Research report · 30 September 2026 · fourth check
Fourth check: does fly wiring help? 31 control studies compared, a FlyLeno verdict and 12 new entries
This check asked one question across the whole catalogue: when a project scrambles the fly's wiring, does the result get worse? We collected 31 control studies into one ledger and compared them on a single scale. We also refereed the week's most-watched new claim, FlyLeno, re-checked all 82 entries and 47 videos, and added 12 entries and 2 videos.
Short answer
Does fly wiring help? Sometimes: real fly wiring beat scrambled wiring in 10 of 23 studies, mostly reflex, sensory and steering circuits; in trained, reservoir and ML tasks it tied or lost (7). Of 31 control studies, 23 have a wiring null: the real wiring helps in 10, makes no difference in 5, does worse in 2, gives mixed results in 5, and 1 is not yet scored. Where the wiring helps, a brainless baseline often still matches the fly: in Doom the fly survives 53.3 s, shuffled wiring 6.8 s and a brainless autopilot 51.9 s.
How sure we are: fairly sure of the split, less sure of any single number. Only 9 of 31 studies reach "strong" method quality. Every number except our own run is the authors'; we recomputed 15 studies from their committed files but re-ran none, and none is replicated.
One claim met our attention bar: FlyLeno ("Tonight's host: Grey Leno, piloted by Drosophila melanogaster"), after a streamer's upload rose to 36,763 views. Graded from its code as B, borderline C: the whole FlyWire brain really runs, untrained, in the browser, but inputs and outputs are mapped by hand, gait, homing and the choice of actions are engineered, and there is no control. The Pokémon stream and the Rainbow Six claim were re-checked and stay U.
What changed
- 82 → 94catalogue entries (12 new: 2 A, 5 B, 2 C, 1 D, 2 not graded)
- 47 → 49gallery videos (2 new, 0 removed)
- 31control studies in a new public ledger, with a new page
- 0main-link failures and grade changes; 3 new releases
| Measure | Value |
|---|---|
| Entries re-checked | 82 of 82, plus 12 new |
| Repositories with new commits / new releases | 7 / 3 |
| Grade changes | 0 (three entries re-checked because of new commits) |
| Main-link failures (4xx, 5xx or error) | 0 (one host showed a bot check, which is not a dead link) |
| Videos re-checked / new / removed | 47 / 2 / 0 |
| Data format | Two new optional fields on every catalogue record: controls and wiring_effect |
New releases: acamilo-flybrain v0.6.7 (a damage reward for the Pokémon battles; it changes the reward signal, not the mechanism, so the grade stays C), FlyWire annotations Version 3.2.0 (29 Sep; an annotation release, not a new connectome) and Kick the Fly 2.13.1 (tests and tooling; stays C).
Other new commits: fly-brain-zero-shot (an install fix in the README), flyconnectome-nulls (author affiliation only), flydoom by mutkuoz (13 commits re-scoring a separate "everything applied" variant; the plain model's scorecard and its grade A are unchanged) and Haltere (round-6 flights passed the Minus Two hairpin for the first time, still with no race finish; stays C). For the ledger studies whose code moved, the numbers were read at the new commit and did not change.
Held-back entries: FlyKart (B) and making-fly-play-chess (C), held back in the third check, were re-verified at their current commits and published.
What people argued about this week
We looked for new fly-brain claims since 29 Sep 2026, 09:50 UTC. The promotion rule is unchanged: a claim must be specific and checkable, have gained attention (30 points on Hacker News, 20,000 video views, 20 GitHub stars within a week, or a news story) and not be covered yet; at most two per check.
| Source | Hits | Relevant in the window |
|---|---|---|
| Hacker News stories / comments | 1 / 60 | No fly-brain story |
| GitHub | 277 (98 unique) | 22 new repositories, each with 0 or 1 star |
| YouTube | 242 (121 unique) | Vinesauce's FlyLeno stream at 36,763 views; in-window uploads all under 40 views |
| Google News | 28 | No new story |
| 0 | Refused (HTTP 403) without a login |
- Promoted (1): FlyLeno. The stream upload grew from 16,201 views (29 Sep) to 36,763 (30 Sep, 09:41 UTC). The viral video is a streamer's playthrough; the claim we graded is the site's and the repository's own.
- Re-checked, no change: the Pokémon stream's creator still has no public code (their GitHub address returns "not found"), and no recording or follow-up article appeared; the Rainbow Six report is still behind a bot check, and a web search found only the original article. X cannot be searched without a login.
