz. fly67 stays grade B: our test is not the author's own control.
| Trial | Shadow on | Giant fibre L / R (mean) | DNa02 L / R | LI | MDN (mean) | Whole-brain spikes | Active neurons | fly67's rule |
|---|---|---|---|---|---|---|---|---|
| fly67, the author's engine (committed, 45f62fd9) | left eye | – (114.8) | 0 / 25 | – | 0.8 | – | – | pass |
| fly67, the author's engine (committed, 45f62fd9) | right eye | – (120.3) | 43 / 0 | – | 3.3 | – | – | pass |
| Real map, seed 0 (our brain gate) (L-real-0) | left eye | 129 / 94 (111.5) | 0 / 23 | 0.96 | 0 | 27,367 | 657 | pass |
| Real map, seed 1 (L-real-1) | left eye | 129 / 99 (114) | 0 / 28 | 0.97 | 0 | 27,538 | 711 | pass |
| Real map, seed 0 (our brain gate) (R-real-0) | right eye | 89 / 149 (119) | 44 / 0 | 0.98 | 5.25 | 27,701 | 677 | pass |
| Real map, seed 1 (R-real-1) | right eye | 85 / 156 (120.5) | 44 / 0 | 0.98 | 3.5 | 27,509 | 682 | pass |
| Scrambled map, shuffle 3 (L-D3) | left eye | 0 / 0 (0) | 0 / 0 | 0.00 | 0 | 13,993 | 349 | fail |
| Scrambled map, shuffle 4 (L-D4) | left eye | 0 / 0 (0) | 0 / 0 | 0.00 | 0 | 13,938 | 336 | fail |
| Scrambled map, shuffle 3 (R-D3) | right eye | 0 / 0 (0) | 0 / 0 | 0.00 | 0 | 13,134 | 351 | fail |
| Run on 6 Oct 2026 (brain only) | The trials missing here (real seed 2, shuffles 4 and 5, the giant fibre silenced and the gain-matched scrambled brain) ran on 6 Oct with a faster input that passed a pre-registered check against these trials, plus a new interior-only scramble. Their numbers are in Where in the wiring is the turn? | |||||||
What it shows
- The map's wiring carries the signal. With the real wiring, the escape neuron and the far-side turning neurons fire; in scrambled maps with the same number of connections per neuron they stay silent, and the fly walks exactly as if it had no brain.
- Not the stimulus or our code alone. Every arm got the same shadow, the same mapping and the same body; only the wiring differed. Left shadows turned the fly right and right shadows turned it left (3 of 3 each).
- Two more controls in the body (8 Oct). "With the escape neuron silenced, the fly still turned away." (2 of 2) "With only the inside of the map scrambled, the fly turned away in 0 of 6 runs." The no-brain floor run repeated the 5 Oct run exactly in a new session.
- fly67's brain-level claim holds in a second engine (our Brian2 run of the same model).
What it does not show
- That a real fly turns this way. The direction rests on our mapping's DNa02 sign and on fly67's claim. Our caption, fixed before the test: "fly67's claim; the direction in real flies was not verified by us." In real flies the giant fibre triggers a take-off, not walking.
- A large test. 6 real runs and 3 scrambled body runs (in effect one walk; 3 more scrambled drives are identical to the no-brain floor, so their walk follows by determinism), and the plain scrambled brains were much quieter (activity ratio 0.50). The activity-matched scramble followed on 6 Oct at brain level (below); its drives equal the floor's, and the ledger rates the row fair, with a strong method.
- That the map is a brain. One model (Shiu et al. LIF on FlyWire v783), open loop, our hand-made mapping, a constant external walking drive, 1 s; not a real fly. As in Who does the walking?, the body does the walking with FlyGym's own controller.
Share it with its caveats: the result card (SVG) carries the sentence, the numbers and the caveat. The ledger row: Does fly wiring help? · build it: Add a sense: looming · method and every run: report of 5 Oct 2026 and, for the runs added on 8 Oct, the report of 8 Oct 2026.
Where in the wiring is the turn?
Added 6 Oct 2026 · made by us · brain only · pre-registered at 09:42:54 UTC before the first trial · it changes no grade
The looming test above left two questions. The scrambled brains were about half as active as the real one, so was the turn lost only because they were too quiet? And which part of the wiring carries it? We went back to the brain model alone (no body) with the same shadow and asked three things: does a scrambled map turned up to the real map's activity produce the turn command? Does scrambling only the inside of the map, while keeping every connection onto the output neurons, remove it? And does the turn need the escape neuron?
We turned the scrambled maps up until they were as busy as the real one. They still did not produce the turn: the wiring, not the amount of activity, makes it.
One model, brain only: the caveats are inside the chart and listed below it. Circles: shadow on the left eye; squares: right eye. Marks are spread up and down where values are equal. Pre-registered at 09:42:54 UTC on 6 Oct 2026, before the first trial; the 5 Oct trials of the real map and the plain scramble are included as they were.
Read this before sharing. One model (Shiu et al. LIF on FlyWire v783), brain only, every LC4/LPLC2 of one eye at 80 Hz for 1 s; descending-neuron rates, not a moving fly; not a real fly.
- Brain level only. No fly walked in these trials. Seven of the 22 new trials sent drives that differ from the no-brain floor (the two new real trials, both giant-fibre-off trials, one inside-only trial and two off-band gain-search trials); they walked in the body on 8 Oct 2026 (the real-map runs turned away; with the giant fibre silenced the fly still turned away; the inside-only drive did not turn; the two off-band gain-search drives, shown for display only, did not turn either; chart). The other 15, including all 3 matched turned-up trials and the 3 new plain scrambles, sent exactly the floor's drive (same sha256), so their body result is the floor's walk by determinism.
- Small n. Real map 6 trials, plain scramble 6 (3 shuffles), turned up 3 matched trials (3 shuffles, one gain; plus 6 search trials), inside-only 6 (3 shuffles), giant fibre silenced 2. Only the left eye was used for the turned-up arm, as pre-registered; its activity rises steeply with gain (about 9,000 to more than 22,000 spikes between gains 2.0 and 2.5).
- A faster input. The 6 Oct trials fed the 80 Hz input through one Poisson group instead of one input per neuron, which cut a trial from about 165 s to about 62 s. It is statistically equivalent, not bit-exact: it passed every criterion of its pre-registered check (gate V) on both eyes, and the plain scrambles agree with 5 Oct in both modes.
- The direction. "Turn command" is the far-side minus the near-side DNa02 rate, the side convention of our 5 Oct test; "away" rests on our hand-made mapping and on fly67's claim. A primary source for "away" in real flies, Card and Dickinson 2008, covers take-off direction only, not walking turns or DNa02.
- Partly expected from the map. The looming neurons have no direct synapses onto DNa02, so an inside-only scramble that breaks every relay was likely to remove the turn command; what the trials add is that no other route rebuilt it, and that the escape signal survived only on the side with direct synapses.
Open the share card (SVG): the sentence, the numbers, the caveats and the date on one image.
- Pre-registered Scrambled and turned up to the real activity.
We turned the scrambled maps up until they were as busy as the real one. They still did not produce the turn: the wiring, not the amount of activity, makes it.
Reading by the rule: "the turn needs the wiring even at the same activity": 3 shuffles matched the real activity (gain 2.25, outside-activity 0.98 × real, whole brain 27,166–27,515 spikes vs 27,686), 0 of them produced the turn command. - Pre-registered Scrambled inside only (every edge onto descending, motor and endocrine neurons and out of sensory and ascending neurons kept; the rest shuffled).
Scrambling only the inside of the map kept about -9% of the turn command and 69% of the escape signal.
In plain words, −9% means none (the far-side command was absent in all 6); it is slightly negative because in one trial the near-side DNa02 fired 1 Hz. Readings by the rules: turn command "the turn needs the wiring in between" (fly67's DNa02 rule passed in 0 of 6); escape signal "partly" (giant-fibre retention 0.69, below the 0.9 bar for "rides on direct connections"; above 50 Hz in 6 of 6).Descriptive, not pre-registered Per neuron, the kept part is one side: the near-side giant fibre, which gets 832 (left) and 1,053 (right) direct synapses from the looming neurons, kept firing at 151–169 Hz (real map: 129 and 149–156 Hz); the far-side giant fibre went to 0 Hz (real map: 84–104 Hz).
- Pre-registered Giant fibre silenced (every synapse out of both giant fibres, DNp01, set to zero; the fibres themselves still spike).
With the escape neuron silenced, the turn command stayed.
Reading: "steering survives silencing the giant fibre" (fly67's rule passed on both sides; far-side DNa02 changed by −2 Hz and 0 Hz against the paired real trials). This was expected from the wiring: the map has no giant fibre → DNa02 connections at all, so the test checks fly67's claim in our engine, not a hidden route. - Pre-registered Plain scramble (same number of connections per neuron): 6 trials over 3 shuffles, every readout 0 Hz, and the brain about 14 times quieter outside the stimulated neurons (0.07 × real).
Numbers behind the chart (all 29 brain trials)
| Trial | Run and input | Arm | Shadow on | Shuffle | Gain | Giant fibre L / R | DNa02 L / R | Turn command | LI | fly67's rule | Whole-brain spikes | Outside activity (× real) | Body |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| LOOML-real-0 | 5 Oct, one input per neuron | Real map | left eye | – | – | 129 / 94 | 0 / 23 | 23 | 0.96 | pass | 27,367 | 0.98 | walked on 5 Oct |
| LOOML-real-1 | 5 Oct, one input per neuron | Real map | left eye | – | – | 129 / 99 | 0 / 28 | 28 | 0.97 | pass | 27,538 | 0.99 | walked on 5 Oct |
| LOOML-real-2-f | 6 Oct, fast input | Real map | left eye | – | – | 129 / 104 | 0 / 25 | 25 | 0.96 | pass | 28,153 | 1.03 | walked on 8 Oct: turned away (+37.2° beyond the floor) |
| LOOMR-real-0 | 5 Oct, one input per neuron | Real map | right eye | – | – | 89 / 149 | 44 / 0 | 44 | 0.98 | pass | 27,701 | 1.00 | walked on 5 Oct |
| LOOMR-real-1 | 5 Oct, one input per neuron | Real map | right eye | – | – | 85 / 156 | 44 / 0 | 44 | 0.98 | pass | 27,509 | 1.00 | walked on 5 Oct |
| LOOMR-real-2-f | 6 Oct, fast input | Real map | right eye | – | – | 84 / 149 | 41 / 0 | 41 | 0.98 | pass | 27,693 | 1.00 | walked on 8 Oct: turned away (+55.4° beyond the floor) |
| LOOML-real-2-gfoff-f | 6 Oct, fast input | Giant fibre silenced | left eye | – | – | 131 / 95 | 0 / 23 | 23 | 0.96 | pass | 27,541 | 0.99 | walked on 8 Oct: turned away (+31.5° beyond the floor) |
| LOOMR-real-2-gfoff-f | 6 Oct, fast input | Giant fibre silenced | right eye | – | – | 84 / 151 | 41 / 0 | 41 | 0.98 | pass | 27,607 | 0.99 | walked on 8 Oct: turned away (+54.8° beyond the floor) |
| LOOML-D-3 | 5 Oct, one input per neuron | Scrambled | left eye | 3 | – | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 13,993 | 0.07 | same drive as the floor (5 Oct) |
| LOOML-D-4 | 5 Oct, one input per neuron | Scrambled | left eye | 4 | – | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 13,938 | 0.07 | same drive as the floor (5 Oct) |
| LOOML-D-5-f | 6 Oct, fast input | Scrambled | left eye | 5 | – | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 13,872 | 0.07 | same drive as the floor |
| LOOMR-D-3 | 5 Oct, one input per neuron | Scrambled | right eye | 3 | – | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 13,134 | 0.06 | same drive as the floor (5 Oct) |
| LOOMR-D-4-f | 6 Oct, fast input | Scrambled | right eye | 4 | – | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 13,152 | 0.07 | same drive as the floor |
| LOOMR-D-5-f | 6 Oct, fast input | Scrambled | right eye | 5 | – | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 13,149 | 0.07 | same drive as the floor |
| LOOML-G-3-g2-f | 6 Oct, fast input | Turned up (search) | left eye | 3 | 2 | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 22,352 | 0.64 | same drive as the floor |
| LOOML-G-4-g2-f | 6 Oct, fast input | Turned up (search) | left eye | 4 | 2 | 1 / 0 | 0 / 0 | 0 | 0.00 | fail | 21,412 | 0.58 | same drive as the floor |
| LOOML-G-5-g2-f | 6 Oct, fast input | Turned up (search) | left eye | 5 | 2 | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 21,967 | 0.62 | same drive as the floor |
| LOOML-G-3-g2.25-f | 6 Oct, fast input | Turned up (matched) | left eye | 3 | 2.25 | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 27,515 | 0.99 | same drive as the floor |
| LOOML-G-4-g2.25-f | 6 Oct, fast input | Turned up (matched) | left eye | 4 | 2.25 | 1 / 0 | 0 / 0 | 0 | 0.00 | fail | 27,252 | 0.98 | same drive as the floor |
| LOOML-G-5-g2.25-f | 6 Oct, fast input | Turned up (matched) | left eye | 5 | 2.25 | 0 / 1 | 0 / 0 | 0 | 0.00 | fail | 27,166 | 0.97 | same drive as the floor |
| LOOML-G-3-g2.5-f | 6 Oct, fast input | Turned up (search) | left eye | 3 | 2.5 | 0 / 0 | 0 / 0 | 0 | 0.00 | fail | 35,282 | 1.52 | same drive as the floor |
| LOOML-G-4-g2.5-f | 6 Oct, fast input | Turned up (search) | left eye | 4 | 2.5 | 3 / 0 | 3 / 0 | -3 | -0.75 | fail | 36,988 | 1.63 | walked on 8 Oct: no turn (−4.0° beyond the floor) |
| LOOML-G-5-g2.5-f | 6 Oct, fast input | Turned up (search) | left eye | 5 | 2.5 | 0 / 0 | 0 / 1 | 1 | 0.50 | fail | 48,682 | 2.42 | walked on 8 Oct: no turn (+0.2° beyond the floor) |
| LOOML-B-301-f | 6 Oct, fast input | Inside only | left eye | 301 | – | 153 / 0 | 0 / 0 | 0 | 0.00 | fail | 17,949 | 0.35 | same drive as the floor |
| LOOML-B-302-f | 6 Oct, fast input | Inside only | left eye | 302 | – | 151 / 0 | 0 / 0 | 0 | 0.00 | fail | 17,711 | 0.33 | same drive as the floor |
| LOOML-B-303-f | 6 Oct, fast input | Inside only | left eye | 303 | – | 153 / 0 | 0 / 0 | 0 | 0.00 | fail | 18,195 | 0.36 | same drive as the floor |
| LOOMR-B-301-f | 6 Oct, fast input | Inside only | right eye | 301 | – | 0 / 169 | 0 / 1 | -1 | -0.50 | fail | 17,282 | 0.34 | walked on 8 Oct: no turn (−1.4° beyond the floor) |
| LOOMR-B-302-f | 6 Oct, fast input | Inside only | right eye | 302 | – | 0 / 169 | 0 / 0 | 0 | 0.00 | fail | 18,227 | 0.40 | same drive as the floor |
| LOOMR-B-303-f | 6 Oct, fast input | Inside only | right eye | 303 | – | 0 / 167 | 0 / 0 | 0 | 0.00 | fail | 20,415 | 0.54 | same drive as the floor |
fly67's own committed numbers (the author's JavaScript engine, not ours, 45f62fd9): left loom giant fibre 114.8 Hz, DNa02 0 / 25; right loom 120.3 Hz, 43 / 0. Gate V for the fast input: giant fibre 116.5 Hz on both eyes (band 94.8–138.6), LI 0.96 and 0.98 (≥ 0.90), stimulated rate 80.1 and 80.3 Hz, whole brain 28,153 and 27,693 spikes. Trials took 55–73 s (mean 61.8 s, 22 trials) against 152–177 s with the 5 Oct input.
