Research report · 29 September 2026 · third check
Third check: two viral fly-brain claims refereed, DesktopFly re-graded to C and a wiring test where scrambled wiring fails
This check started with the arguments of the week: which fly-brain claims got attention since our last check, and do they hold up? Two did, and neither has published code. We also re-checked all 73 entries and 44 videos, re-graded DesktopFly, added 9 entries and 3 videos, and found the week's strongest evidence in a small new control study.
Short answer
Two viral claims met our bar, and both are graded U (unverified) because no code has been published. The Raspberry Pi 5 "fly brain plays Pokémon" stream may well run a full FlyWire brain, but by the creator's own account the brain reads positions from the game's memory, the decoder is unpublished and chat steers the learning, so the stream cannot show how much of the progress comes from the fly. The "fly brain plays Rainbow Six Siege" claim is unlikely to be a fly brain at work: its "over 1,000,000 neurons" matches no fly connectome, and training on 20 TB of human gameplay would explain any skill.
DesktopFly, the most-starred demo (1,053 stars), moves from B to C after a line-by-line reading of unchanged code: its neurons mostly switch coded animations, and sleep follows the computer's clock.
The strongest new evidence is a grade A control study, fly-brain-zero-shot: a whole FlyWire brain steers a simulated body to targets, even after two legs are cut, while scrambled wiring or disconnected turn neurons fail (2.96 against 0.34 and 0.12 poles per 10 s, the authors' numbers). It is the positive counterpoint to the Doom test.
What changed
- 73 → 82catalogue entries (9 new: 1 A, 4 C, 2 U, 2 not graded)
- 44 → 47gallery videos (3 new, 0 removed)
- 1grade change: DesktopFly B → C
- 0main-link problems; 3 projects with new releases
| Measure | Value |
|---|---|
| Entries re-checked | 73 of 73, plus 9 new |
| New default-branch commits / new releases | 3 / 3 (Haltere, Kick the Fly, neuprint-python) |
| Grade changes | 1 (DesktopFly, B to C) |
| Licence / archive changes | 0 / 0 |
| Main-link problems (4xx, 5xx or error) | 0. One new main link (dev.ua, for the Rainbow Six claim) shows a bot check to our script |
| Existing entries with a material / minor update | 4 / 4 |
| Videos re-checked / new / removed / restored | 44 / 3 / 0 / 0 |
| Video link problems | 0 |
Material updates: Kick the Fly release 2.12.0 adds a larva brain for headless tests only (its two larva validation tests fail, as the author reports), flies that differ and a pet mode; the game still uses the adult brain and stays C. Haltere release fast-brain-11 (experimental) flew one full Straw Bale lap in 1:42.988 with a braking motor readout, but every flight ended in a crash and there is no race finish; documentation only, stays C. neuprint-python 0.6.4 is a client update, not a dataset change. FlyBrain-HalfLife is unchanged at commit 0ba5548, so its D grade stands.
Minor updates: FlyBridge's star count (6 to 7), the status of secondary links for the larva connectome and the Shiu et al. model (publisher and preprint sites showing bot checks), and Neural Canvas's re-check note (unchanged since 6 Sep 2026).
Link check: 188 addresses over 82 entries. 156 answered directly and 11 through a redirect. 19 journal, preprint, Reddit and news pages (doi.org, nature.com, eLife, science.org, bioRxiv, reddit.com, dev.ua) showed bot checks, one secondary source (liquipedia.net) refused our script, and one bioRxiv connection was reset. None of these is a dead main link.
What people argued about this week
We looked for new fly-brain claims since the end of our last check (28 Sep 2026, 15:00 UTC). A claim is promoted to a verdict when it is specific and checkable, has gained attention (at least 30 points on Hacker News, 20,000 video views, 20 GitHub stars within a week of creation, or a news story), and is not covered yet. At most two claims are promoted per check.
| Source | Queries | Hits | Relevant in the window |
|---|---|---|---|
| Hacker News stories and comments | 6 + 6, plus 4 by date | 8 + 67 + 80 | 2 stories, both with 1 point: the Roblox maze video and a fly tic-tac-toe project |
| GitHub, created or pushed since 28 Sep | 18 | 282 (106 unique) | 29 new repositories; the most starred has 1 star |
| YouTube, sorted by upload date | 16 | 241 (117 unique) | Vinesauce's "Fruit Fly Brain pilots Grey Leno" (16,201 views) and the Roblox maze race (35 views) |
| Google News (past 7 days) | 5 | 36 | XDA Developers (Pokémon on a Pi 5) and dev.ua (Rainbow Six Siege), both 28 Sep |
| Reddit search | 4 | 0 | Refused (HTTP 403) without a login |
| Dataset release pages and directory lists | 6 + 2 | – | No new dataset release; 13 new leads on Fly Brain Hub |
- Promoted (2): the Pokémon stream and the Rainbow Six claim. Both news stories appeared on 28 Sep, a few hours before the window opened; our previous check had no news scan, so both were new to us.