- Not promoted: "I Taught a Fly How to Play Gorilla Tag" (415,612 views) says the fly "learned" with "its open source brain" but links no code. It was uploaded on 26 Sep, before the window, and our date-sorted video search had missed it; it is flagged for a later check. Two other older videos (24,575 and 11,934 views) link no project.
- Data releases: FlyWire annotations Version 3.2.0 (29 Sep). No new connectome release.
Verdict: "Tonight's host: Grey Leno, piloted by Drosophila melanogaster"
FlyLeno (TUURD Talk) is a browser show by AgitationSkeleton in which "a whole-brain spiking model of the adult fruit fly (FlyWire v783, 138,639 neurons, 15.1M connections) runs live and puppeteers Grey Leno" (README). We read the code at commit 078d8fa:
- Wiring: the Shiu et al. v783 files, packed with every connection kept (
tools/build_connectome.py). - Neurons: the Shiu et al. leaky integrate-and-fire equations, event-driven in a Web Worker (
js/brain-worker.js). Dopamine-gated plasticity, on by default, changes only the synapses onto the voice neurons. - Inputs and outputs, hand-set: show events, thrown objects, music and food stimulate chosen sensory groups; descending-neuron rates become walk, turn, startle, groom and feed commands through fixed gains of 40 Hz (15 Hz for startle) (
js/motor.js). - Engineered layers, labelled by the author: a walking rhythm generator, balance "puppet strings", saccades, food seeking, homing and sleep (
js/instincts.js), and an action selector outside the connectome that picks spontaneous actions and stimulates the matching descending neurons (js/mind.js). - Controls: none. A search of the code for shuffle, scramble or rewire finds only playlist and animation code.
B Verdict, borderline C: the fly brain is real and really running, but Leno's show is mostly built around it. The brain's reflexes move him; his walking rhythm, homing and what he does next are engineered; and nothing tests whether a scrambled brain would look the same. Up to A with a scrambled-wiring or no-brain run of the same show and a measured behaviour; down to C with evidence that the engineered layers drive most on-screen behaviour. Verdict row · catalogue entry
Other rows on Is it real?: the Doom row now adds that the Doom control study's results file (results/a2/summary.json) holds a sweep of 12 autopilot settings, the best of which survives 53.44 s, slightly longer than DOOMFLY's 53.34 s, next to the matched autopilot's 51.93 s; we checked this in the file. The controls section now points to the new ledger instead of its earlier 20-study table.
Does fly wiring help?
Full page: Does fly wiring help? Data: controls.json and its schema.
| Task family | Helps | No difference | Worse | Mixed | Not tested |
|---|---|---|---|---|---|
| Reflex circuits | 4 | 0 | 0 | 0 | 0 |
| Sensory models | 2 | 0 | 0 | 0 | 1 |
| Steering a body | 2 | 0 | 0 | 1 | 2 |
| Playing games | 1 | 2 | 0 | 2 | 4 |
| Machine-learning benchmarks | 0 | 3 | 1 | 1 | 0 |
| Reservoir computing | 0 | 0 | 1 | 0 | 0 |
| Language models | 0 | 0 | 0 | 0 | 1 |
| Forecasting | 0 | 0 | 0 | 0 | 1 |
| Graph analysis | 1 | 0 | 0 | 0 | 0 |
| Evolved controllers | 0 | 0 | 0 | 1 | 0 |
| All 31 studies | 10 | 5 | 2 | 5 | 9 |
Of the 9 "not tested", 8 studies have no wiring null (only baselines or ablations). The ninth, bioreservoir, has one, but its forecast questions resolve only on 15 Dec 2026.
Which null is fair
The ledger lists 16 kinds of null and control. Where the choice of null changed the result: in flyconnectome-nulls, standard nulls beat the connectome (1.78 and 1.77 against 1.57) while nulls that keep the sensory-motor boundary tie it (1.58 and 1.62); in the Arkanoid chapter of fly-plays-games a weight shuffle keeps the play and a rewiring loses it; in FLY-lab, re-fitting the readout for each null shrinks the gap from 69 to 47 points; reservoir results depend on each wiring's gain, and the larva study, which fixes it, finds no difference; and in Doom the brainless autopilot matches the brain. The headline chart uses the degree-preserving null first (19 of the 23 studies with a wiring null). The four-point fair-control checklist: keep degrees and signs; tune, train and re-fit the null like the real wiring; use 3 or more shuffles and report the spread; include a no-brain baseline and say whether the fly beat it.