What the map's own counts already said
Before the first trial we counted where the looming neurons' synapses go in the map. They reach the turning neurons DNa02 only through relays, the strongest being PVLP141 on the eye's side, while they synapse directly onto the giant fibre of their own side. That is why the inside-only scramble could break the route to DNa02 while the direct route to the giant fibre stayed, and why silencing the giant fibre could not touch the turn: the map has no giant fibre → DNa02 connections.
What it shows
- Not just a quieter brain. Scrambled maps turned up to the real map's activity (0.98 × outside the stimulated cells, and the same whole-brain spike count) still sent no turn command in 3 of 3. In this model the turn needs the wiring, not only the amount of activity.
- The turn runs through the inside of the map. "Scrambling only the inside of the map kept about -9% of the turn command and 69% of the escape signal." The −9% means none: the far-side command was absent in all 6 trials.
- Steering does not need the escape neuron in this map: with the giant fibre's outputs silenced, the turn command stayed (23 and 41 Hz).
What it does not show
- A moving fly. These are spike counts of descending neurons in one model over 1 s. No body ran in these trials; the drives that differed from the no-brain floor walked on 8 Oct (above).
- A surprise about the escape neuron. The map has no giant fibre → DNa02 connections, so "steering survives" was expected; it checks fly67's claim in our engine.
- A real fly. One model on one fly's map, with a strong artificial stimulus; "away" rests on our mapping and fly67's claim. The top relay candidate, PVLP141, was silenced on 8 Oct (below).
Share it with its caveats: the share card (SVG). Every trial, the gain search and the deviations: report of 6 Oct 2026 · the ledger row: Does fly wiring help?
How strong must the shadow be?
Added 8 Oct 2026 · made by us · brain only · pre-registered at 09:45:48 UTC and re-stamped with the deviations at 09:57:49, before the first brain trial · it changes no grade
Our looming tests drove every looming neuron of one eye at 80 spikes per second, a strong input. A common objection: with that much input, any wiring that connects the eye to the turning neurons would turn. So we lowered the input step by step, to 60, 40, 30, 20 and 10 spikes per second, and asked where the turn command starts, whether the escape neuron starts before or after it, and whether a second random seed agrees. Then we silenced the relay neurons that the map's own counts point to.
We lowered the input to the looming neurons step by step, from 80 to 10 spikes per second each. The turn command appeared from about 30 per second (38% of our strongest input) on both sides; below that the model did not turn. The escape neuron started firing at the same input as the turn command.
One model (Shiu et al. LIF on FlyWire v783), brain only; the input is a rate we set for every LC4/LPLC2 of one eye, not a measured response to a real shadow; 'silenced' means the neuron's output synapses were removed; not a real fly. We found no measured real-fly LC4/LPLC2 firing rate during looming to compare these rates with (abstract search, 8 Oct 2026).
"Escape" here is the pre-registered line: the giant fibre's mean rate above 50 Hz (at 20 per second it fired 42 and 47.5 Hz, below the line). Seed check: repeated with a second random seed at 20 and 30 per second on both eyes, with the same pass and fail (pre-registered reading: "stable across seeds"). The 80 per second points are our 6 Oct trials with the same seed. Pre-registered at 09:45:48 UTC on 8 Oct 2026 and re-stamped with the deviations at 09:57:49, before the first brain trial.
Open the share card (SVG): the sentence, the threshold, the caveat and the date on one image.
Descriptive, pre-registered as optional, n 1 per eye We also walked the threshold drives in the body: At 30 spikes per second per looming neuron, the fly turned away by 11° (left eye) and 17° (right eye).
Both above the 10° line of our body test (11.1° and 16.8° beyond the no-brain floor, which repeated our 5 Oct floor run exactly in this environment). One run per eye; not in the ledger.
Numbers behind the chart (all 16 brain trials)
| Trial | Shadow on | Input (per second) | Seed | Giant fibre L / R | Giant fibre mean | DNa02 L / R | Turn command | LI | Turn rule | Escape above 50 Hz | Whole-brain spikes |
|---|---|---|---|---|---|---|---|---|---|---|---|
| LOOML-real-0-r20-f | left eye | 20 | 0 | 39 / 35 | 37 | 0 / 0 | 0 | 0.00 | fail | no | 6,377 |
| LOOML-real-0-r30-f | left eye | 30 | 0 | 74 / 68 | 71 | 0 / 9 | 9 | 0.90 | pass | yes | 10,357 |
| LOOML-real-2-r10-f | left eye | 10 | 2 | 7 / 7 | 7 | 0 / 0 | 0 | 0.00 | fail | no | 2,547 |
| LOOML-real-2-r20-f | left eye | 20 | 2 | 47 / 37 | 42 | 0 / 1 | 1 | 0.50 | fail | no | 6,339 |
| LOOML-real-2-r30-f | left eye | 30 | 2 | 71 / 65 | 68 | 0 / 8 | 8 | 0.89 | pass | yes | 10,073 |
| LOOML-real-2-r40-f | left eye | 40 | 2 | 91 / 82 | 86.5 | 0 / 20 | 20 | 0.95 | pass | yes | 13,770 |
| LOOML-real-2-r60-f | left eye | 60 | 2 | 117 / 94 | 105.5 | 0 / 20 | 20 | 0.95 | pass | yes | 21,165 |
| LOOML-real-2-f | left eye | 80 | 2 (6 Oct) | 129 / 104 | 116.5 | 0 / 25 | 25 | 0.96 | pass | yes | 28,153 |
| LOOMR-real-0-r20-f | right eye | 20 | 0 | 33 / 64 | 48.5 | 2 / 0 | 2 | 0.67 | fail | no | 6,565 |
| LOOMR-real-0-r30-f | right eye | 30 | 0 | 52 / 92 | 72 | 6 / 0 | 6 | 0.86 | pass | yes | 10,457 |
| LOOMR-real-2-r10-f | right eye | 10 | 2 | 19 / 33 | 26 | 0 / 0 | 0 | 0.00 | fail | no | 2,778 |
| LOOMR-real-2-r20-f | right eye | 20 | 2 | 36 / 59 | 47.5 | 0 / 0 | 0 | 0.00 | fail | no | 6,315 |
| LOOMR-real-2-r30-f | right eye | 30 | 2 | 52 / 90 | 71 | 12 / 0 | 12 | 0.92 | pass | yes | 10,429 |
| LOOMR-real-2-r40-f | right eye | 40 | 2 | 66 / 115 | 90.5 | 24 / 0 | 24 | 0.96 | pass | yes | 14,257 |
| LOOMR-real-2-r60-f | right eye | 60 | 2 | 82 / 136 | 109 | 39 / 0 | 39 | 0.97 | pass | yes | 21,637 |
| LOOMR-real-2-f | right eye | 80 | 2 (6 Oct) | 84 / 149 | 116.5 | 41 / 0 | 41 | 0.98 | pass | yes | 27,693 |
Thresholds by the pre-registered definition (the lowest tested input at which the rule passes there and at every higher tested input): turn 30 and 30, escape 30 and 30 per second (left, right eye); no pass below a threshold. Readings: strength "moderate" (threshold 40 or less on both eyes), escape order "together", seed check "stable across seeds". Whole-brain spikes rose with the input, from 2,547 (10 per second, left eye) to 28,153 (80, left eye). Trials took 57–66 s each.
14 Oct, updated 17, 18, 19 and 21 Oct: the looming turn in degrees, at each input strength
Our own test · pre-registered before the first body run (14 Oct 2026, 09:28 UTC) · simulated body (FlyGym) driven through our own mapping · one model · a simulation, not an animal
On 8 Oct the brain model's turn command appeared from about 30 spikes per second per looming neuron. On 14 Oct we walked the remaining brain drives of that series in the simulated body, so every drive that differs from the no-brain floor now has a body run. The sentences were chosen by rules fixed in advance; the two eyes read differently, so there is one sentence per eye.
Left eye (pre-registered sentence): In the body, the fly turned away from 30 spikes/s per looming neuron upward on this side, the strength at which the brain's turn command appeared; the turn was about 11 deg at 30 and about 37 deg at 80 (our mapping turns the brain's rates into steering).
Right eye (pre-registered sentence): At 30 spikes/s the turn was borderline: the fly turned away in 1 of 2 runs on this side; from 40 upward it turned away in every run.
This borderline rests on one run: seed 0 at 30 spikes/s turned 9.18° beyond the no-brain floor, 0.8° under our 10° rule (seed 2 at 30 turned 16.8°).
These body runs turn the brain's firing rates into steering through our own hand-made mapping, so they show the brain result in degrees; they are not a separate test of the wiring. We ran scrambled brains at 30 and 80 spikes/s, not at 20, 40 or 60.
17 Oct 2026: scrambled brains at 30 spikes/s (pre-registered sentences; brain only, because their steering drives equal the no-brain drive):
17 Oct: at 30 spikes/s, the weakest tested input at which the real brain's turn command appeared, we also ran 2 scrambled brains per side (each neuron keeps its number of connections; one 1-second trial each, with the same trial seed as the real-wiring body runs).
- Left eye:
On the left side, neither scrambled brain sent a turn command, and their steering signals were identical to having no brain, so in the body they walk like the no-brain fly (the real wiring turned it about 11 deg).
- Right eye:
On the right side, neither scrambled brain sent a turn command, and their steering signals were identical to having no brain, so in the body they walk like the no-brain fly (the real wiring turned it about 17 deg).
The scrambled brains were not matched for overall activity: they fired 48% (left) and 44% (right) as many spikes as the real brain at this strength. Most of those spikes were the stimulated input neurons themselves; outside the input, the scrambled brains fired about 1% as many spikes as the real brain.
18 Oct 2026: the same scrambled brains, strengthened (pre-registered sentences; brain only):
18 Oct: we then strengthened the connections inside scrambled brains (not the input) until their neurons outside the stimulated input fired about as many spikes as in the real brain (within 30%; same input, one 1-second trial each).
- Left eye:
On the left side, 2 scrambled brains matched this way, and neither sent a turn command.
- Right eye:
On the right side, 2 scrambled brains matched this way, and neither sent a turn command.
Matching the number of spikes does not match the pattern: 1,152 to 2,228 neurons fired in the matched scrambled brains, against 379 (left) and 432 (right) in the real brain.
19 Oct 2026, in the body (pre-registered; one run): 19 Oct: walked in the body, the one matched scrambled brain with a small steering signal changed the fly's heading by about 1.4 degrees away from the shadow compared with no brain input, under our 10-degree turn rule; the real brain, with the same input, turned it away by about 16.8 degrees (one run each).
This body run turns the brain's firing rates into steering through our own hand-made mapping; it is not a separate test of the wiring.
As at 80 spikes/s, none of the scrambled brains we ran at 30 spikes/s sent a turn command, including those strengthened to about the real brain's activity outside the input.
21 Oct 2026: a second input draw (pre-registered sentences; the same 4 scrambled brains, one 1-second brain trial each, then one body run):
21 Oct: we repeated the matched test with a second, independent draw of the same input on the same scrambled brains (one 1-second trial each).
- Left eye:
On the left side, both scrambled brains matched again, and neither sent a turn command.
- Right eye:
On the right side, both scrambled brains matched again, and neither sent a turn command.
Matching the number of spikes still did not match the pattern: 1,256 to 2,741 neurons fired in the matched scrambled brains, against 427 (left) and 446 (right) in the real brain with this input.
With this input the real brain's turn signal on the right side was 6 spikes/s against our threshold of 5, so the real brain itself sits close to our turn rule at this stimulus strength.
21 Oct 2026, in the body (pre-registered; one run): Walked in the body, the matched scrambled brain with a small steering signal changed the fly's heading by about 1.3 degrees away from the shadow compared with no brain input, under our 10-degree turn rule; the real brain, with the same input, turned it away by about 9.2 degrees (one run each).
With this input the real brain's own turn was also under our 10-degree rule.
This body run turns the brain's firing rates into steering through our own hand-made mapping; it is not a separate test of the wiring.
With a second input draw, the matched scrambled brains again sent no turn command (2 input draws on the same 4 scrambled brains).
- seed 2 (line joins its points)
- seed 0
- same as no brain
- scrambled brain, 30 spikes/s (17 Oct)
- scrambled, gain-matched, 30 spikes/s, right eye (walked 19 Oct)
- the same brain, second input draw (walked 21 Oct)
- 10° rule
Horizontal axis: input in spikes per second for every looming neuron of one eye. Vertical axis: degrees turned away beyond the no-brain floor. Each point is one body run: how far the simulated fly turned away from the shadow after it appeared, minus the turn of the same body with no brain input (the floor). Below the dashed line our rule reads "did not turn away". Hollow diamonds at 30 spikes/s: the 17 Oct scrambled brains, drawn at the no-brain value because their drives equal the no-brain drive. Filled triangle (right eye only): the one strengthened scrambled brain with a small steering signal, walked on 19 Oct (1.39°, below the rule; the real brain at the same input 16.78°). Hollow triangle: the same brain with a second input draw, walked on 21 Oct (1.30°, below the rule; the real brain with that input is the seed 0 point at 30, 9.18°, also below the rule). The turn did not grow at every step: seed 2 at 30, 40, 60 and 80 spikes/s turned 11.1, 27.2, 26.9, 37.2° (left eye) and 16.8, 33.8, 62.2, 55.4° (right eye).
Every point in the charts (16 body readings, 4 scrambled brains and 1 strengthened scrambled brain walked on 19 Oct and, with a second input draw, on 21 Oct)
| Shadow on | Input (spikes/s) | Brain and seed | Turn beyond the floor | Turned away (10° rule) | Body run |
|---|---|---|---|---|---|
| left eye | 10 | real, seed 2 | 0° | no | same as no brain |
| left eye | 20 | real, seed 2 | 1.67° | no | walked 14 Oct 2026 |
| left eye | 20 | real, seed 0 | 0° | no | same as no brain |
| left eye | 30 | real, seed 2 | 11.13° | yes | walked 8 Oct 2026 |
| left eye | 30 | real, seed 0 | 12.68° | yes | walked 14 Oct 2026 |
| left eye | 30 | scrambled (shuffle 4), trial seed 2 | 0° | no | same as no brain (brain trial 17 Oct 2026) |
| left eye | 30 | scrambled (shuffle 5), trial seed 2 | 0° | no | same as no brain (brain trial 17 Oct 2026) |
| left eye | 40 | real, seed 2 | 27.18° | yes | walked 14 Oct 2026 |
| left eye | 60 | real, seed 2 | 26.88° | yes | walked 14 Oct 2026 |
| left eye | 80 | real, seed 2 | 37.18° | yes | walked 8 Oct 2026 |
| right eye | 10 | real, seed 2 | 0° | no | same as no brain |
| right eye | 20 | real, seed 2 | 0° | no | same as no brain |
| right eye | 20 | real, seed 0 | 2.52° | no | walked 14 Oct 2026 |
| right eye | 30 | real, seed 2 | 16.78° | yes | walked 8 Oct 2026 |
| right eye | 30 | real, seed 0 | 9.18° | no | walked 14 Oct 2026 |
| right eye | 30 | scrambled (shuffle 4), trial seed 2 | 0° | no | same as no brain (brain trial 17 Oct 2026) |
| right eye | 30 | scrambled (shuffle 5), trial seed 2 | 0° | no | same as no brain (brain trial 17 Oct 2026) |
| right eye | 40 | real, seed 2 | 33.78° | yes | walked 14 Oct 2026 |
| right eye | 60 | real, seed 2 | 62.15° | yes | walked 14 Oct 2026 |
| right eye | 80 | real, seed 2 | 55.42° | yes | walked 8 Oct 2026 |
| right eye | 30 | scrambled, gain-matched (shuffle 4, connections ×2.75), trial seed 2 | 1.39° | no | walked 19 Oct 2026 |
| right eye | 30 | scrambled, gain-matched (shuffle 4, connections ×2.75), trial seed 0 (second input draw) | 1.30° | no | walked 21 Oct 2026 |
- One run per drive. Seed 2 was walked at each strength; seed 0 only at 20 and 30 spikes/s. The right-eye reading turns on one run 0.8° under the line.
