# Article source guide Checked 12 September 2026. These notes support your historical and critical sections; the experiment's separate methods record reports its current results. ## Distinguish the datasets before telling the story The 2024 FlyWire publication describes an adult **female brain**, containing 139,255 neurons and 54.5 million synapses. The accompanying annotation paper describes about 15.1 million weighted neuron-to-neuron edges. These numbers measure different things: several synaptic contacts can connect one ordered pair of neurons. [Wiring paper](https://www.nature.com/articles/s41586-024-07558-y), [annotation paper](https://www.nature.com/articles/s41586-024-07686-5). MaleCNS is a separate male central-nervous-system reconstruction, including the brain and ventral nerve cord. The official release notes date v1.0 to June 8, 2026. The homepage lists the paper's publication on September 3, 2026. The homepage dates v0.9 to October 3, 2025, whereas the release notes say October 5; use “October 2025” unless that discrepancy is resolved. Do not present the June 2026 version as the initial 2024 FlyWire release. [Project page](https://male-cns.janelia.org/), [release notes](https://male-cns.janelia.org/release/). For our actual prepared model, quote the verified local counts: 166,700 nodes, 25,582,938 directed edges and 124,177,617 represented contacts. State that these are counts in the retained prepared graph, so readers need not assume that any different count in a publication indicates an error. ## Two entertaining demos, inspected at the source **fly-wirehead** feeds real video pixels into the neural model. However, its feed advances on a timer and its front-leg swipe is choreographed. Each accepted video frame also supplies an artificial current to 15 PAM11 dopamine neurons. The README explicitly says learned preference, pleasure and addiction have not been established. That makes it an excellent example of the distinction between real calculated activity and an unsupported interpretation of that activity. Its overlay reports firing rates, not dopamine concentration. [Repository and explanations](https://github.com/mattyhempstead/fly-wirehead). **Stonkfly** turns public market prices into an RGB chart. A fixed neural readout proposes buy, sell or hold. Positive portfolio change stimulates PAM11 cells; negative change stimulates PPL101 cells. A candidate rule changes existing KC-to-MBON connections, but the project explicitly does not claim demonstrated profitable learning. In your article, distinguish the financial feedback supplied by the program from the network's response to the chart. Neither is evidence of subjective financial anxiety or satisfaction. [Repository and model notes](https://github.com/nftechie/stonkfly). Descriptions should be pinned to inspected revisions when the article is published. Our reused fly-wirehead source is revision `fcefe9441f80e25aab713411ebced53f5e5ea172`. A current README can change after a viral video was recorded; avoid assuming that it describes every older clip. ## More examples worth separating carefully **DOOMFLY** connects actual game frames to the retained graph and uses a fixed neuron-to-button interface. Damage supplies an artificial aversive input, and a candidate rule changes existing synapses. Its current README reports failed visual, conditioning and survival validation gates. A running Doom player is therefore not the same claim as demonstrated improvement through learning. [Repository](https://github.com/nftechie/doomfly). **Flyhard** reports a trained requested-angle steering skill through simulated foreleg contact with a wheel. Its README reports a held-out stationary-wheel test and a connected CARLA sequence, but explicitly says visual driving remains untested. The current connected sequence uses requested turns and scripted speed. The indicator-stalk experiment is described as a planned, untested fallback. Do not report it as an already demonstrated ability to choose when to signal in traffic. These are the repository's reported results, not tests we independently reproduced. [Repository](https://github.com/MarkUnthank/flyhard). **Eon's embodied fly** is particularly relevant to the “upload” debate. Its March 10, 2026 technical account describes an integration of existing neural and biomechanical components, including modified NeuroMechFly walking controllers. It describes invisible taste cues and fictive dust in the virtual environment, and acknowledges simplified neural dynamics and a limited set of inputs and behaviors. The authors defend their use of “upload” as a graded term. Attribute that definition to them; do not silently treat it as a demonstrated identity between the simulation and the original animal. The technical account is more informative for your criticism than a reposted headline. [Authors' explanation](https://eon.systems/updates/embodied-brain-emulation). ## A useful counterweight to “this is pointless” The amusing application can be arbitrary while the scientific approach remains useful. A connectome provides anatomical constraints for hypotheses that can be checked against living animals. Shiu and colleagues' 2024 computational-model paper is a relevant scientific starting point, distinct from the social-media demonstrations. [Primary paper](https://www.nature.com/articles/s41586-024-07763-9). There are also explicitly engineered uses of connectomic architecture. The Fly-connectomic Graph Model preprint reports training a graph-structured controller for simulated fly locomotion using deep reinforcement learning. It was first submitted February 20, 2026, with the inspected third version dated June 14. Treat its reported performance as the authors' research claim and identify it as a preprint, rather than proof of an uploaded animal. [Preprint](https://arxiv.org/abs/2602.17997). Our interpretation: the sharp criticism should target unjustified claims about what a demonstration establishes. “A graph constrained by anatomy can control a simulated task” and “the original fly now inhabits a computer” are profoundly different propositions. The former does not supply evidence for the latter. Likewise, success in a task can validate an engineering system without validating all of its assumptions about biology. ## Questions to ask of every spectacular clip 1. Which cells and connections are actually simulated? 2. What information enters the model, and through which adapter? 3. Is reward calculated from an external label, a game score, elapsed playback, or the model's own subsequent behavior? 4. Which parameters change during training? 5. What maps neural outputs to actions? Are visible movements authored, mechanically simulated, or a mixture? 6. Does the demonstrated behavior improve with training when reward is removed? 7. Are there original-weight, unrelated-reward and held-out controls? 8. Which claims concern measured behavior, and which are metaphors? These questions also apply to our demonstration. Its current failed learning tests belong in the laboratory record. The successful body-physics repair and the functioning pixel–neuron–head loop do not turn those failures into successful learning. ## Wording for the personal-experiment section The completed narrow demonstration supports “I trained a connectome-constrained model to turn toward an intact Habr article rather than its scrambled control.” The successful model uses differentiable rate equations and supervised optimization of existing weight factors; it is not the earlier spiking plasticity model. `EXPERIMENT_AND_RESULTS.md` explains that change and the reserved tests. Do not imply that the as-yet unwritten article itself was used in training: the current captures come from a different, identified Habr article. The playful title can still use “likes” or “enjoys,” provided the text explains the operational definition and does not substitute it for subjective experience. Avoid claiming that ordinary website screenshots demonstrate reading comprehension or semantic understanding.