Replayed vs live, brain vs no brain
Replayed brain output scores 0 of 30 on FLY-lab's turning task against 100% live; in the zero-shot study, disconnected turn neurons give 0.125 poles per 10 s, scrambled wiring 0.34 and the live real wiring 2.96. No study shows replay matching live performance. 13 studies have a no-brain or no-graph baseline, and the fly network beats it in only 2: fly-dino (179 s against 46 s for a hand rule, but with no wiring null) and the Fly's Hash Function (beats SimHash while losing to its wiring nulls).
| Measure | Studies |
|---|---|
| Quality: strong / moderate / weak | 9 / 6 / 16 |
| Null fairness: fair / partly fair / unclear / unfair | 11 / 15 / 4 / 1 |
| Numbers from result files / papers / catalogue text | 27 / 3 / 1 |
| Recomputed by us from committed files | 15 |
| Our own measurement | 1 (the Build your own shuffle run) |
| Peer-reviewed | 2 (the Shiu et al. model and flyvis) |
What would change the answer: a behavioural study with a fair null (3 or more degree-preserving shuffles, tuned equally) and a no-brain baseline in which the real wiring beats both (none exists); boundary-preserving nulls applied to the reflex-circuit studies; independent replications; and our own fair-null test of the Shiu et al. model with weight, degree-preserving and sign shuffles, 5 or more each, which is a candidate for a coming check.
New entries
Every entry was graded from files we opened. The numbers are the authors' own.
| Entry | Kind | Grade | Controls | Deciding evidence |
|---|---|---|---|---|
| fly-cartpole | Research | A | Wiring null: no difference | Mushroom bodies wired from MaleCNS learn CartPole (233.8 steps over 20 held-out seeds); degree-preserving shuffles learn about as well (218.3, p = 0.20; recomputed by us); a TD learner does better (302.3). |
| flybrain-connectome-benchmark | Research | A | None (checked against fly data, not a null) | A pre-registered test of the Shiu et al. whole-brain model against published experiments: knockout sensitivity 0.769, specificity 0.989 (preprint). |
| FlyLeno | Browser demo | B | None | See the verdict above. |
| FlyKart | Game | B | None | The whole MaleCNS brain steers a go-kart, but the game computes what the fly "sees" and the controls are hand-mapped; a new CPU path runs without a GPU. |
| Open Fly | Browser demo | B | None (a shuffle control is pre-registered, with no result yet) | The whole Shiu et al. brain plays a strategy game in the browser; game events go to taste neurons and 39 arbitrary action groups. |
| Fly With Me (Fly Brain DJ) | Art | B | None | The FlyWire brain hears music through its antennal neurons and triggers DJ moves from named descending neurons; the mixing is ordinary audio code. |
| Fly-NAF | Game | B | None | The FlyWire brain plays Five Nights at Freddy's; the seeing is a pixel-difference threshold and each action comes from one chosen neuron. |
| making-fly-play-chess | Game | C | Wiring null: no difference | A fixed FlyWire patch with a trained readout ties its rewired twin (0.506) in a run the author marks as a code check, and loses to a material-count player. |
| FlyCNS Tic-Tac-Toe | Game | C | None | By default a minimax solver picks the playable moves; the fly circuit only breaks ties. |
| FLYBRAIN Bad Apple x DOOM | Art | D | None | A connectome viewer that uses neurons as pixels; no brain activity is simulated. |
| connectome_interpreter | Data tool | n/a | None | A library for turning wiring diagrams into testable circuit hypotheses (MIT, with a preprint). |
| FlyBrainLab | Data tool | n/a | None | An older interactive platform for fly brain data and circuits (BSD-3); last commit September 2025. |
Left out: 31 other GitHub leads, each with a reason in our exclusion log: most have 0 or 1 star and no demo or control; two promote crypto tokens; two are off-topic; and two repositories that copy FlyGym's name were not opened and are not linked. The first in line for the next check include a project with its own controls section. The Fly Brain Hub directory added 10 repositories.