- Checks before reading. The no-brain floor repeated our 8 Oct floor run exactly (0 of 700 trajectory values differ), and every brain drive matched its saved fingerprint (12 of 12), so these runs read the same drives as the brain series. Before the 17 Oct scrambled brains, our 8 Oct real-brain trial at 30 spikes/s was repeated exactly (spike counts, active neurons and every descending-neuron rate), and each scrambled graph was the same graph as its 80 spikes/s twin.
- Scrambled brains at 30 spikes/s. Two shuffles per side (shuffle seeds 4 and 5), one 1-second trial each. In all 4 trials of 17 Oct the turn neuron on both sides (DNa02), both giant fibres (DNp01) and MDN fired 0 Hz. On 18 Oct we multiplied the connection strengths of the same 4 graphs by 2.5 to 2.75 (6 trials in all) until each brain's spikes outside the stimulated input were within 30% of the real brain's (left 5,266, right 5,901); the matched counts were 3,750 and 3,991 on the left and 6,370 and 4,587 on the right. No scrambled brain was run at 20, 40 or 60 spikes/s.
- Limits of the 18 Oct check. One 1-second trial per scrambled brain and one trial seed. Two of the four matches sit near the lower edge of the band (0.71 and 0.76 of the real count, both on the left). Matching the count did not match the pattern: about 3 to 6 times as many neurons fired as in the real brain. One matched brain (right eye, shuffle 4) sent a small steering signal (far-side turn neuron 1 Hz in one 100 ms bin); on 19 Oct we walked it in the body (next item).
- The 19 Oct body walk. One deterministic body run of that brain's steering drive, with the same body, floor and mapping as the real-brain runs; the no-brain floor repeated exactly first (0 of 700 values differ), and the drive was regenerated and matched its saved fingerprint. The 1.39° is the turn beyond the no-brain floor (−1.49°); the real brain at the same input turned 16.78° (8 Oct body run). One run, one trial seed, one scrambled graph.
- The 21 Oct second input draw. The same 4 scrambled graphs (each checked to be the same graph as on 18 Oct), with a second, independent draw of the same 30 spikes/s input (trial seed 0), one 1-second trial each; all 4 matched at connections ×2.75 in 5 trials (left 4,103 and 5,816, right 5,960 and 7,568 spikes outside the input, against the real brain's 5,465 and 5,870 with this draw). Two matches sit near the band edges (0.75 and 1.29 of the real count). Before the trials, the real brain's right-eye trial with this input was repeated and equalled our 8 Oct file on every compared count. One matched brain (right eye, shuffle 4) again sent a small steering signal (far-side turn neuron 1 Hz); walked in the body it turned 1.30° beyond the floor, and the real brain with this input turned 9.18°, so at this input our 10° body rule does not separate them; only the brain-level turn signal does (6 spikes/s against our threshold of 5 in the real brain, at most 1 in the scrambled ones). The match is on spike count only, over 2 input draws on the same 4 graphs.
- What it measures. The heading of a simulated FlyGym body over 1.2 s, steered by the brain model's turning-neuron rates through our own hand-made mapping, with the walking drive held by us. The 30 and 80 spikes/s seed 2 points are our 8 Oct body runs and the 4 grey squares take the no-brain walk, as do the 4 diamonds (the 17 Oct scrambled brains); the other 8 points are 14 Oct runs (the remaining 4 runs of 14 Oct are the relay lines on this page).
Every run, gates and timings: report of 14 Oct 2026; the scrambled brains at 30 spikes/s: report of 17 Oct 2026 and, strengthened, report of 18 Oct 2026; the 19 Oct body walk: report of 19 Oct 2026; the second input draw: report of 21 Oct 2026; the brain series: How strong must the shadow be?
What it shows
- A threshold, not a smooth fade. At 10 and 20 spikes per second the turn rule failed on both eyes; from 30 up it passed on both, and the turn command grew with the input (8 to 25 Hz on the left eye, 12 to 41 Hz on the right).
- Escape and turn start together. "The escape neuron started firing at the same input as the turn command." Here "started firing" means above the pre-registered 50 Hz line.
- Not a fluke of one seed. A second random seed gave the same pass and fail at 20 and 30 per second on both eyes.
What it does not show
- How strong a real shadow is. The input is a rate we set for every LC4/LPLC2 of one eye, not a measured response to a real shadow. We found no measured real-fly LC4/LPLC2 firing rate to compare against, so "38% of our strongest input" is a statement about this model, not about flies.
- A moving fly. These are descending-neuron rates in one model over 1 s. Two threshold drives also walked in the body (one run per eye, descriptive, below the chart).
- A real fly. One model (Shiu et al. LIF on FlyWire v783), brain only. The body and the brain stepping together, with the eye seeing the fly's own movement, is not yet tested.
13 Oct, updated 15 Oct: the relays against an activity-matched control
Our own test · pre-registered before the first trial (13 Oct 2026, 10:18 UTC; the 5 added right-side controls on 15 Oct 2026, 09:38 UTC) · brain only · one model · not a real fly
The pre-registered sentence, chosen by the rule fixed in advance:
The two sides disagree: on one side silencing the relays lowered the turn neuron's rate more than matched activity elsewhere, on the other it did not. We do not draw a conclusion.
15 Oct 2026 (pre-registered update line): 15 Oct: we ran 5 more matched controls on the right side, 10 in all. Every one of them left the turn neuron's rate higher than silencing the relays did (controls 38-47 Hz, relays 25.7 Hz), so the right-side reading did not change.
Our rule: Our rule compares the relays with the lowest matched control, so extra controls can only make the relays look less special, never more.
The 15 Oct trials and checks
A stricter control, counted (17 Oct 2026): We also looked for a stricter control: other active inputs that excite the turn neuron with a similar firing rate and a similar number of connections onto it. In this model there are none for any of the active relay neurons on either side (count of 15 Oct: the busiest relay, PVLP141, fires about 100 spikes/s; no other such input fires above 43), so that stricter test cannot be built here.
23 Oct 2026 (pre-registered before the trials; brain only): 23 Oct: our earlier controls matched each relay neuron's activity, but not how strongly it connects to the turn neuron. This time the control is a group of other active neurons that excite the turn neuron, chosen so that together they fire about as much as the active relays and send about as many synapses to the turn neuron (both within 20%).
Left eye: in 1 of 5 matched groups the turn neuron's rate fell as low as, or lower than, with the relays silenced (matched groups 13-31 Hz, relays 16.0 Hz, nothing silenced 21-25 Hz). By our pre-registered rule, this test does not show that the relays matter more than other inputs with the same combined connection strength and activity.
Right eye: among the 9 other active neurons that excite the turn neuron, no group came within 20% of the relays on both combined activity and synapse count (we had set 5 groups as the minimum in advance), so this control could not be built for this eye.
- Our rule:
Our rule compares the relays with the lowest matched group, so extra groups can only make the relays look less special, never more.
Left eye: outside the stimulated looming neurons, the brain fired 13,581-14,040 spikes with a matched group silenced, against 13,901-14,760 with the relays silenced and 14,502-15,172 with nothing silenced.
- Limits:
One model, brain only, one stimulus strength (80 spikes/s), 5 matched groups on the left eye only, each run once; the 5 groups all share the same three most active neurons, so they are not independent draws; not a real fly.
In the body, 14 Oct 2026 (descriptive, not a reading of the relay question above; shadow at 80 spikes/s; our own mapping):
- Left eye:
With PVLP141 silenced, the fly still turned away, by about 24 deg instead of 37 deg (one run per side; our mapping).
- Left eye:
With all 21-23 two-step relay neurons silenced, the fly still turned away, by about 37 deg instead of 37 deg (one run per side; our mapping).
Unrounded: 37.3° vs 37.2°. - Right eye:
With PVLP141 silenced, the fly still turned away, by about 34 deg instead of 55 deg (one run per side; our mapping).
- Right eye:
With all 21-23 two-step relay neurons silenced, the fly still turned away, by about 40 deg instead of 55 deg (one run per side; our mapping).
These body runs turn the brain's firing rates into steering through our own hand-made mapping, so they show the brain result in degrees; they are not a separate test of the wiring. We ran scrambled brains at 30 and 80 spikes/s, not at 20, 40 or 60.
What is new. On 9 Oct we silenced the relay neurons between the looming-sensitive cells and the turn neuron (DNa02) and compared that with silencing randomly picked active neurons. That random set removed 2-6 times as much activity as the relays (2.3 times for a shadow on the left eye, 5.8 times on the right), so the two were not like for like. This time each silenced relay neuron got its own partner outside the relays that fired about as often in the same trial (within 20% of its spike count), and the control silences those partners instead. Both arms of the comparison now remove about the same amount of activity.
Blue strip: the turn neuron's rate with nothing silenced (5 seeds). Dashed box and open circles: the matched-control draws (5 on the left; 10 on the right, the 5 added on 15 Oct with a thicker outline). Bottom row (23 Oct): the matched-group draws, shaded circles, left eye only; on the right eye no matched group could be built. Filled dots: the relay arms, one per seed; the bar is their mean. By the pre-registered rule a relay value below both the normal strip and every matched draw reads "fell more than matched activity"; below the strip but inside the matched draws it reads "fell, like matched activity". Left eye: E2 16.0 Hz, "fell, like matched activity". Right eye: E2 25.7 Hz, "fell more than matched activity". PVLP141 alone reads the same as E2 on each side.
| Shadow on | Nothing silenced (5 seeds) | Matched control (5 draws left, 10 right) | Relays E2 (3 seeds) | PVLP141 alone (3 seeds) | Matched group (23 Oct, 5 draws) |
|---|---|---|---|---|---|
| Left eye (far side: right DNa02) | 24, 25, 25, 25, 21 band 21-25 | 19, 24, 19, 34, 13 band 13-34 | 15, 14, 19: mean 16.0 (about 67% of normal) fell, like matched activity | 19, 17, 17: mean 17.7 (about 74%) fell, like matched activity | 22, 22, 13, 23, 31 band 13-31; 1 of 5 at or below the relays |
| Right eye (far side: left DNa02) | 43, 44, 41, 40, 49 band 40-49 | 40, 41, 41, 43, 40, 43, 47, 38, 47, 46 band 38-47; the last 5 added 15 Oct | 26, 25, 26: mean 25.7 (about 59% of normal) fell more than matched activity | 26, 31, 24: mean 27.0 (about 62%) fell more than matched activity | could not be built 0 groups among 9 neurons |
- Small samples. 5 seeds per side; 5 matched draws on the left and 10 on the right; the relay arms have 3 seeds each. The left side's matched draws spread widely (13-34 Hz: three below the normal strip, one above it).
- Some matched partners touch the turn neuron. The matching rule excludes the two-step relays, not every input to the turn neuron: 1 of 5 left-side draws and 2 of 10 right-side draws (none of the 5 added on 15 Oct) include a neuron that synapses directly onto the far-side DNa02.
- E2 is mostly PVLP141. Of the activity removed by silencing the excitatory relays (E2), PVLP141 accounts for 102 of 149 spikes (left) and 103 of 121 (right). E2 and PVLP141 alone read the same on each side, so this test says little about the other relay neurons.
- What it measures. A simulated firing rate in one brain-only model (Shiu et al. LIF on FlyWire v783) with a looming input we set; not fly behaviour and not a real fly. Percentages are a value divided by the mean with nothing silenced.
The 9 Oct line stays, with a note. On 9 Oct the pre-registered summary was: Silencing other active neurons also changed the turn command, so this relay result is not specific to the relay neurons.
Note added 13 Oct 2026: the random control behind that line removed 2-6 times as much activity as the relays (2.3 times left, 5.8 times right); the activity-matched test above gives the reading on each side. Trials, deviations and code: report of 13 Oct 2026; the 9 Oct design: report of 9 Oct 2026.
Which relays carry the turn? Silencing them, read by firing rate (9 Oct)
The map has no direct synapses from the looming neurons to the turning neurons DNa02; its two-step routes to the far-side DNa02 run through 21 relay neurons (left eye) and 23 (right eye), and the top one by synapse count is PVLP141 on the same side as the eye (the census above). On 8 Oct we silenced PVLP141 alone, then all two-step relays of that side, at the full 80 spikes per second. On 9 Oct, pre-registered again before the first trial, we silenced PVLP141, the relays that excite the turn neuron and the relays that inhibit it separately, plus the same number of randomly picked active neurons as a control, and read how often the far-side turn neuron fired against three runs per eye with nothing silenced.
The honest headline: Silencing the relays lowered the turn command to about 60-70% of normal, but silencing random active neurons also moved it, so this is not yet specific to the relays.
In the pre-registered words: Silencing other active neurons also changed the turn command, so this relay result is not specific to the relay neurons.
- The normal range is narrow. For a shadow on the left eye it is only 1 Hz wide (24–25 Hz, three runs with nothing silenced); on the right it is 41–44 Hz. One random-neuron run at 26 Hz, just above the left range, is what triggered the summary above.
- The random control is one draw per eye (17 and 20 active neurons, one run each). More draws and more runs with nothing silenced are needed before any relay is called specific.
- What 8 Oct said, and how it reads now. On 8 Oct we read the same kind of test by the left/right contrast:
With one relay neuron (PVLP141) silenced, the turn command stayed (by the left/right contrast (8 Oct)).
With all 21-23 relay neurons between the looming neurons and the turn neuron silenced, the turn command stayed (by the left/right contrast (8 Oct)).
Read by firing rate instead:Read by firing rate, our 8 Oct test with all 21 (left eye) / 23 (right eye) relay neurons silenced raised it to about 105% of normal for a shadow on the left eye, and lowered it to about 65% of normal on the right (1 run per eye).
(the open marks below).
Silencing other active neurons also changed the turn command, so this relay result is not specific to the relay neurons.
Turn command = how often the turn neuron on the far side (DNa02) fires during the 1-second shadow. Normal range = our three runs per eye with nothing silenced.
- With one relay neuron (PVLP141) silenced, the turn command fell to about 70% (left eye) / 65% (right eye) of normal (3 runs per eye).