- Watched, not promoted: Vinesauce's Flyleno playthrough had 16,201 views about 8 hours after upload, below the 20,000 bar at our check, and it is a streamer's upload rather than the project author's claim.
- Logged only: the Roblox maze race (catalogued, see below), the tic-tac-toe project, and older items from before the window: a reaction video with 135,504 views, a second Half-Life video, a Left 4 Dead 2 mod and several "fly brain mines bitcoin" stories (crypto promotion, which we do not catalogue).
Verdict: "A digital fly's brain on a Raspberry Pi 5 plays Pokémon"
XDA Developers covered a Twitch stream by Leetzerzz on 28 Sep 2026. Quoted claims: "all 139,255 neurons of the FlyWire FAFB v783 reconstruction", simulated on the Pi next to the mGBA emulator; walls, doors and people are "read out of the game's memory" into looming-detector neurons, and "it isn't doing image processing — there's no retinotopy, which is the honest limitation"; "descending neuron activity decodes into button presses"; and viewers can zap reward and punishment neurons, with effects that "last for 10 minutes".
U Verdict: a full FlyWire brain may well be running on the Pi. But by the creator's own account, the brain is fed positions read from the game's memory and does not see the screen, the button decoder and any helper scripts are not published, and chat steers the learning. So the stream cannot show how much of the Pokémon progress comes from the fly. We found no code under the creator's name, and neither the article nor the stream links any.
Two open-source "fly plays Pokémon" projects exist. We graded both from their files to see whether either is the stream's code. Neither matches the creator's description, and no link names either of them. This rests on those mismatches, not on proof: the creator could still publish different code.
| The Pi 5 stream (as described) | acamilo/flybrain | fly-plays-games | |
|---|---|---|---|
| Dataset | FlyWire v783, connections of 5+ synapses | FlyWire v783, whole brain (139,255 neurons) | MaleCNS v1.0, through an external package |
| Input | Game-memory positions of walls, doors and people into looming detectors | Screen pixels, one per L1 column | Nearest sprite position read from game memory |
| Output | Descending neurons decoded into buttons (decoder not described) | Descending neurons split into 8 buttons by index; hand-set decoder | A trained readout picks only left or right |
| Emulator and hardware | mGBA on a Raspberry Pi 5 | Linux service; no Raspberry Pi build found | PyBoy on a laptop |
| What moves the character | Unknown | The author's 43-minute check found the direction scores frozen, with only the decoder changing the winner; the stream log records progress through scripted macros | The long walk in its video is a scripted planner |
| Grade | U | C (borderline: in raw-button mode without macros it would sit on the B/C border) | C |
What would change the verdict: published code, graded from its files; or a logged comparison on the same save of real wiring, shuffled wiring and a random button presser without chat input.
Verdict: "A fly's digital brain plays Rainbow Six Siege"
dev.ua reported on 28 Sep 2026 an X post of 26 Sep by the Rainbow Six Siege caster Oleksandr "GHOOD_BHOY" Samoilenko: a fly brain model with "over 1,000,000 simulated neurons" was trained for about a month on "more than 20 TB" of professional gameplay and scored its first kill. dev.ua blocks our fetches with a bot check, so we know its text only through search summaries; X cannot be read without a login.
U Verdict: unverified, and unlikely to be a fly brain at work. Nothing is published: no code, model or write-up.
- Where the skill would come from: training on human gameplay is ordinary machine learning. Unless a control shows otherwise, the skill comes from the training, as the Doom studies with shuffled wiring suggest.
- One kill is not a result: a single kill in a team shooter has no baseline.
- What would change the verdict: published code naming the connectome release, the neuron model and how frames and controls are mapped, plus a comparison with shuffled or no fly wiring.
The strongest new evidence: a closed loop where the wiring matters
fly-brain-zero-shot joins the FlyVis eye model, the Shiu et al. whole-brain model on FlyWire v783 with unchanged weights, and a NeuroMechFly body. The left-right difference of the DNa02 turning neurons steers the body toward dark poles. Two legs are removed mid-walk from cloned flies, and the same flies are rerun with scrambled wiring and with the turn neurons disconnected.
- Result: intact flies reach 2.96 poles per 10 s with real wiring, 0.34 with scrambled wiring and 0.12 with the turn neurons disconnected. After the two-leg cut: 3.14, 0.12 and 0.52.
- Why it matters: in Doom, a brainless constant autopilot played as well as the fly; here, a fly that walks without brain steering almost never reaches a pole. It is one of the few positive wiring controls in a closed loop.
