{"slug": "desktopfly-gives-macos-a-fruit-fly-driven-by-668-mapped-neurons", "title": "DesktopFly gives macOS a fruit fly driven by 668 mapped neurons", "summary": "Denis Shiryaev released DesktopFly on August 18th, an open-source macOS toy that simulates a 3D fruit fly on application windows using a 668-neuron circuit with about 19,000 synaptic connections extracted from the FlyWire consortium's FAFB v783 dataset. The fly reacts to cursor movement, grooms, sleeps, and can take flight when approached quickly, with the simulation running at 1 kHz and including an interactive brain viewer. Shiryaev, a group product manager at JetBrains and co-founder of neural.love, built the tool to apply agent-assisted development to a scientific dataset, but the project does not simulate the full 139,255-neuron brain.", "body_md": "[Denis Shiryaev (@literallydenis)](https://x.com/literallydenis) released [DesktopFly](https://github.com/DenisSergeevitch/desktop-fly) on August 18th, an open-source macOS toy that puts a 3D fruit fly on top of application windows and uses a simulated slice of the FlyWire brain map to decide how it moves.\n\nThe fly walks along window edges, grooms, sleeps and reacts to the cursor. A fast approach can send it into flight, while a slow movement may leave it undisturbed. DesktopFly also includes an interactive brain viewer that lets users stimulate regions of the simulated circuit and watch the resulting behavior.\n\nShiryaev is a group product manager working on JetBrains' AI Assistant and a co-founder of generative AI service neural.love, according to his X profile. His public GitHub work has recently concentrated on coding agents, repository workflows and developer instructions for AI systems. DesktopFly applies that agent-assisted development approach to a scientific dataset rather than another coding utility. ([x.com](https://x.com/literallydenis/with_replies?utm_source=openai))\n\n### A real wiring diagram, with modeled behavior\n\nThe word \"real\" needs a boundary here. DesktopFly does not simulate all 139,255 neurons in the female adult fruit fly brain mapped by the [FlyWire consortium](https://home.flywire.ai/). Shiryaev extracted a 668-neuron circuit containing about 19,000 synaptic connections from FlyWire's FAFB v783 dataset. The circuit covers populations associated with looming detection, escape, steering, walking, grooming and backward movement. ([github.com](https://github.com/DenisSergeevitch/desktop-fly))\n\nAnother 23,210 neuron positions appear in the project's rotating brain visualization. Those points provide anatomical context; they are not all participating in the live simulation. The full FlyWire resource contains more than 50 million synapses across roughly 140,000 proofread neurons, the product of work by hundreds of scientists and proofreaders. ([home.flywire.ai](https://home.flywire.ai/))\n\nDesktopFly runs its smaller circuit as a 1 kHz leaky-integrate-and-fire simulation. Cursor movement is converted into input for LC4 and LPLC2 looming-detector cells. Activity then passes through connections taken from FlyWire toward the two DNp01 giant-fiber neurons associated with escape. When the simulated giant fiber spikes, the rendered insect takes off.\n\nShiryaev says roughly 1,200 inhibitory synapses in the selected circuit suppress gradual approaches, while abrupt looming input can produce a giant-fiber spike in about four milliseconds. That timing is an output of DesktopFly's implementation and diagnostic tests, rather than an independent biological measurement of the application. The repository includes simulation tests and 17 end-to-end checks connecting neuron stimulation to actions such as grooming, turning and escape. ([github.com](https://github.com/DenisSergeevitch/desktop-fly))\n\nThe repository is unusually explicit about where the connectome ends and Shiryaev's engineering begins. FlyWire supplies neuron identities, positions, connections and synapse counts. DesktopFly supplies the leaky-integrate-and-fire dynamics, baseline activity, synaptic delays, electrical-coupling boosts and the conversion from mouse movement into sensory input. The fly's 3D body, gait, wings and sleep postures are procedural because FlyWire maps the brain rather than the animal's body. ([github.com](https://github.com/DenisSergeevitch/desktop-fly))\n\nThat separation keeps DesktopFly from making the larger claim that it has reproduced a fruit fly mind. It is an interactive model built on measured anatomy, with substantial behavior logic and physiology layered on top.\n\n### A desktop pest with system APIs\n\nDesktopFly uses macOS window geometry as terrain. The insect can land on the upper edge of a window, walk with it when the window moves and react when its perch disappears. New windows become looming objects, nearby clicks become substrate taps and rapid cursor motion acts as wind input.\n\nShiryaev also tied the simulation to time of day and the Mac's thermal state. The fly has morning and evening activity peaks, a midday slowdown and nighttime sleep behavior. A hotter computer produces a faster fly, following the broad premise that flies are ectothermic. These environmental rules are application features rather than outputs derived from the FlyWire graph.\n\nDesktopFly requires macOS 13 or later and Swift 5.9. Installation currently involves cloning the repository and compiling it with the included build script. The resulting application runs as a transparent, click-through overlay and adds a menu-bar control for pausing the fly, opening the brain viewer and injecting a test looming stimulus. The project says it does not require accessibility or input-monitoring permissions.\n\nThe code is available under the MIT license. The derived FlyWire data files carry a separate CC BY-NC 4.0 license, preserving the research dataset's noncommercial terms.\n\n### Claude gets a co-author credit\n\nDesktopFly's initial commit credits [Claude Fable 5](https://www.anthropic.com/claude/fable) as a co-author, and the repository includes a detailed `CLAUDE.md`\n\nfile documenting the architecture, test requirements and previous simulation failures. Anthropic positions Fable 5 as a model for long-running coding work that can plan, delegate and test its output. The repository does not separate which portions were written by Shiryaev from those generated by the model. ([anthropic.com](https://www.anthropic.com/claude/fable?utm_source=openai))\n\nThe result is a compact demonstration of two resources that have recently become easier for individual developers to use: a research-grade connectome published as accessible data and a coding agent capable of assembling the surrounding simulation, interface and tests. The fly remains a desktop toy. Its underlying ingredients are considerably less trivial.", "url": "https://wpnews.pro/news/desktopfly-gives-macos-a-fruit-fly-driven-by-668-mapped-neurons", "canonical_source": "https://runtimewire.com/article/desktopfly-macos-fruit-fly-flywire-connectome", "published_at": "2026-08-18 22:59:32+00:00", "updated_at": "2026-08-18 23:13:34.795663+00:00", "lang": "en", "topics": ["artificial-intelligence", "neural-networks", "ai-research", "ai-tools"], "entities": ["Denis Shiryaev", "DesktopFly", "FlyWire", "JetBrains", "neural.love", "FAFB v783"], "alternates": {"html": "https://wpnews.pro/news/desktopfly-gives-macos-a-fruit-fly-driven-by-668-mapped-neurons", "markdown": "https://wpnews.pro/news/desktopfly-gives-macos-a-fruit-fly-driven-by-668-mapped-neurons.md", "text": "https://wpnews.pro/news/desktopfly-gives-macos-a-fruit-fly-driven-by-668-mapped-neurons.txt", "jsonld": "https://wpnews.pro/news/desktopfly-gives-macos-a-fruit-fly-driven-by-668-mapped-neurons.jsonld"}}