- With the 17 (left eye) / 20 (right eye) relay neurons that excite the turn neuron silenced, the turn command fell to about 65% (left eye) / 60% (right eye) of normal (3 runs per eye).
- With the 4 (left eye) / 3 (right eye) relay neurons that inhibit the turn neuron silenced, the turn command changed, but not clearly beyond its normal range for a shadow on the left eye, and rose to about 125% of normal on the right (3 runs per eye).
- With 17 (left eye) / 20 (right eye) other active neurons, picked at random, silenced, the turn command rose to about 105% of normal for a shadow on the left eye, and stayed within its normal range on the right (1 run per eye).
- Read by firing rate, our 8 Oct test with all 21 (left eye) / 23 (right eye) relay neurons silenced raised it to about 105% of normal for a shadow on the left eye, and lowered it to about 65% of normal on the right (1 run per eye).
One model (Shiu et al. LIF on FlyWire v783), brain only; the input is a rate we set for every LC4/LPLC2 of one eye, not a measured response to a real shadow; 'silenced' means the neuron's output synapses were removed; the normal range comes from 3 runs per eye; not a real fly.
Not shown in the chart: the excitatory relays without PVLP141 (16 left / 19 right, one run per eye: 30 and 41 Hz), descriptive only and never a reading. Pre-registered 2026-10-09T09:36:50Z, before the first trial; every run and the deviations: report of 9 Oct 2026.
Numbers behind the chart (30 brain runs)
| Trial file | Shadow on | Silenced | Seed | DNa02 far / near (Hz) | Reading (eye) | Giant fibre mean | MDN mean | Whole-brain spikes | Active neurons |
|---|---|---|---|---|---|---|---|---|---|
| LOOML-real-0-f.json | left eye | nothing | 0 | 24 / 0 | normal range 24–25 | 116.5 | 2.25 | 27,467 | 685 |
| LOOML-real-1-f.json | left eye | nothing | 1 | 25 / 0 | normal range 24–25 | 111.5 | 0 | 27,187 | 682 |
| LOOML-real-2-f.json | left eye | nothing | 2 | 25 / 0 | normal range 24–25 | 116.5 | 1.25 | 28,153 | 745 |
| LOOML-real-0-P-f.json | left eye | PVLP141 (1) | 0 | 19 / 0 | fell, ~70% | 116 | 0 | 27,413 | 747 |
| LOOML-real-1-P-f.json | left eye | PVLP141 (1) | 1 | 17 / 0 | fell, ~70% | 115 | 0 | 27,346 | 740 |
| LOOML-real-2-P-f.json | left eye | PVLP141 (1) | 2 | 17 / 0 | fell, ~70% | 113.5 | 0 | 27,438 | 709 |
| LOOML-real-0-E2-f.json | left eye | excitatory two-step relays (17) | 0 | 15 / 0 | fell, ~65% | 115.5 | 0 | 26,973 | 699 |
| LOOML-real-1-E2-f.json | left eye | excitatory two-step relays (17) | 1 | 14 / 0 | fell, ~65% | 111.5 | 0 | 26,575 | 653 |
| LOOML-real-2-E2-f.json | left eye | excitatory two-step relays (17) | 2 | 19 / 0 | fell, ~65% | 114.5 | 0 | 27,741 | 769 |
| LOOML-real-0-I2-f.json | left eye | inhibitory two-step relays (4) | 0 | 28 / 0 | unclear, ~110% | 117.5 | 0 | 27,563 | 675 |
| LOOML-real-1-I2-f.json | left eye | inhibitory two-step relays (4) | 1 | 23 / 0 | unclear, ~110% | 115 | 1.25 | 27,446 | 725 |
| LOOML-real-2-I2-f.json | left eye | inhibitory two-step relays (4) | 2 | 32 / 0 | unclear, ~110% | 115 | 0.5 | 28,196 | 732 |
| LOOML-real-2-RND-f.json | left eye | random active neurons (17) | 2 | 26 / 0 | rose, ~105% | 113.5 | 0.5 | 27,711 | 660 |
| LOOML-real-2-R2-f.json | left eye | all two-step relays (8 Oct) (21) | 2 | 26 / 0 | rose, ~105% | 116.5 | 0 | 27,991 | 772 |
| LOOML-real-2-E2mP-f.json | left eye | excitatory relays without PVLP141 (descriptive only) (16) | 2 | 30 / 0 | descriptive only | 116 | 1 | 28,244 | 737 |
| LOOMR-real-0-f.json | right eye | nothing | 0 | 43 / 0 | normal range 41–44 | 120 | 2.75 | 27,446 | 659 |
| LOOMR-real-1-f.json | right eye | nothing | 1 | 44 / 0 | normal range 41–44 | 122.5 | 3.75 | 27,252 | 675 |
| LOOMR-real-2-f.json | right eye | nothing | 2 | 41 / 0 | normal range 41–44 | 116.5 | 5 | 27,693 | 677 |
| LOOMR-real-0-P-f.json | right eye | PVLP141 (1) | 0 | 26 / 0 | fell, ~65% | 121 | 0 | 26,631 | 635 |
| LOOMR-real-1-P-f.json | right eye | PVLP141 (1) | 1 | 31 / 0 | fell, ~65% | 117.5 | 0 | 26,572 | 655 |
| LOOMR-real-2-P-f.json | right eye | PVLP141 (1) | 2 | 24 / 0 | fell, ~65% | 121.5 | 0 | 27,008 | 652 |
| LOOMR-real-0-E2-f.json | right eye | excitatory two-step relays (20) | 0 | 26 / 0 | fell, ~60% | 120 | 0 | 26,739 | 655 |
| LOOMR-real-1-E2-f.json | right eye | excitatory two-step relays (20) | 1 | 25 / 0 | fell, ~60% | 118.5 | 0 | 26,351 | 648 |
| LOOMR-real-2-E2-f.json | right eye | excitatory two-step relays (20) | 2 | 26 / 0 | fell, ~60% | 122 | 0 | 26,947 | 640 |
| LOOMR-real-0-I2-f.json | right eye | inhibitory two-step relays (3) | 0 | 55 / 0 | rose, ~125% | 121.5 | 6 | 27,896 | 702 |
| LOOMR-real-1-I2-f.json | right eye | inhibitory two-step relays (3) | 1 | 52 / 0 | rose, ~125% | 123.5 | 5 | 27,551 | 686 |
| LOOMR-real-2-I2-f.json | right eye | inhibitory two-step relays (3) | 2 | 52 / 0 | rose, ~125% | 121 | 4.5 | 28,032 | 668 |
| LOOMR-real-2-RND-f.json | right eye | random active neurons (20) | 2 | 42 / 0 | within normal range, ~100% | 105 | 0.25 | 23,586 | 572 |
| LOOMR-real-2-R2-f.json | right eye | all two-step relays (8 Oct) (23) | 2 | 28 / 0 | fell, ~65% | 123.5 | 0 | 27,016 | 677 |
| LOOMR-real-2-E2mP-f.json | right eye | excitatory relays without PVLP141 (descriptive only) (19) | 2 | 41 / 0 | descriptive only | 119.5 | 6 | 27,799 | 684 |
Described, never read: the turn rule passed and the near-side DNa02 was 0 Hz in all 30 runs; giant fibre 113.5–122 Hz except the right-eye random run (105 Hz); MDN 0 Hz in every PVLP141 and excitatory-relay run. Arm readings: PVLP141 fell, excitatory relays fell, inhibitory relays and random neurons "sides differ".
Share the threshold result with its caveats: the share card (SVG). Method, every trial and the deviations: report of 8 Oct 2026 (threshold) and report of 9 Oct 2026 (relays by firing rate) · the ledger row: Does fly wiring help? · teaching with it: Use in class
Who does the walking?
Added 2 Oct 2026 · our own test, pre-registered · context for embodied claims; it changes no grade
"The fly brain learned to walk" is the most common embodied claim: Eon's embodied fly, the Gorilla Tag video ("learned to move around as a gorilla") and FLY67, whose browser fly turns the rates of the walking neurons P9, DNa02 and MDN into an animated walk with hand-set gains. We could not inspect Eon's or Gorilla Tag's code, so we built the setup ourselves: the Shiu et al. whole-brain model (FlyWire) drives FlyGym's walking fly, the most-used simulated fly body, through two numbers every 100 ms.
What makes the fly walk. (1) Legs, rhythm and balance come from FlyGym's walking controller. (2) The brain sends two numbers, left and right, every 100 ms. (3) How neuron rates become those two numbers is our hand-made mapping. Paths are drawn to scale from the thorax position every 10 ms; forward is to the right. The three scrambled-brain paths coincide because their drives were identical. Our test of 2 Oct 2026, pre-registered before the first trial.
Open the chart as an image (SVG) to share it; it carries its own title, numbers and source line.
Read this before sharing. The fly walks because we stimulate its forward-walking neurons (P9); P9 fires at the same rates whether the rest of the brain is real or scrambled. Our mapping caps each drive number at 1.2, and 90 of the 120 drive values of the six P9 walks (75%) sit exactly at the 1.2 clip, so forward speed could hardly differ between conditions by construction. Another hand-made mapping could let the wiring matter more. One stimulus, one model, one body, 1 s of walking, open loop (the body never feeds back to the brain).
The brain is not doing nothing. Only the real wiring recruited the turning neurons DNa02, more on the right than the left in all three seeds; the scrambled brains recruited no mapped neuron. That turned the real-brain fly by −0.4° to −26°, but random no-brain drives turned it by −7° to −27° in the same second, so 1 s of walking cannot tell the two apart. Sugar: the same brain tasting sugar fires its feeding neuron MN9 at 82 Hz while every walking command neuron stays at 0 Hz, so the fly stands.
Numbers behind the chart (all 18 runs)
| Run | Group | Forward (mm) | Heading (°) | Path (mm) | Drive L / R | Brain sends (Hz) | Files |
|---|---|---|---|---|---|---|---|
| Real brain, P9, trial seed 0 | Real brain | 13.26 | -26.3 | 18.89 | 1.20 / 1.00 | P9 130/162, DNa02 7/29, MDN 0/0 | body/P9-real-0.json drives/P9-real-0.json brain/P9-real-0.json |
| Real brain, P9, trial seed 1 | Real brain | 14.35 | -0.4 | 19.64 | 1.16 / 1.11 | P9 146/132, DNa02 11/21, MDN 0/0 | body/P9-real-1.json drives/P9-real-1.json brain/P9-real-1.json |
| Real brain, P9, trial seed 2 | Real brain | 14.17 | -14.4 | 19.50 | 1.19 / 1.06 | P9 141/136, DNa02 10/25, MDN 0/0 | body/P9-real-2.json drives/P9-real-2.json brain/P9-real-2.json |
| Scrambled brain, P9, shuffle 3 | Scrambled brain | 14.15 | -14.6 | 19.52 | 1.20 / 1.06 | P9 130/162, DNa02 0/0, MDN 0/0 | body/P9-D3.json drives/P9-D3.json brain/P9-D-3.json |
| Scrambled brain, P9, shuffle 4 | Scrambled brain | 14.15 | -14.6 | 19.52 | 1.20 / 1.06 | P9 130/162, DNa02 0/0, MDN 0/0 | body/P9-D4.json drives/P9-D4.json brain/P9-D-4.json |
| Scrambled brain, P9, shuffle 5 | Scrambled brain | 14.15 | -14.6 | 19.52 | 1.20 / 1.06 | P9 130/162, DNa02 0/0, MDN 0/0 | body/P9-D5.json drives/P9-D5.json brain/P9-D-5.json |
| No brain, constant drive (L 1.18, R 1.06) | No brain | 14.28 | -11.3 | 19.49 | 1.18 / 1.06 | no brain | body/matched.json drives/matched.json |
| No brain, random drive, seed 0 | No brain | 14.01 | -9.7 | 19.07 | 1.16 / 1.05 | no brain | body/random-0.json drives/random-0.json |
| No brain, random drive, seed 1 | No brain | 14.39 | -7.5 | 19.65 | 1.18 / 1.08 | no brain | body/random-1.json drives/random-1.json |
| No brain, random drive, seed 2 | No brain | 13.48 | -27.2 | 18.81 | 1.19 / 0.97 | no brain | body/random-2.json drives/random-2.json |
| Real brain tasting sugar | Sugar | 0.00 | -0.0 | 0.12 | 0.00 / 0.00 | P9 0/0, DNa02 0/0, MDN 0/0 | body/SUG-real-0.json drives/SUG-real-0.json brain/SUG-real-0.json |
| Zero drive | Floor | 0.00 | -0.0 | 0.12 | 0.00 / 0.00 | no brain | body/zero.json drives/zero.json |
| FlyGym default drive (1, 1), sanity check | No brain (sanity) | 13.01 | 6.7 | 17.48 | 1.00 / 1.00 | no brain | body/const-1.json drives/const-1.json |
| Scrambled brain tasting sugar, shuffle 3 | Sugar (fill) | 0.00 | -0.0 | 0.12 | 0.00 / 0.00 | P9 0/0, DNa02 0/0, MDN 0/0 | body/SUG-D3.json drives/SUG-D3.json brain/SUG-D-3.json |
| Scrambled brain tasting sugar, shuffle 4 | Sugar (fill) | 0.00 | -0.0 | 0.12 | 0.00 / 0.00 | P9 0/0, DNa02 0/0, MDN 0/0 | body/SUG-D4.json drives/SUG-D4.json brain/SUG-D-4.json |
| Scrambled brain tasting sugar, shuffle 5 | Sugar (fill) | 0.00 | -0.0 | 0.12 | 0.00 / 0.00 | P9 0/0, DNa02 0/0, MDN 0/0 | body/SUG-D5.json drives/SUG-D5.json brain/SUG-D-5.json |
| Real brain, left P9 only | Real brain (fill) | -4.77 | 254.2 | 12.75 | -0.07 / 1.11 | P9 131/0, DNa02 14/0, MDN 0/0 | body/P9L-real-0.json drives/P9L-real-0.json brain/P9L-real-0.json |
| Scrambled brain, left P9 only, shuffle 3 | Scrambled brain (fill) | -2.18 | 210.3 | 10.16 | 0.00 / 1.11 | P9 131/0, DNa02 0/0, MDN 0/0 | body/P9L-D3.json drives/P9L-D3.json brain/P9L-D-3.json |
Retention, (null − zero) / (real − zero) on the mean forward distance: scrambled 1.016, brainless constant 1.025, brainless random 1.002. Mapping (ours, hand-made, not fitted): drive_s = clip(pbar − 0.5·(p_s − p_o) − m_s − 0.5·(a_s − a_o), −1.2, 1.2) with p, m, a = DNp09, MDN and DNa02 rates ÷ 100 Hz.
What it shows
- The legs walk by themselves. FlyGym's controller makes the steps, rhythm and balance; with no brain at all, its default drive walks 13.0 mm in 1 s.
- The brain does not need to be real for this walk. Stimulating P9 makes the fly walk about 14 mm whether the rest of the brain is real, scrambled or absent.
- The real brain does add turning commands (DNa02) that scrambled brains lack, the same neurons the steering study above uses. In 1 s their effect is within what random drives produce.
What it does not show
- It is not a test of Eon's, Gorilla Tag's or any other demo's code. It shows where the walking comes from in the most common setup, so a claim of a brain that "learned to walk" needs its own control.