- Caveats (why it is borderline A): walking comes from a hand-built rhythm generator, and the brain only steers through one hand-set formula with a gain tuned on separate flies. Scrambling every connection is a strong perturbation, the scrambled comparison after the cut has 16 flies from a single shuffle, and only the per-fly summary table is stored.
Re-grades and re-checks
| Entry | Before | After | Why |
|---|---|---|---|
| DesktopFly | B borderline C | C | Code unchanged at 32b0001. Neuron rates only cross hand-set thresholds, then coded programs run: flight follows a random-target timed path, grooming is a sine animation and darting a scripted turn. Sleep follows the computer's idle time and the clock. Only walking is driven by simulated leg neurons. |
| Kick the Fly | C | C | Re-checked after release 2.12.0, which touches data loading and the neuron model. The larva mode is headless only and fails its own tests; the game's reactions are still scripted programs that thresholds trigger. |
| FlyBrain-HalfLife | D | D | Still commit 0ba5548, with no new connectome loader. The video has 139,436 views (29 Sep). A second video by the same author says "There is no scripted game AI", which the same code contradicts. |
| Haltere | C | C | Only documentation changed; the new lap result was added to its measured results. |
Two corrections to our Is it real? row for DesktopFly: sleep is switched by idle time and the clock, not by a neuron rate, and the quoted claim is the GitHub repository description, not the README. The DOOMFLY control numbers were read again from doom-fly-control's results at the same commit and still match.
The 9 new entries
All numbers were measured by the projects' authors; we reproduced none of them. The full file citations are in each entry's "Evidence, limits and sources".
| Project | Kind | Grade | Why, in one line |
|---|---|---|---|
Are fruit flies zero-shot adapters? fly-brain-zero-shot | Control study | A | Eye model, whole FlyWire brain and NeuroMechFly body find poles, also after a leg cut; scrambled wiring or disconnected turn neurons fail. Walking is a hand-built rhythm. |
flybrain (Game Boy fly) acamilo-flybrain | Game | C | A whole FlyWire brain sees screen pixels and plays Pokémon Red for a stream, but the author's own check found the decoder choosing directions in long runs, and the stream uses scripted macros. |
fly-plays-games fly-plays-games | Game | C | A MaleCNS brain in Pokémon Red: the long walk in its video is a scripted planner, and in fly mode the brain only chooses left or right from sprite positions. |
Fruit fly vs Jev: Roblox maze race fruitfly-roblox | Game | C | 65% of the steering command (80% near walls) comes from a scripted helper that also remembers visited maze cells. |
NeuroWeave neuroweave | Research | C | 150 MaleCNS neurons as a trained mask. Against a random graph it scores 0.98 vs 1.00 (one seed, at the ceiling); the memory task has no such control. |
Pokémon on a Raspberry Pi 5 fly-pokemon-pi5-stream | Game | U | The Twitch stream above: no code; input read from game memory by the creator's account. |
Fly brain plays Rainbow Six Siege fly-brain-rainbow-six-siege | Game | U | The claim above: no code; the stated size matches no fly connectome. |
fly-connectome-template fly-connectome-template | Data tool | n/a | A browser template that renders MaleCNS cell positions. Its custom licence requires a linked credit on every web interface built with it. Its example activity file is made-up values on real IDs. |
fafbseg fafbseg | Data tool | n/a | An R package for FlyWire data: release downloads (v783 and v630), meshes and annotations. GPL-3. |
Our plan limits how many entries a check adds when viral claims need verdicts, so two further graded projects wait for the next check: flykart (B) and making-fly-play-chess (C).
Gallery changes
- All 44 published videos are still available, and so are the 3 new ones. Nothing was removed.
- 3 new videos: the 47-second video from fly-brain-zero-shot (A), the Roblox maze race on YouTube (C) and Haltere's Straw Bale race at 4× speed (C). The two project files are links only; nothing is re-hosted.
- 15 candidates left out: the Vinesauce Flyleno playthrough (not the author's upload; project not catalogued yet), two reaction and commentary videos, the second Half-Life video (it shows nothing new), videos of projects we have not catalogued (a Left 4 Dead 2 mod, a Roblox project, a robot), a compilation, an off-topic video, the Pokémon live stream (not a recorded video), and three project videos whose files are not committed or not linked from the project.
- Gaps: there is no video we could admit for the two new verdicts yet.
Backlog, copies and a safety finding
- Triaged: fly-connectome-template (published), fafbseg (admitted), NeuroWeave (admitted, C), 6 new repositories from the lead scan (2 admitted, 2 graded and held back, 2 deferred), and 5 deferrals from the first check (the DesktopFly Android port is described on the DesktopFly page; 4 deferred).