Gallery
49 videos, 2 new: Vinesauce's FlyLeno stream, labelled as a streamer's playthrough, not the author's upload (listed because the project links no video of its own), and Fly-NAF's best run, the author's upload. All 47 earlier videos were available, and every card's grade matches the catalogue. Not admitted: the Gorilla Tag video (no project to map it to) and uploads without a project link. The Pokémon and Rainbow Six rows still have no video we could admit, and 9 of the new entries have none. Thumbnails are unchanged.
Build your own: pins and the "Add a body" path
Pins unchanged. No pinned package of the beginner guide has had a release since 27 Sep 2026; newer brian2 2.10, numpy 2.5 and pandas 3.0 predate our pin test. The next full re-test is due by 12 Oct 2026.
FlyGym: the latest release is still 2.1.0 (24 Jun 2026). It requires Python 3.12 to 3.14 (>=3.12,<3.15); the last release for Python 3.11 is 1.2.1, with the old interface. It has no hard GPU dependency, and its download is about 170 MB (an estimate from package metadata). Our research host has Python 3.11, so a brain-body test is not possible there without an extra Python 3.12 install; that choice is open. The body path on Build your own therefore stays as it was: installing FlyGym 2.1.0 and running its physics on a CPU were tested on 28 Sep 2026 with Python 3.12, and connecting it to the brain is untested.
How often things change
Three daily checks so far (28, 29 and 30 Sep 2026): 0 main-link failures in 213 link checks and 0 video removals in 131 video checks, while the share of repositories with new commits rose from 0% to 4% to 9% per day, mostly young demo repositories. We keep checking links and videos on every run, re-grade entries when their commits suggest a change (no upstream change has altered a grade so far), and refresh stars weekly except for young repositories.
Method
- Lead scan: Hacker News, GitHub, YouTube and Google News searches since the end of the last scan, plus video pages and re-checks of the two open claims.
- Tracking: every entry's main link, repository commits and releases, and every video link. For repositories that changed, we read the commit messages and re-checked grades where they suggested a change.
- Controls ledger: one study per catalogue entry with a wiring null or a no-brain baseline, plus our own run. Numbers were read from committed result files, READMEs or papers at the commits listed in the data file. For 15 studies we recomputed the reported numbers from the committed files, and 8 studies were spot-checked a second time against files fetched separately; all matched. Retention, wiring effect, fairness and quality follow fixed rules; a gap counts only if it exceeds twice the combined standard error. One label was set by hand (Fly.exe is "mixed", because its index can be negative), one headline arm was moved to a note (chess: its interval is for the fly's score, not the difference) and one task family was set by hand (flydoom by mutkuoz counts as a sensory model, because its only wiring null is on the smell response).
- GitHub data: anonymous API access. 59 tracked records keep stars and dates from earlier checks (27, 28 or 29 Sep), labelled with those dates; stars were fetched fresh for the 12 new entries and 11 young repositories.
- Our own measurement: the Build your own shuffle run of 28 Sep 2026 is the only number we produced ourselves.
Limitations
- X and TikTok cannot be searched without a login, and Reddit refuses scripts. View counts are single snapshots. Our date-sorted video search misses older videos that grow fast (the Gorilla Tag case).
- Every ledger number except our own run is the authors'. We recomputed 15 studies from committed files but re-ran none. 7 studies rest on README or docs tables and 3 on papers; the flyvis ablation values sit in a figure and were not extracted.
- Fairness and quality labels come from yes/no fields read from code. They are consistent but coarse: for example, the Shiu et al. paper result is "weak" only because the paper text does not give the shuffle count and spread.
- The FlyGym download size is an estimate from package metadata; the 602 MB installed size is our measurement of 28 Sep 2026.
- 29 GitHub leads were deferred without being opened.
Next
- Decide the "Add a body" path (a Python 3.12 environment, the older FlyGym 1.2.1, or neither), and whether to referee the Gorilla Tag video.
- Admit the next leads, including a project with its own controls section, and check the two FlyGym name copies before any link.
- Add a "most viewed this week" video search, and keep the ledger current: new control studies, Open Fly's pre-registered result, and bioreservoir's scores after 15 Dec 2026.
- Our own fair-null test (a candidate for a coming check): no study yet pairs a fair null with a no-brain baseline on a behaviour.
Main sources
- FlyLeno: repository (commit 078d8fa), show, Vinesauce stream.
- Doom control study: results/a2 (commit 6b22922).
- fly-cartpole: repository. flyconnectome-nulls: results.
- FlyGym on PyPI: flygym.
- Every study's sources, arms and commits: controls.json; every entry's sources: projects.json.