- Our mapping caps the drive, so speed could hardly differ by construction. One stimulus, one model, one body, 1 s, open loop.
Questions to ask of a "brain walks" video: who chose which neurons to stimulate? Does the body have its own walking controller? How do neuron rates become leg commands, and who wrote that rule? Does the same body walk with a scrambled brain, or with no brain? Full test: Does fly wiring help? · build it: Add a body.
Five questions to ask about any fly brain demo
-
Which wiring does it load?
Look for a named dataset and version, such as a FlyWire connectome release. "Inspired by the fly brain" or "based on the connectome" without a file or version is a warning sign.
- Measured, if a real release is loaded
-
What turns the wiring into activity?
There must be a neuron model that runs over time. If the code only draws the network, or picks neurons at random for display, the brain is decoration.
- Modelled
-
Who chose the inputs and outputs?
In most games, the author decides which neurons "see" the screen and which neurons press which keys. That choice can matter more than the brain itself.
- Hand-made
-
Was anything trained, or scripted?
A trained readout or controller can learn to play well with almost any network behind it. Animation or game code can make a fly move without any neuron firing.
- Trained
- Scripted
-
Was it compared with anything?
The strongest projects compare their result with real fly experiments, or run the same test with shuffled or random wiring. If the shuffled brain does just as well, the real wiring did not help.
Our evidence grades
- ATested against data or a control
A named connectome release runs in a neuron model, and the output is compared with real fly experiments or with a proper control such as shuffled wiring. Methods and code or data are public. A careful negative result also gets grade A.
- BConnectome runs, hand-made inputs and outputs
The code loads a named connectome release and simulates activity that drives the behaviour. The author chose which neurons get the input and which neurons control the output. Trained readouts are allowed but flagged. No control or benchmark.
- CPartial, stand-in or decorative network
A small hand-picked circuit with set or trained weights, a network that is not taken from the connectome, or a real connectome whose activity only triggers scripted movements. The network is not the main source of the behaviour.
- DScripted or ordinary game logic
The behaviour comes from normal code, animation, a controller without connectome limits, or a language model. Any neurons on screen are cosmetic.
- UUnverified claim
We could not inspect how it works: there is no code, the code does not match the claim, or there is only a video or post. We record the claim and its source.
- n/aInfrastructure
Datasets, bodies, simulators, analysis tools and directories. We do not grade them for behaviour; they get licence, maintenance and setup notes instead.
We give the lowest grade whose conditions all hold, and we grade from code, configuration, data loading or paper methods that we inspected, never from a README headline alone.
Verdicts on viral fly brain claims
Each row splits a viral claim into its parts: what the simulated brain really does, and what a person chose, trained or scripted. We inspected the code on 28 September 2026 and re-checked the rows on 29 and 30 September and 1, 2, 3, 5, 6 and 8 October; the last column shows each row's latest check. On 8 October a claim that a fly brain model "learns to hack" was added at the top (grade U: no code linked). On 2 October the Terraria fly was added at the top, the first viral video claim this week with code; on 1 October two video claims were added, Gorilla Tag and a Roblox fight, both without code; on 30 September the FlyLeno claim was added and the Doom row gained the autopilot sweep. Results are the authors' own numbers; we did not re-run these projects. "Watch" links open the matching cards in the Video gallery.
Two claims checked on 29 Oct 2026
- A walking circuit found by connectome simulation (Nature, 28 Oct 2026). Pugliese and colleagues (Nature, 28 Oct 2026) report that simulations of four fly nerve-cord wiring maps found a three-neuron circuit, one inhibitory and two excitatory cells, that produced leg rhythms in the model, and that a descending pathway the simulations pointed to drove rhythmic leg movements when they activated it with light in behaving flies. The simulation code (MIT licence) and simulation data are public. We did not re-run it, and our tools could not open the journal page; we read the preprint abstract and the code repository (checked 29 Oct 2026). Sources: Nature article (not readable by our tools); preprint; code; simulation data (not opened). Update, 30 Oct 2026: now in our catalogue, graded A. The authors' code includes a test that shuffles the wiring while keeping the balance of excitatory and inhibitory connections; their saved notebook output shows the number of three-neuron circuits of the walking-rhythm type among front-leg nerve-cord neurons: 21,544 with the real wiring and 306.95 ± 29.10 (mean ± SD) across 100 shuffles. We did not re-run it (checked 30 Oct 2026). This is a count of circuit patterns in the wiring, not a walking test with shuffled wiring, so our ledger does not score it. Verdict row. Update, 31 Oct 2026: We checked whether the authors' walking model can be re-run with shuffled wiring on an ordinary 2-processor server without a GPU (31 Oct 2026): its software would need an estimated 620 MB once installed, plus 86 MB of wiring data, more than our 600 MB test space. The test was not run. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. Whether the real wiring is needed for the simulated walking rhythm stays not tested; the earlier circuit count is a pattern in the wiring, not a test of walking. Update, 9 Nov 2026: We re-ran the authors' nerve-cord walking model on an ordinary 2-processor server without a GPU (9 Nov 2026): with the real front-leg wiring, 3 motor neurons showed a regular rhythm; with the wiring shuffled by the authors' own shuffle, 4 of 5 tries still showed a rhythm (0, 12, 40, 3 and 1 motor neurons in the 5 tries). One run each, faster solver settings than the paper, no noise. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. The shuffled circuits were far more active than the real one (1,645 to 1,995 active cells against 97; top firing rates about 230 against 16 per second), so this is not a like-for-like comparison; a test with activity matched has not been run. By the rule we fixed before the run (if at least half of the shuffles still show a rhythm, our row reads no difference), our own row reads no difference. The catalogue entry for the paper stays not tested. (report of 9 Nov 2026; our ledger row).
- Videos saying a fly brain was "uploaded". Two new videos use the word upload for a fly brain: "The First Brain Was Just Uploaded To A Computer" (AI Uncovered, 27 Oct 2026; about 26,000 views on 29 Oct) and a Japanese video whose title says a fly's brain was copied whole and run (25 Oct 2026; about 50,000 views on 29 Oct), whose description names Eon Systems' fly demo. Primary sources show a wiring map traced from one preserved fly brain and run as a simplified model, with parameters from the map, predicted neurotransmitters and earlier studies; Eon's body code is not public, so we grade its demo U. We did not watch the videos; see our answer "Did a dead fly come back to life in a computer?" (checked 29 Oct 2026). Update, 9 Nov 2026: A new Urdu and Hindi video, 'Scientists Uploaded a Fly's Brain Into a Computer Are Humans Next?' (Takhti, 3 Nov 2026; about 202,000 views on 9 Nov), uses the word uploaded in its title, while its own description says this does not mean a fly's mind was copied and calls Eon Systems' demos company claims with limits. That matches the primary sources: a wiring map traced from one preserved fly brain, run as a simplified model. We read the title and description, not the video (checked 9 Nov 2026). Verdict row.
| Claim and source | What the simulated brain does | Hand-made, trained or scripted | Measured result (by the authors) | Grade | Last checked |
|---|---|---|---|---|---|
| New 10 Nov 2026 "What Is It Like to Be a Fly?" Free browser gamebook by Andrew Sheerin, shared from 6 Nov 2026. Its note says the taste, motor, grooming and looming-threat neurons are simulated and the activity shown is genuine for those. gamebook |
Replays recorded spike output: spike times embedded in the page, from runs of the Shiu et al. 2024 model on FlyWire data and from DesktopFly's model. No neuron model runs in the page. | Written by the author, as its own note says: the movements, the stimulus schedules and two of its six scenes (the letter and the final choice). | None; it is a gamebook, not a test. Credits: Shiu et al. (2024), DesktopFly, FlyWire (CC BY-NC 4.0), MaleCNS (CC BY 4.0). A line says "code MIT", but no code repository is linked. Read from the page's files on 10 Nov 2026; the page was not rendered. | Not graded Recorded replay, not a live simulation; no code repository to grade. |
No code to re-check Checked ; not graded: no code repository linked |
| New 8 Oct 2026 "Can A Conscious Fly Brain Learn how to Hack?" YouTube video by dzuma, uploaded 7 Oct 2026. The channel rotates the title ("I Taught A Fruit Fly Brain how to Hack", "POV: You Teach A Conscious Fly Brain Hacking") video · entry dzuma-fly-hack Watch: YouTube, length not recorded |
Claimed, from the title and description only: a whole fly brain model built from FlyWire and simulated in Brian 2. The description cites papers on odour learning and dopamine neurons, and links the FlyWire Codex and Brian 2. | Claimed: the model was taught with dopamine-style reward and learned to attack two deliberately vulnerable practice targets (Damn Vulnerable Web Application and the VulnHub Kioptrix level 1 machine, standard training exercises). How neurons become web requests or commands is not stated; a fly brain model has no notion of web requests, so any such mapping would be made by the author. | None. No code linked: the description links papers, the FlyWire Codex, Brian 2 and the two practice targets, and no repository or data for the project; the GitHub searches of 8 Oct 2026 found no matching repository. We have not watched the video. What would change it: a public repository with the brain-to-action mapping and logs, so we can see what is trained, what is scripted and whether any step comes from the model rather than a script. |
U Unverified (8 Oct 2026). Nothing public to check; a video alone never sets a grade above U. |
No code to re-check Checked ; graded 8 Oct 2026; grade U: nothing to inspect approximately 92,236 views on 29 Oct 2026 (video page; gallery card) |
| Added 2 Oct 2026 "I put a fly's brain into Terraria. And tried to train it" YouTube video by George Ostrobrod (in Russian), uploaded 29 Sep 2026 video · code (GitLab) · entry fly-terraria Watch: YouTube, length not recorded |
The MaleCNS v1.0 connectome (166,700 neurons, 25,088,107 signed connections; the input files' checksums are enforced) runs as a sparse spiking model with adaptation. Synapses change with reward-modulated plasticity, rewarded by the body's own energy, water and harm signals. | Game light, smell, touch, pain and balance are injected into hand-picked sensory neuron groups (the author calls the receptor assignment "approximate"); motor neuron groups are decoded into turning, walking, feeding, grip and adhesion with fixed hand-set functions. Grip and adhesion have documented scripted fall-backs when the graph lacks those neurons. The 2-D body is a Terraria mod. | One 13.7-hour run, 176 lives; we recomputed the log: mean lifespan 278.4 s, and the median rose from 102.8 s (first 44 lives) to 163.3 s (last 44; our permutation p = 0.037). No control: the author's own report says there was no plasticity-off control and that the world changed during the run. What would change it: paired plastic and plasticity-disabled runs from the same start (the author's own suggestion), or the same body with scrambled wiring; code that matches the video. |
B Graded 2 Oct 2026 from the files at commit 9052966d (26 Sep), which is older than the video. Real and candid, but the "learning" is not separated from luck. |
Code unchanged since graded Checked ; graded 24 Oct 2026; GitLab HEAD still the graded commit approximately 51,400 views on 29 Oct 2026 (video page; gallery card) |
| New 1 Oct 2026 "I Taught a Fly How to Play Gorilla Tag (seriously)" YouTube video by SuperCatCrazeGT, uploaded 26 Sep 2026 video · entry fly-gorilla-tag Watch: YouTube, length not recorded |
Not stated. The description says: "Using Its OPEN SOURCE BRAIN, The Fly Learned To Move Around As A Gorilla … It Even Learned How To Play INFECTION!" No connectome release, neuron model or project is named. "Open source brain" fits several catalogued projects, such as the Shiu et al. model and its browser ports, but none is named. | Claimed: the fly "learned" over seven chapters; the method is not stated. How game input reaches neurons and how neurons move the avatar is not stated. The same channel's earlier video, "I Trained an AI to Beat Gorilla Tag's Fastest Player", trained an ordinary AI on the game. | None. No code found: the only description link is a linktr.ee page with 76 links (game maps, socials, sponsors), none naming a fly, brain or connectome; four GitHub searches returned nothing. We have not watched the video, so its on-screen claims are not recorded. What would change it: the controller code with the brain model named, so it can be graded from files; a named base project alone would not raise the grade. |
U Unverified (1 Oct 2026; re-checked 2 Oct: still no code). The most-viewed fly-brain claim in our scans, with no project or code behind it; "the fly learned" may describe a trained controller rather than fly wiring. For how much of a simulated walk a brain itself provides, see our body test (context only). |
No code to re-check Checked ; graded 1 Oct 2026; grade U: nothing to inspect approximately 1,204,127 views on 29 Oct 2026 (video page; gallery card) |
| New 30 Oct 2026 "Connectome simulations identify a central pattern generator for fly walking" Pugliese, Tuthill, Brunton and colleagues, Nature, 28 Oct 2026; the claim is the article title article · preprint · code · entry pugliese-2026-walking-cpg |
Rate-model simulations of the fly's nerve cord, built from four wiring maps (MANC, FANC, BANC and the male CNS), pruned to a small circuit that still produced leg rhythms in the model: two excitatory neurons in a chain and one inhibitory neuron inhibiting both. A nerve-cord study, not a whole-brain or body simulation. | Nothing trained: the network is pruned, not trained; the stimulus and the pruning procedure are chosen by the authors. | The authors' code includes a test that shuffles the wiring while keeping the balance of excitatory and inhibitory connections; their saved notebook output shows the number of three-neuron circuits of the walking-rhythm type among front-leg nerve-cord neurons: 21,544 with the real wiring and 306.95 ± 29.10 (mean ± SD) across 100 shuffles. We did not re-run it (checked 30 Oct 2026). This is a count of circuit patterns in the wiring, not a walking test with shuffled wiring, so our ledger does not score it. 31 Oct 2026: We checked whether the authors' walking model can be re-run with shuffled wiring on an ordinary 2-processor server without a GPU (31 Oct 2026): its software would need an estimated 620 MB once installed, plus 86 MB of wiring data, more than our 600 MB test space. The test was not run. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. Whether the real wiring is needed for the simulated walking rhythm stays not tested; the earlier circuit count is a pattern in the wiring, not a test of walking. 9 Nov 2026: We re-ran the authors' nerve-cord walking model on an ordinary 2-processor server without a GPU (9 Nov 2026): with the real front-leg wiring, 3 motor neurons showed a regular rhythm; with the wiring shuffled by the authors' own shuffle, 4 of 5 tries still showed a rhythm (0, 12, 40, 3 and 1 motor neurons in the 5 tries). One run each, faster solver settings than the paper, no noise. The authors' shuffle keeps each neuron's excitatory or inhibitory sign and shuffles connections within each cell group; it differs from the shuffle used for their circuit-pattern count. The shuffled circuits were far more active than the real one (1,645 to 1,995 active cells against 97; top firing rates about 230 against 16 per second), so this is not a like-for-like comparison; a test with activity matched has not been run. By the rule we fixed before the run (if at least half of the shuffles still show a rhythm, our row reads no difference), our own row reads no difference. The catalogue entry for the paper stays not tested. (report of 9 Nov 2026; our ledger row). | A Peer-reviewed; code public (MIT). We could not open the journal page; we read the code, the preprint abstract and the saved notebook output. |
Peer-reviewed paper, code public Checked ; graded 30 Oct 2026; code and saved notebook output read, not re-run by us |
| New 29 Oct 2026 "The First Brain Was Just Uploaded To A Computer" AI Uncovered on YouTube, 27 Oct 2026, and a Japanese video of 25 Oct 2026 whose title says a fly's brain was copied whole and run; the claims are the videos' own video 1 · video 2 · video 3 (9 Nov) · entry eon-embodied-fly |
The wiring map was traced from one preserved fly brain and is run as a simplified model, with parameters from the map, predicted neurotransmitters and earlier studies. See Did a dead fly come back to life in a computer? | Eon Systems' fly demo, which the Japanese video's description names: its body code is not public, so how the body is driven cannot be inspected. | None to measure: these are videos about a demo. Two new videos use the word upload for a fly brain: "The First Brain Was Just Uploaded To A Computer" (AI Uncovered, 27 Oct 2026; about 26,000 views on 29 Oct) and a Japanese video whose title says a fly's brain was copied whole and run (25 Oct 2026; about 50,000 views on 29 Oct), whose description names Eon Systems' fly demo. Primary sources show a wiring map traced from one preserved fly brain and run as a simplified model, with parameters from the map, predicted neurotransmitters and earlier studies; Eon's body code is not public, so we grade its demo U. We did not watch the videos; see our answer "Did a dead fly come back to life in a computer?" (checked 29 Oct 2026). 9 Nov 2026: A new Urdu and Hindi video, 'Scientists Uploaded a Fly's Brain Into a Computer Are Humans Next?' (Takhti, 3 Nov 2026; about 202,000 views on 9 Nov), uses the word uploaded in its title, while its own description says this does not mean a fly's mind was copied and calls Eon Systems' demos company claims with limits. That matches the primary sources: a wiring map traced from one preserved fly brain, run as a simplified model. We read the title and description, not the video (checked 9 Nov 2026). | U Eon's demo: unverified, body code not public. |
No code to re-check Checked ; graded 28 Sep 2026; grade U: nothing to inspect |
| New 18 Oct 2026 People ask: Did a fly brain fly through an AI world on its own?Our dated status, 30 Oct 2026: not checkable yet (no method, code or connectome version published). "We placed the fruit fly brain inside Odyssey-3, a frontier world model. The lil' guy is now flying around totally imagined, simulated worlds all by itself" Post by Odyssey (@odysseyml) on X, 17 Oct 2026, with a short video; the claim is Odyssey's own post · entry odyssey3-fly-world-model-claim |