- Download lures (do not use): of 8 likely copies of catalogued projects, 6 are not copies at all: four repositories under other accounts that reuse the names DesktopFly, FLM and FlyDrones, each with a README and a zip file containing a Windows program, and two download pages on github.io. We never link them. The DesktopFly page explains how to recognise the one that copies DesktopFly.
- Plain copies left out: a copy of fly-connectome-template and a copy of Lulzx's fly-brain.
- Not triaged yet: about 211 repositories listed only in project directories, 13 new Fly Brain Hub leads and 94 other search hits.
Data files
The catalogue's enumerated fields (platform, gpu, peer_review) are now filled with allowed values at the source, with the fuller wording in the *_note fields. In the previous release our publish step had to convert free-text values in 13 records; now 21 values in 14 records are corrected at the source, our publish step changed nothing, and the published values are the same as before. projects.json (82 records) and videos.json (47 records) validate against their published JSON Schemas without errors.
Method
Tracking
- Last commits come from fresh shallow clones of each default branch, last releases from each repository's release feed, and archive status from the repository page, for every entry.
- Stars, creation dates and last-push dates for 56 records were reused from our checks of 27 or 28 Sep 2026; each record names the date it was actually fetched. The 8 recently created repositories and the 6 new ones got fresh values.
- Every main link and every video was checked again over HTTP. No simulation was run, nothing was installed, and no video was watched frame by frame.
Grading
- Grades rest on files we opened, cited with line numbers. The two U grades rest on the absence of code, with the claims recorded.
- Source labels: [direct] we opened it; [fetch-summary] and [search-summary] we read it through a summarising fetch or a web search (used for the XDA and dev.ua texts, Reddit and Twitch); [video] video metadata only.
- No grade above U rests on a video, a README headline or a search summary.
Limits and uncertainty
- Blind spots in the lead scan. X and TikTok cannot be searched without a login, and Reddit refuses scripts, so claims that live only there are missed unless news or YouTube picks them up. The dev.ua text is known only through search summaries.
- Attention figures are snapshots taken once at a stated time. Flyleno may pass 20,000 views later. Reddit and Twitch numbers could not be read.
- Stream identity. "Not the stream's code" rests on mismatches between the creator's description and the two repositories, and on the absence of any link. The creator could still publish different code.
- acamilo-flybrain's C partly rests on the project's own run log, which describes macro-driven progress; the live stream's mode setting cannot be seen from the repository.
- How closely grades were read. We re-opened the deciding lines for the DesktopFly re-grade and for fly-brain-zero-shot, fruitfly-roblox, acamilo-flybrain, NeuroWeave and fly-plays-games. The fafbseg citations and the two held-back grades were read once and not re-verified. fly-plays-games runs its neurons in an external package (fly.ai) that we did not inspect; we exclude fly.ai itself because it promotes a crypto token.
- Reused dates. Stars, creation and last-push dates for 56 records date from 27 or 28 Sep 2026 and are labelled.
- No browser, no builds. No demo was played and no project was built or run in this check.
- Not done this time: the FlyGym install check that would decide whether Build your own can add a fly body was postponed, because the two verdicts took priority under our plan.
Next check
- Re-check Flyleno's views; if it passes 20,000 views or gets news coverage, grade the Flyleno page's code. Look again for a code release or recording of the Pokémon stream and for the Rainbow Six clip.
- Publish the two held-back entries, flykart (B) and making-fly-play-chess (C).
- Run the short FlyGym install check for Build your own.
- Add fly-brain-zero-shot and the Arkanoid chapter of fly-plays-games to our summary of whether fly wiring helps.
- Work through the deferred backlog, and look for creator videos for the two new verdicts.
Key sources
- XDA Developers, 28 Sep 2026 [fetch-summary]; the creator's Reddit post [search-summary]; the Twitch channel [fetch-summary, title and description only].
- acamilo/flybrain (commit 9f2345d; docs/limitations.md, docs/stream-mvp-plan.md, the retina and decoder files) and blackicon-eth/fly-plays-games (commit ecb95f7; pokemon-red README, record_journey.py, encoder_b.py, play_live.py) [direct].
- dev.ua, 28 Sep 2026 [search-summary; bot check to our fetch].
- vib2810/fly-brain-zero-shot (commit 6c3a5e7; wiring.py, vision_loop.py, download_data.sh, assets/results_per_fly.csv) [direct].
- DenisSergeevitch/desktop-fly (commit 32b0001; FlyModel.swift, main.swift, Sim.swift, Locomotor.swift, etl.py, EVALUATION.md) [direct].
- Yusuftmle/FlyBrain-HalfLife (commit 0ba5548), its first video and second video [direct; video metadata].
- gabrycina/doom-fly-control (commit 6b22922; results/a2/summary.json) [direct].
- The other new repositories, each linked from its catalogue row with the files we read. Machine-readable data: projects.json and videos.json.