Not stated. The post does not say which fly brain model was used (for example a connectome simulation or a learned policy). | Claimed: the fly flies through worlds generated by Odyssey-3, its generative world model. What drives the flight and how the world model feeds the brain input are not stated. | None. No method, code or write-up was published as of 30 Oct 2026 (checked 30 Oct); no public fly mode found on the preview page (29 Oct). Until then the claim cannot be checked. Odyssey's Odyssey-3 announcement page does not mention a fruit fly, a fly brain or a connectome, and we found no paper, repository or demo page about the fly. We read the post's text from the page without logging in; we did not see the video. What would settle it: a write-up or code that names the fly brain model, how its outputs drive the flight and how the world model feeds it sensory input. |
U Unverified (re-checked 30 Oct 2026): nothing can be inspected. |
No code to re-check Checked ; graded 18 Oct 2026; grade U: nothing to inspect |
| New 15 Oct 2026 "Roblox is secretly testing fruit fly brains on their avatars" From the description of a YouTube recap video by SharkBlox, uploaded 13 Oct 2026; the claim's source, not a project video recap video · entry roblox-fly-avatar-claim |
Not stated. No builder, connectome release, neuron model or project is named. | Claimed: fruit-fly brains control Roblox avatars, including a user named "Admin (Fly)", in BloxStrike and Natural Disaster Survival. Any coupling would be in code that is not published. | None. The description credits an earlier NiftRex video and links the two games and the Roblox user, but no builder, code, paper or project page. The builder is unknown; we have not watched the video and found no code. It is a separate claim from fruitfly-roblox and fly-jjs-fight. What would change it: the builder, the controller code and the brain model it calls. |
U Unverified (15 Oct 2026): nothing can be inspected. |
No code to re-check Checked ; graded 15 Oct 2026; grade U: nothing to inspect |
| New 1 Oct 2026 "I Made Two Real Flies Brain Fight Each other in JJS" YouTube video by Evoke, uploaded 21 Sep 2026; Jujutsu Shenanigans is a Roblox fighting game video · entry fly-jjs-fight Watch: YouTube, length not recorded |
Not stated ("real flies brain"). No connectome release, neuron model or project is named. | Claimed: two fly brains control two Roblox fighters. Any coupling would be the creator's own client script, which is not published. | None. The description links only a Discord invite and a Roblox group; three GitHub searches found nothing related (the Roblox script repositories they return are unrelated and not linked). We have not watched the video. What would change it: the controller script and the brain model it calls; the same fight with scrambled wiring or a random-input bot. |
U Unverified (1 Oct 2026): nothing can be inspected. |
No code to re-check Checked ; graded 1 Oct 2026; grade U: nothing to inspect approximately 151,586 views on 21 Oct 2026 (video page; gallery card) |
| New 30 Sep 2026 "Tonight's host: Grey Leno, piloted by Drosophila melanogaster" FlyLeno (TUURD Talk) by AgitationSkeleton, a browser show; Vinesauce streamed it (view count in the Last checked column) show · repository · entry flyleno Watch: YouTube, streamer's playthrough, 1:55:44 |
The whole FlyWire v783 brain (138,639 neurons, every connection kept) runs live and untrained in the browser as the Shiu et al. spiking model. Its descending neurons drive walking, turning, startle, grooming and feeding, and a dopamine signal from its own neurons changes the synapses onto the voice neurons while the show runs. | Show events, thrown objects, music and food stimulate hand-picked sensory neurons; descending-neuron rates are turned into movements with fixed hand-set gains. Engineered layers, labelled as such by the author: a walking rhythm generator, balance "puppet strings", saccades, food seeking, homing, sleep, and an action selector outside the connectome that picks what Leno does next (talk, stroll, turn, rest…) and stimulates the matching neurons. The show itself is ordinary code. | None: no scrambled-wiring or no-brain control. Our browser check (3 Oct 2026): the page fetched the 31.4 MB connectome into a Web Worker and its counters started, then it froze in our headless browser without a GPU (most likely software graphics). This fits the row: the brain runs in the page. Not a play test; details. What would change it: up to A with a scrambled-wiring or no-brain run of the same show and a measured behaviour (for example startle latency, eating decisions or time spent in each action); down to C with evidence that the engineered layers drive most of what is on screen. |
B Borderline C (graded 30 Sep 2026 from the code at commit 078d8fa). The brain is real and really running, but the show is mostly built around it. The viral video is a streamer's playthrough; the claim graded here is the site's and the repository's own. |
Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date approximately 55,574 views on 29 Oct 2026 (video page; gallery card) |
| New 29 Sep 2026 Re-graded 15 Oct 2026 "A digital fly's brain on a Raspberry Pi 5 plays Pokémon" Twitch stream by Leetzerzz; XDA Developers, 28 Sep 2026; code published 14 Oct 2026 by ianwes92 (its README names the stream) XDA article · stream · code · entry fly-pokemon-pi5-stream |
From the code: the whole FlyWire FAFB v783 graph (connections of five or more synapses) runs as a leaky integrate-and-fire network on the Pi, next to the mGBA emulator. | The fly does not see the screen: walls, doors and people are read from the game's memory (RAM) and mapped onto visual neurons (LPLC1, LC10, LPLC4). Descending-neuron rates pass through a hand-written priority arbiter that picks the key. Chat keys are added to the fly's keys, and a paid override can replace the fly's keys for a while. Only Kenyon-cell to MBON synapses learn (chat !pain / !pleasure and code-side rewards). | None for the fly's share of the play: there is no shuffled-wiring control, and the stream title's game progress does not say how much comes from the fly, the arbiter or chat. We read the code; we did not watch the stream. Two other open-source "fly plays Pokémon" projects, both grade C and separate from this stream: acamilo/flybrain (the author's own check found the decoder, not the network, picking directions in long runs, and its streamed runs move by scripted macros) and fly-plays-games (the long walk in its video is a scripted planner). What would change it: the same play with scrambled wiring, or with the arbiter fed random rates. |
B Borderline C. Re-graded 15 Oct 2026 from the public code (our judgement call; a stricter reading of the arbiter, chat keys and paid override could give C). Until 15 Oct: U, no code to inspect. |
Code unchanged since graded Checked ; graded 30 Oct 2026; last commit on or before the grade date |
| New 29 Sep 2026 "A fly's digital brain plays Rainbow Six Siege" dev.ua, 28 Sep 2026, reporting an X post of 26 Sep by the caster Oleksandr "GHOOD_BHOY" Samoilenko dev.ua report · entry fly-brain-rainbow-six-siege |
Claimed: "over 1,000,000 simulated neurons". That matches no fly connectome: MaleCNS v1.0 has 166,700 neurons and FlyWire v783 has 139,255, so either most of the network is not fly wiring or the number is wrong. | Claimed: about a month of training on "more than 20 TB" of professional players' gameplay. That is ordinary machine learning; any skill would come from the training unless a control shows otherwise. | One in-game kill, described on social media, with no baseline. We read the dev.ua text only through search summaries, because the site blocks our fetches, and X cannot be read without a login. | U Nothing published: no code, model or write-up (searched 29 Sep 2026; re-checked 30 Sep 2026: no clip, code or follow-up). Unlikely to be a fly brain at work. |
No code to re-check Checked ; graded 29 Sep 2026; grade U: nothing to inspect |
| "First embodied fly upload" Eon Systems, March 2026 Eon page · entry eon-embodied-fly Watch: X, 0:43, project site, 0:40, YouTube repost, 1:16, TikTok repost, 1:30 |
By Eon’s own description: the Shiu whole-brain LIF model (FlyWire) plus flyvis vision in a NeuroMechFly body. | A few hand-picked neurons switch pre-trained, imitation-learned walking controllers; sensory rates and gains were "chosen by hand"; Eon calls the vision "somewhat decorative". | None for the embodied fly. The "91%" Eon cites is the brain-only 2024 result. | U Body code not released (checked 28 Sep 2026). If the description is accurate, the mechanism would be grade C. Context, not evidence about Eon's code: in our own body test the walk came from the body controller and the stimulated neurons. |
No code to re-check Checked ; graded 28 Sep 2026; grade U: nothing to inspect |
| "Fly brain plays Doom" DOOMFLY repository · project page · entry doomfly Watch: X, 0:07, X, 0:10 |
The full MaleCNS v1.0 model (166,700 neurons) receives game frames, and its spikes drive the game. | Screen-to-photoreceptor input is an engineered adapter. The four read-out neurons were chosen because they responded to the screen; the biologically expected steering neurons gave zero actions. | Negative: a dopamine "learning" version survived 3.66 s against 5.83 s without learning (two episodes each). The author says learning failed. Removing all connections stops all control. Third-party control (new, 28 Sep 2026): real wiring 53.3 s survival, shuffled wiring 6.75 s, no connections 5.5 s, and a brainless constant autopilot 51.9 s (51.93 s, with its commands matched to the fly's average). The same results file (results/a2/summary.json) also holds a sweep of 12 autopilot settings: the best one survives 53.44 s, slightly longer than DOOMFLY's 53.34 s (checked by us on 30 Sep 2026). The autopilot's constants come from the real fly's average output, so this shows that a constant policy reproduces the behaviour, not that the brain does nothing. Chart · entry doom-fly-control |
B Honest, but no sign of skilled play. The outside controls do not raise the grade. Other Doom projects with controls: doomfly-rl (A: real, shuffled and no wiring score the same) and flydoom by mutkuoz (A: beats a matched random agent only narrowly). |
Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date |
| "I put a real fruit fly brain into Half-Life" FlyBrain-HalfLife, Yusuf Temel; the video says it uses "the biological connectome … MaleCNS v1.0" with 12,260 neurons. A second video (23 Sep 2026) adds: "There is no scripted game AI or traditional reinforcement learning at play here." repository · entry flybrain-halflife Watch: YouTube, 6:26 |
No connectome by default. The released code builds a synthetic 12,260-neuron network with brain-region labels from region sizes. It uses real data only if the user supplies a connectome file (an npz file) themselves. | Steering, attacks and escapes are computed outside the network from image rules and scripted reflexes, then pushed into the output neurons, so the neurons mostly pass the decision on. | None beyond speed benchmarks; no controls. | D Both video claims are contradicted by the code. Re-checked 29 Sep 2026: still commit 0ba5548, unchanged since our 28 Sep check, with no new connectome loader. |
Code unchanged since graded Checked ; graded 28 Sep 2026; last commit on or before the grade date approximately 163,535 views on 29 Oct 2026 (video page; gallery card) |
| "A fly’s brain in Minecraft" NeuroCraft Fly post · entry neurocraft-fly Watch: X, 0:09 |
Claimed: all 166,700 MaleCNS v1.0 neurons. | The README says read-outs "select and modulate scripted body programs"; the body moves even with inputs off; some responses survive weight shuffling. | Only numbers reported by the author: with shuffled weights (one seed), 5 of 6 target responses remained, a one-take illustration the author says is not a benchmark. | C Re-graded 10 Nov 2026 (was U): the source is now public (release candidate 0.2.0-rc.7, MIT); a leaky-rate model on a retained MaleCNS v1.0 graph, whose readouts select scripted body programs. Not run by us. |
Code unchanged since graded Checked ; graded 10 Nov 2026; last commit on or before the grade date |
| Fly Brain Minecraft mod blendi-remade repository · entry fly-brain-minecraft Watch: X, 0:13 |
MaleCNS v1.0 LIF model, using connections with 5 or more synapses. | A reflex layer (odour following, random bouts, collision turns) is on by default and acts when the brain is silent. Looming and motion signals are computed from the game and injected into visual neurons. | Sanity checks against the Shiu model (sugar input drives MN9 at 30–90 Hz); no wiring control. | B | Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date |
| "A fly's brain learned to farm in Minecraft" FlyBridge, SWOT; a third Minecraft project repository · entry flybridge Watch: YouTube, 34:00 |
A 1,536-neuron smell-pathway part of the FlyWire v783 wiring (about 1% of the brain) works as one layer inside a larger policy network. | Every connection strength in that layer is trained, together with trained image encoders and a trained readout. A scripted controller does the walking, aiming and clicking. | The author reports 12 of 12 checks passed in fresh worlds. No stored result compares it with the shuffled or disconnected versions in the code. | C New 28 Sep 2026. The file citations for this grade were spot-read by us, not re-checked line by line. |
Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date approximately 258,800 views on 29 Oct 2026 (video page; gallery card) |
| "No behaviour is scripted" FlyCraft for Minecraft Education repository · entry flycraft |
The FlyWire v783 brain (138,639 neurons); feeding and escape really pass through the network. | Walking, seeking, homing and sleep are decided in Python and injected into the same descending neurons that are read out. | Removes a scripted drive, not the connectome. | C The headline is contradicted by the code. |
Code unchanged since graded Checked ; graded 22 Oct 2026; last commit on or before the grade date |
| "Fly brain plays Beat Saber" viral post, approximately 23,059,544 views on 28 Sep 2026 post · entry beat-saber-fly Watch: X, 0:31 |
Unknown. | In replies (seen only in search summaries), the author says the motor output was fitted to a replay of one song and that visual training was still running. | None. | U No code, weights or method. |
No code to re-check Checked ; graded 28 Sep 2026; grade U: nothing to inspect |
| "Bad Apple on a fly brain" viral post, about 20.8 million views post · entry bad-apple-fly Watch: X, 0:38 |
A MaleCNS v1.0 LIF model runs. | The picture is mostly the stimulus: neurons under dark video pixels are forced to fire. The README says it is "not a real brain simulation". The body uses pre-programmed FlyGym steps and hand-written gestures. | None. | C | Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date |
| "Playing Mario 64 using a fly’s brain" Fly64 repository · entry fly64 Watch: X, 1:20 |
MaleCNS v1.0 LIF model (166,700 neurons); the stick and buttons come from chosen neurons. | The eye-to-pixel map is partly estimated; stick smoothing, dead zone and jump cooldown are ordinary code; the author’s own coarse model settings. | None: no goal, benchmark or control. | B | Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date |
| "The fly trades bitcoin" Stonkfly repository · entry stonkfly Watch: X, 0:10 |
MaleCNS v1.0 LIF model; profit and loss are sent as dopamine pulses. | The price-to-neuron input and the buy/sell decoder are hand-made; risk limits can block orders; paper trading by default. | The README says "profitable learning has not been demonstrated". | B | Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date |
| "Driven by a live spiking simulation of the real FlyWire connectome" DesktopFly (macOS, with a Windows port); the sentence is the GitHub repository description ("About"), not the README repository · project page · entry desktop-fly |
A 668-neuron FlyWire v783 circuit, plus a 1,045-neuron MaleCNS leg circuit since v1.1.0. | The 23,210 "neurons" on screen are display points. An ordinary state machine switches walk, idle, groom, fly and sleep when neuron rates cross hand-set thresholds; flight, grooming and darting are coded animation, and sleep follows the computer's idle time and the clock, not a neuron. Only walking is driven by the simulated leg neurons. | Only in-model checks. | C Re-graded from B (borderline C) on 29 Sep 2026 after a line-by-line check; the code is unchanged since 5 Sep 2026. No official app downloads exist: see the safety note on the project page. |
Code unchanged since graded Checked ; graded 23 Oct 2026; last commit on or before the grade date |
| "A fruit fly’s wiring diagram is flying this drone" FlyDrones demo · entry flydrones |
The browser demo runs MiniFly, a synthetic 850-neuron network with real cell-type names, not connectome data. | A fixed flight bias, a fixed escape programme, a safety governor and the drone’s own flight controller. | None. | C MaleCNS is only an optional path with no published run. |
Code unchanged since graded Checked ; graded 22 Oct 2026; last commit on or before the grade date |
| "A fly learns to drive and park a car" Flyhard Re-shared on 2 Oct 2026 in a compilation, "Scientists Teach A Fly How To Drive" (Daily Dose of Science), described as "a simulation of a fly's brain is driving a car". Its thumbnail appears to show Flyhard in its CARLA town; we matched it from the thumbnail only, without watching the video. On 6 Oct 2026 the search page showed this video under another title, "How Your Brain Changes What You Hear"; whether the video itself changed is not known, and the match is flagged for a re-check. compilation repository · entry flyhard Watch: X, 0:25 |
The MaleCNS v1.0 wiring (165,122 traced neurons) as a rate network. | All of its about 25.7 million connection gains were trained to copy a teacher; random fixed input and output maps; turn requests are scripted; no vision. | Steering 100/100 after training against 0/100 before; parallel parking 0/50. | B Trained parts flagged. |
Code unchanged since graded Checked ; graded 24 Oct 2026; last commit on or before the grade date |
The pattern: a real connectome is usually loaded and simulated, but a person decides which neurons see the game and which neurons press which button. In several demos, scripts or trained parts produce most of the visible movement, and one (Half-Life) does not load a connectome at all. None of the viral game demos tests its own behaviour against shuffled wiring; the first outside test of one (DOOMFLY) found that a constant autopilot does about as well. The Rainbow Six claim publishes no code at all; the Pokémon-on-a-Raspberry-Pi stream published its code on 14 Oct, and a hand-written arbiter, chat keys and a paid override shape its key presses (B, borderline C). FlyLeno is the opposite case: its code is open and its brain really runs, but engineered layers shape most of the show. Gorilla Tag and a Roblox fight are videos with hundreds of thousands of views and no project or code behind them that anyone can inspect. The newest, the Terraria fly, is the opposite: open code and a real connectome, but no control. And our own body test shows why a walking body is weak evidence on its own: the body controller walks with or without a real brain.
Claims we could not check (grade U)
- UA fly brain model "learns to hack" (YouTube, dzuma)
New 8 Oct 2026. The description links papers, the FlyWire Codex, Brian 2 and two practice targets, but no code or data; approximately 92,236 views on 29 Oct 2026. We have not watched the video. Verdict row
- UGorilla Tag (YouTube)
No model, data or code is named; the only link is a page of game maps and socials, and four GitHub searches found nothing (1 Oct 2026). Re-checked 11 Oct 2026: still no code; approximately 1,204,127 views on 29 Oct 2026. We have not watched the video.
- UTwo fly brains fight in a Roblox game (YouTube)
Only Discord and Roblox-group links; three GitHub searches found nothing related (1 Oct 2026).
- UPokémon on a Raspberry Pi 5 (Twitch)
No code is public: we found no code under the creator's name, and neither the article nor the stream page links any (searched 29 Sep 2026). The input is game memory, not vision, by the creator's own account.
- URainbow Six Siege fly brain
No code, model or write-up; known only from one news report of a social-media post, which we could read only through search summaries.
- UBeat Saber fly
No code, weights or method were published. The author's explanation (motor output fitted to a replay of one song) is known to us only through search summaries of replies.
- UNeuroCraft Fly (Minecraft)
The public repository holds documents, a licence and two videos. The simulation code and mod are "coming soon". The README itself says read-outs select scripted body programs.
- UEon's embodied fly
The body and embodiment code are not released; the public Eon repository is brain-only. Eon's own description points to pre-trained controllers switched by a few hand-picked neurons.
- ULoihi 2 fly brain (Sandia)
A preprint with no code link that we could find.
- Uflybrain.info Lab
The in-browser Lab plays back eight pre-computed runs; the model code behind them is not published.
Common questions
Plain answers to the questions people ask most, each with its evidence and what we are unsure about. Last checked 6 Oct 2026; the open-source answer and three new answers checked 29 Oct 2026.
In your words: the short answers
- Is it alive? Is it sentient or conscious? We found no study that measured consciousness or suffering in these models, which leave out much biology. Fly larvae detect harmful heat and pressure; whether flies feel pain is open.
- Is it a simulation of a mind? It is a map (a fly brain scan) run with a simple neuron model. In our tests its wiring drove a reflex and a turn; body software did the walking.
- Was a fly uploaded? Did a dead fly come back to life? No. The map was traced from one preserved fly brain, cut into thin slices and imaged; models run it as a simplified simulation, without recordings of that fly's activity.
- Is it AI? The basic model is not trained AI: it copies the fly's wiring and sets one strength by hand. Many viral demos add trained AI parts around it.
- Is the fly gamebook a live simulation? No: it replays recorded spike output from published models; no neuron model runs in the page.
- Was the virtual fly brain a lie? The 10 Physics video “They Lied to you about the Virtual Fly Brain” makes that case; our gallery card links it, and the answers above summarise what the code and papers show.
Is the fruit fly brain simulation AI?The basic model is not trained AI: it copies the fly's wiring and sets one strength by hand. Many viral demos add trained AI parts around it.
The best-tested model (Shiu et al. 2024) learns nothing: every connection comes from the fly's wiring map, and one strength value is set by hand. Many demos add trained parts. Of 172 graded projects in our catalogue (11 Oct 2026), 103 have a trained, learning or search-tuned part in the main path: 17 a trained readout, 14 train the whole network or some of its synapses, 7 a trained vision or text front end, 4 a policy or language model, 13 learning rules or tuned settings; 39 have none. One example: flybrain.app's page says a language model ("Claude") acts as the fly's caretaker; it changes the fly's world (food, light), not its brain model.
trained_class (10 Nov 2026), our reading of each project's code and README. "Learning or tuning" mixes learning rules that run during play (STDP, dopamine plasticity) and a few settings tuned by search. Optional or unused trained parts are noted per entry, not counted. Filter the list with "Trained parts (is it AI?)" on Projects.Since 6 Oct 2026 every catalogue entry carries this as a field, trained_class; on Projects you can filter by "Trained parts (is it AI?)".
Evidence: the field trained_class on the 91 graded entries (report of 6 Oct 2026); the Shiu et al. model entry (trained parts: none; one synapse weight, 0.275 mV, set by hand); the flybrain.app entry (page read 5 Oct 2026).
How sure: the classes come from our reading of each project's code and README; "learning rules or tuned settings" mixes several kinds of tuning, and projects change after we read them.
Is it a brain or just a map?It is a map (a fly brain scan) run with a simple neuron model. In our tests its wiring drove a reflex and a turn; body software did the walking.
The connectome is a map of one adult fly's brain, traced from electron-microscope images: which neurons connect, and through how many synapses. Models run this map with simple neurons and leave out gap junctions, neuromodulators and internal state. Run on our computer, the map gives a sugar reflex (the feeding neuron fires at 82 Hz) that scrambled maps lose (0 Hz), and in our looming test:
We ran the map. A looming shadow on one eye turned our simulated fly away in 6 of 6 runs. With the same map scrambled (same number of connections per neuron) it turned away in 0 of 3. Here the map's wiring, not the stimulus or our code, makes the turn.
A brain-only follow-up on 6 Oct asked whether that was only because scrambled brains are quieter: "We turned the scrambled maps up until they were as busy as the real one. They still did not produce the turn: the wiring, not the amount of activity, makes it." And where in the map: "Scrambling only the inside of the map kept about -9% of the turn command and 69% of the escape signal." (−9% means none: the far-side command was absent in all 6 trials.) And how strong the input must be (8 Oct, brain only): "The turn command appeared from about 30 per second (38% of our strongest input) on both sides; below that the model did not turn." That input is a rate we set, not a measured response to a real shadow.
What the map leaves out, part by part: our table on How it works lists what the map leaves out, with one number and one source for each part. A body simulator, not the brain, does the walking: with its forward-walking neurons stimulated, the FlyGym fly walked 13.9 mm in 1 s with the real brain, 14.1 mm with a scrambled brain and 14.3 mm with no brain.
Evidence: Shiu et al. 2024, which "does not account for gap junctions, non-spiking neurons, internal state or long-range neuropeptides"; the FlyWire entry; our sugar test (four scrambles, re-run bit-exact on 5 Oct 2026); Who does the walking?; We ran the map; Where in the wiring is the turn?; How strong must the shadow be?
How sure: one model, one fly's map, short open-loop tests with our own hand-made input and output choices; other models or real flies may differ. The looming body test has 6 real runs and 3 scrambled body runs; the 6 and 8 Oct follow-ups are brain-only, with 3 turned-up scrambles and one input series per eye.
Did a dead fly come back to life in a computer?No. The map was traced from one preserved fly brain, cut into thin slices and imaged; models run it as a simplified simulation, without recordings of that fly's activity.
No. The wiring map comes from the brain of one adult fruit fly, which was preserved, cut into very thin slices and imaged with an electron microscope (Zheng et al. 2018, Cell). Researchers traced its neurons and the connections between them (Dorkenwald et al. 2024, Nature). Models such as the Shiu et al. 2024 brain model turn the map into a simplified simulation: every neuron is a simple spiking unit (leaky integrate-and-fire), with connection strengths taken from the map's synapse counts and predicted neurotransmitters (Shiu et al. 2024, Nature). Its parameters come from the wiring map, predicted neurotransmitters and earlier modelling studies; we found no recordings of that fly's activity among them (Shiu et al. 2024, Methods). Whether the map keeps any trace of that fly's own experience is an open question; we found no study that tests it (checked 29 Oct 2026).
Several videos describe this as a fly brain being "uploaded" (see our verdict row, 29 Oct 2026). A commentary video argues the same point from the other side: “They Lied to you about the Virtual Fly Brain” (10 Physics), a video explainer by 10 Physics (16 Oct 2026; about 411,000 views on 29 Oct 2026) that argues many people online have been overstating what the fruit fly connectome actually is. We did not watch it (gallery card).
Evidence: Zheng et al. 2018, Cell ("complete brain of a female adult fruit fly", 7,062 serial thin sections); Dorkenwald et al. 2024, Nature; Shiu et al. 2024, Nature (model and Methods); Is it a brain or just a map?
How sure: "no recordings of that fly's activity" is our reading of the model's parameter list together with the fixed tissue; the paper does not state it in those words. "No study that tests it" rests on one web search on 29 Oct 2026.
Is the digital fly being tortured? Can it suffer? Is it alive or conscious?We found no study that measured consciousness or suffering in these models, which leave out much biology. Fly larvae detect harmful heat and pressure; whether flies feel pain is open.
The fly in these demos is a computer program, a model: simplified neurons that copy the wiring map of one fly's brain. We found no study that measured or even claimed consciousness or suffering in these models, and they leave out internal state, neuromodulators and non-spiking dynamics. A game demo usually runs between a few hundred and 166,700 simplified model neurons, with inputs and outputs chosen by a person. Real fruit flies are a different question: their larvae have nociceptors, sensory neurons that detect harmful heat and pressure. Detecting harm is not the same as feeling pain, and science has not settled whether flies feel it.
Evidence: Shiu et al. 2024 (what the model leaves out); PLoS ONE 2012: "The Drosophila TRPA channel Painless is required for the function of polymodal nociceptors which detect noxious heat and noxious mechanical stimuli."
How sure: whether insects feel pain is debated; detecting harm is shown in fly larvae, felt pain is not shown either way. Our statement about the models covers the studies and projects we read, not every project.
Could this become a 'hell' for future simulated brains?An essay argues that shared, steerable brain models could be misused to cause suffering. Today's fly models leave out much biology; whether any of them, or a future one, can feel anything is not known.
Some writers argue that as people get better at reading and steering simulated brains, and as these models are shared openly, some people will use them to try to cause suffering ("There Will Be Hells", LessWrong, 20 Oct 2026). Today's fly models leave out gap junctions, neurons that do not spike, internal state and long-range neuropeptides, and the brain model starts every neuron at zero background activity (Shiu et al. 2024); the essay itself calls the fly model oversimplified. We found no study that measured consciousness or suffering in these models, and whether flies feel pain is open. Whether any of these models, or a future one, can feel anything is not known.
Evidence: "There Will Be Hells" (LessWrong, 20 Oct 2026; about 290 karma and about 70 comments on 29 Oct 2026, approximate); Shiu et al. 2024: the model "does not account for gap junctions, non-spiking neurons, internal state or long-range neuropeptides, and assumes that the basal firing of each neuron is zero"; Is the digital fly being tortured?
How sure: whether insects feel pain is debated; detecting harm is shown in fly larvae, felt pain is not shown either way. Our statement about the models covers the studies and projects we read, not every project. The essay's argument is about the future and is its authors' view, not a measured result.
Can a fly brain learn maths, to hack or to play a game?Some projects say so. By our reading of 13 of them, practice changes a trained readout (3), trained settings across the network (4) or a learning rule on some connections (3); 1 trains nothing; 2 give no code. 3 compare with scrambled wiring.
Several projects say a fly brain learns maths, a game or hacking (their words). We read what changes with experience in 13 projects (some make such claims, others are research tools), at one fixed version each. By our reading: 1 trains nothing; 3 train only a readout on a fixed network; 3 use a learning rule that changes some connections during a run; 4 train settings across the network (strengths, gains or biases) by self-play, search or copying a teacher; 2 videos, including the hacking one, give no code, so we could not check. The maths project's README says "no fly learned anything". 3 compare with shuffled or random wiring: the authors report the real wiring did about the same (ChessFly, malecns-keratitis-pilot) or did worse (fly-arithmetic).
Update, 24 Oct 2026: we re-read fly-chess. Its committed results now keep the measured connectome frozen and train only a linear readout, and under the same short readout probe a shuffled wiring scored higher than the real one (20.9% against 18.0% top-1, author's numbers, not re-run by us). By our reading it moves from "trained settings across the network" to "a trained readout", so 4 of the 13 now compare with shuffled or rewired wiring.
In people's words:
- Can a Fruit Fly's Brain Learn Maths? (krezio/fly-arithmetic, README title)
- Can A Conscious Fly Brain Learn how to Hack? (dzuma, YouTube video title)
- A real fruit-fly brain learns to play Clash Royale, just from winning and losing against itself. (Zaib111/fly-clash-royale, README)
- A trained readout on a fixed network: 4
- Trained settings across the network: 3
- A learning rule on some connections: 3
- Nothing trained: 1
- Video only, no code: not checked: 2
| Project | What is trained (our class) | What the project's files say | Wiring comparison |
|---|---|---|---|
| fly-arithmetic | readout only | The README says only a linear readout is trained to answer the sums while the roughly 2,000-neuron FlyWire network stays fixed, that no fly learned anything, and that a random network of the same size did better. Source | shuffled or rewired |
| fly-chess | readout only | The flies keep the measured connectome frozen and train only a linear readout on descending neurons to imitate Stockfish moves (training the connection strengths too made held-out accuracy worse); the dopamine version changes only Kenyon cell to MBON weights. Under the same short readout probe, a shuffled wiring scored higher than the real one (20.9% against 18.0% top-1). Source | shuffled or rewired |
| malecns-keratitis-pilot | readout only | Only a PCA and logistic readout on frozen-circuit spike counts is fitted to culture labels of training patients; the circuit does not change. Source | shuffled or rewired |
| Let It Flyppen (five FlyWire brains with trained readouts play a Tame Impala song) | readout only | The README says the connectome is frozen and only a linear readout learns, mapping output-neuron spike counts to the song's notes, loudness and length. Source | none |
| FlyRoyale (Zaib111/fly-clash-royale) | whole network | The README says self-play reinforcement learning (PPO) trains per-neuron or per-synapse gains, per-neuron biases, the input encoder and the output decoder, while the connections and their signs stay fixed. Source | none |
| fly-jumpshot ("ts fly ballin") | whole network | The code comments say the hidden-layer wiring is never trained and a cross-entropy search tunes per-input gains, per-cell gains and biases of the hidden cells, and the read-out weights. Source | none |
| Flyhard (The Driving Fly) | whole network | The code gives every measured connection a trainable gain and every neuron a trainable leak, and the README says the trained model turned a wheel on 100 of 100 held-out targets after training against 0 of 100 before (one seed); the only comparison is untrained versus trained on the same wiring. Source | none |
| adfly-connectome | plasticity rule on part of the network | In the simulated odour conditioning assay, Kenyon cell to MBON synapses in compartments whose dopamine neuron type fired are scaled down by a fixed factor when the Kenyon cell spiked during pairing; the study measures how pathology reduces this learning index. Source | none |
| Stonkfly | plasticity rule on part of the network | The documentation says profit or loss pulses into chosen dopamine cells drive a candidate memory rule that changes 7,835 existing Kenyon cell to MBON07/MBON11 connections, and the README says this does not establish profitable trading. Source | none |
| ChessFly | plasticity rule on part of the network | The README says a dopamine-like prediction-error rule changes only the Kenyon cell to output-neuron synapses (plus a board-to-output shortcut) after each self-play game, and that the real wiring did as well as the same cells with shuffled inputs. Source | shuffled or rewired |
| flyonenomics | none | The README says the model has no learning: input is injected into chosen neurons and the results are spike activity under dopamine and transmitter perturbations of the same wiring. Source | none |
| Fly brain learns to hack (dzuma, video claim) | not checked | The video description says an adapter turned computer actions into simulated odour cues, the fly brain chose between them, and correct choices were rewarded with a dopamine signal over three weeks; it does not say which part of the model changes. Source | none |
| Fly plays Gorilla Tag (video claim) | not checked | The video description says the fly 'learned' to move as a Gorilla Tag avatar and to play Infection over seven phases; it does not say what is trained or how. Source | none |
Evidence: the project files linked in each row (read at the commit in the link); the catalogue, where each of these entries has a "What is learned" note; wiring comparisons in the controls ledger (ChessFly, fly-arithmetic, malecns-keratitis-pilot). See also Is it AI?
How sure: The classes are our reading of each project's code or README at one commit; projects change after we read them, and some have optional modes that train other parts. Video claims without code are restated as the creator describes them and not checked. A shuffled-wiring comparison reported by an author was not re-run by us.
Can I put the fly brain in my game?Yes for a non-commercial project, with credit (our reading of FlyWire's CC BY-NC 4.0). You must design yourself how the game talks to the brain.
The FlyWire brain data is licensed CC BY-NC 4.0 in FlyWire's guidelines: credit, non-commercial. The male fly data (MaleCNS) is CC BY 4.0, which allows commercial use with credit. The Shiu et al. model code and fly67 are MIT, FlyGym is Apache-2.0. A browser demo downloads 3.8–36 MB in its first minute and can run the whole brain in your browser; in Python, one simulated second took us 55–177 s on 2 CPUs, and the model needs two files, 104.1 MB (see What do I have to download?). You choose which neurons the game stimulates and which neurons move the character, and that choice is yours, not the fly's. In our catalogue 51 entries use data under CC BY only, 37 include CC BY-NC data, and 21 use other or unstated terms (a rough bucketing of the licence texts).
Evidence: FlyWire guidelines; licences in Build your own; our browser checks (download sizes, 3 and 6 Oct 2026); our looming steps (run times); the catalogue (each entry's licences).
How sure: not legal advice. FlyWire's guidelines say CC BY-NC 4.0 while its Zenodo v783 record says CC BY 4.0; which applies is an open question with FlyWire, so we follow the stricter, non-commercial reading. Check each project's own licence too.
What do I have to download, and how big is it?To run the brain model yourself: two files, 104.1 MB. The full official FlyWire release is 10.6 GB; the biggest public fly dataset is 536 GB.
The Shiu et al. model needs two files from its repository: Connectivity_783.parquet (100,804,642 bytes, the wiring) and Completeness_783.csv (3,327,347 bytes, the neuron list), plus the Python packages. You do not need the full FlyWire v783 release, which is 10.6 GB in 5 files. The male fly's brain and nerve cord (MaleCNS v1.0) is 31.3 GB in 11 files, of which the wiring alone is 508 MB; BANC is 536.1 GB in 379 files, 308.3 GB of them on request. A browser demo downloads 3.8–36 MB in its first minute instead. Once installed, the beginner project took about 660 MB of disk: a 464 MB Python environment and a 196 MB model folder with its data (our re-test of 2 Nov 2026).
Evidence: our download facts table of 6 Oct 2026, read from each host's storage metadata without downloading the data (How it works; Build your own); our browser checks.
How sure: sizes are the hosts' own figures on 6 Oct 2026 and can change with new releases. The MANC and larva L1 sizes are not stated: we did not list the MANC bucket and the larva supplement page did not answer. FlyWire's licence is open: its Zenodo record says CC BY 4.0, its guidelines CC BY-NC 4.0; we follow the stricter reading.
Is the fruit fly brain open source?The main wiring maps are free to download, under different licences. FlyWire's two sources disagree (CC BY-NC 4.0 on its own site, CC BY 4.0 on its Zenodo record), so we follow the stricter non-commercial reading (checked 29 Oct 2026). Code licences vary by project.
Updated 29 Oct 2026. The main wiring maps are free to download, under different licences: FlyWire (FAFB): CC BY-NC 4.0 on FlyWire's own site (flywire.ai/guidelines), while its Zenodo record lists CC BY 4.0 (zenodo.org/records/10676866); both checked 29 Oct 2026, and we follow the stricter non-commercial reading. MaleCNS: CC-BY (male-cns.janelia.org, checked 29 Oct 2026). Hemibrain: CC-BY (janelia.org hemibrain page, checked 29 Oct 2026). BANC: CC BY 4.0 (BANC project README, checked 29 Oct 2026). Code licences vary by project, and each catalogue entry lists its own. Free to download does not always mean free for any use: FlyWire's own site gives a non-commercial licence, so we treat its data as not allowing commercial use without permission.
- Data and code are licensed separately. "The fruit fly brain" usually means a measured wiring map (a connectome) plus a model that runs on it. The Shiu et al. whole-brain model code is MIT. 30 code repositories in our catalogue have no licence file, so reuse is not clearly allowed.
- Whole projects can mix licences: for example, neural-canvas's LICENSE file names MIT for the app code, CC BY 4.0 for the MaleCNS data and Apache-2.0 for the body and kernels.
- Where to download: sizes and links are in our download facts table (How it works) and the beginner guide.
Evidence (each checked 29 Oct 2026): FlyWire guidelines: "FlyWire's public release data is made available under license CC BY-NC 4.0"; Zenodo record of FlyWire v783 (licence CC BY 4.0); MaleCNS download page; hemibrain page; BANC project README; the code licence of each project in the catalogue; Licences: code and data are different.
How sure: licences are what each host or repository stated when we read it, and they can change. The FlyWire conflict is still there at source on 29 Oct 2026, and we have not resolved it. This is not legal advice.
What do I need to run it?For our beginner project we measured one machine on 2 Nov 2026: 2 CPUs and no GPU, a simulation process that peaked at 0.59 GB of memory, and about 660 MB of disk. Larger projects need more.
Our beginner project ran on a computer with 2 CPUs and no graphics card: no NVIDIA device or driver was present, and Brian2 used its RuntimeDevice with the numpy code target, which the demo picks because the computer has no C++ compiler. The simulation process itself peaked at 0.59 GB of memory for one 1-second trial, and the tests at 0.61 GB. The folder with the install, the code and the data took about 660 MB of disk: a 464 MB Python environment and a 196 MB model folder with its data. One sugar trial took 96 to 112 seconds from start to finish on 2 Nov 2026 (68 seconds on 19 Oct; we do not know why it was slower). We tested one machine. Larger projects need more; for example, FlyGym states that its new version runs at about 2x real time on a CPU and about 50x real time on a GPU (FlyGym README, checked 29 Oct 2026).
- Memory: These figures are the largest single process (measured with getrusage), not the whole test container, which peaked at about 1.7 GB including file cache and the parent process. Our guide suggests about 1 GB free.
- Graphics card: not needed for this project: on 2 Nov 2026 it ran on the CPU of a machine without one. We did not test whether a graphics card would make it faster.
Evidence: our beginner re-test of 2 Nov 2026 (Build your own; memory chart in How much memory, and of what); FlyGym README: "~2x real-time throughput … ~50x real-time throughput"; Shiu et al. brain model code.
How sure: one machine, one beginner project, measured on 2 Nov 2026; other projects, data versions and computers will differ, and run times on our shared 2-CPU machine varied between test days. The memory figures are single processes; a version that ran trials side by side would need more. The FlyGym speeds are the project's own statement, not our measurement.
Has a fruit fly been "uploaded" to a computer?
No. What exists is (1) a measured wiring diagram of one fly's brain (FlyWire, 139,255 neurons in v783) or whole nervous system (MaleCNS v1.0, 166,700 neurons), and (2) simplified models that run on that wiring.
The best-tested model, Shiu et al. 2024, "does not account for gap junctions, non-spiking neurons, internal state or long-range neuropeptides, and assumes that the basal firing of each neuron is zero", and "neuromodulation … is not accounted for", in the authors' words. By Eon's own account, its embodied fly moves with pre-trained controllers switched by a few hand-picked neurons, and that code is not public.
Sources: Shiu et al. 2024; Eon Systems.
What do the "91%" and "95%" numbers refer to?
Both come from Shiu et al., Nature 2024, and describe the brain-only model, not an embodied fly or a game.
- 91%: "Across 164 predictions we were able to test empirically, 91% were consistent with our empirical results." Without the optogenetic screen, the accuracy is 84%.
- 95% is not an accuracy. "Decreasing the strength of inhibition results in predictions consistent with 95% of those of the default model; accuracy drops from 91% to 88%." It compares two versions of the model with each other.
A third-party repository (not Eon's) claims "95% motor behavior accuracy" for an embodied fly. We found no measurement behind that number.
Source: Shiu et al. 2024, full text.
Will a mouse or human brain be next?
Not soon. The largest mammal connectome so far, MICrONS, covers about one cubic millimetre of mouse visual cortex, with more than 200,000 cells and about 523 million synapses (Nature, April 2025; figures from a search summary). That is a tiny part of a mouse brain, and it was made for anatomy, not whole-brain simulation.
The fly results depend on the complete wiring of a whole nervous system plus experiments that check the model's predictions. Neither exists yet for a mouse or a human. For a field-wide view, see the State of Brain Emulation Report 2025.
Controlled tests: does the real wiring matter?
The most useful projects run the same test with shuffled or random wiring. If the shuffled brain does just as well, the real wiring did not help. We keep every such study in one ledger: 107 studies, 97 of them with a wiring null. The real wiring helps in 36, makes no difference in 27, does worse in 11, gives mixed results in 18, and 5 is not yet scored. New on 11 Oct 2026: five rows by the authors' own numbers, among them fly-arithmetic (worse: a random network of the same size did better at addition) and MelanoGraph (helps, but weak: one rewired graph). New on 8 Oct 2026: Connectocopter ("helps": 15 of 15 looming escapes with the real wiring vs 0 of 15 rewired, but a hand-written controller also escapes 15 of 15) and Synaptera (not scored: its shuffle is not a wiring null we could classify); our looming test now has 6 real runs and an inside-only arm. On 6 Oct: our looming test became "helps" with a fair null and strong method, after the activity-matched scramble; fly tennis and FlyBrain Flappy ("helps", one shuffle each, weak), Fly-Racer ("mixed"); and Fly Dino corrected: a network with no connectome does as well.
- Reflexes, senses and steering: mostly yesFor fly-like reflexes, such as sugar taste driving the feeding neuron, for sensory circuits and for steering a simulated body toward a target, the real wiring beats scrambled wiring. It is a low bar: most of these tests ask whether a pathway still carries a signal after rewiring.
- Trained, reservoir and ML tasks: noFor reservoir computing, hashing, chess (two studies) and trained Doom players, the fly wiring did no better than shuffled wiring, and sometimes worse. The one machine-learning exception, a redesigned CartPole flight circuit (3 Oct 2026), beats its shuffled wiring but ties its own linear map. And in Doom, a brainless autopilot plays about as well as the untrained fly brain.
Only 39 of the 107 studies reach "strong" method quality, every number except our own three tests is the authors', and none is replicated. Our own pre-registered reflex test counts as "mixed" with an unfair null and weak method by our fixed rules, because its gain-matched check did not match on 3 shuffles; no rule was changed to rescue it. The chart, the counts by task family, a filterable table of all 107 studies with links, and a checklist for fair controls are on Does fly wiring help? The table that stood here until 29 Sep 2026 (20 studies) is replaced by that ledger.