A fly plays Pokémon, Part 1

Thu, 17 Sep 2026 00:00:00 GMT

So, a fly neural map just dropped, and the internet has been getting up to some delightful, borderline horrifying shenanigans. A team led by HHMI Janelia, working with Google Research and collaborators, mapped a male fruit fly’s brain and nerve cord. Thin slices, electron microscope images, a whole lotta analysis, and the dataset is now on the internet.

Some of the reconstructed cells in the male fly's brain and nerve cord, viewed from two angles. Image: Google Research, from the connectome work by HHMI Janelia, the Cambridge Connectomics Group, Google Research, and collaborators.

I was introduced to this madness with Infinite Sugar, where William Kelly (cnqso) uses the earlier FlyWire connectome, published in 2024, to simulate a fly’s brain. Some neural activity drives movement in a simulated body, and a button “feeds” the fly infinite sugar by triggering sweet-sensing neurons. Most motor output is disconnected, and the fly doesn’t get visual input from its terrarium.

Infinite Sugar, with the sugar stimulus on and neural activity shown at lower left. Screenshot of William Kelly's Infinite Sugar; the project uses FlyWire connectome data and the FlyBody model.

I was on the road when I read about this and began doom-coding (prompting an LLM on my phone instead of binging social media) a prototype. It started with questions about the dataset Infinite Sugar used, and the conversation meandered into project territory.

I wanted to do something with this. People were already hooking these simulations up to games, including Alex Wormuth’s Doom project, which maps game frames to sensory input and neural activity to controls. There’s also a Minecraft fly mod that uses the newer MaleCNS connectome, with simplified integrate-and-fire neurons driving a fly mob.

The first thing that came to mind was watching SethBling’s MarI/O, which evolves neural networks to play Super Mario World. Then I remembered the various Twitch Plays Pokémon streams, specifically Fish Plays Pokémon.

So, fly plays Pokémon. But how? What would this look like?

A Twitch stream, obviously, an emulator, and mapping the fly to the game. Initially, the plan was to map motor neurons to buttons, the screen to visual neurons, and set up some kind of reward-based feedback mechanism. Additionally, some tracking mechanism to prevent regressions and take periodic snapshots of the game.

My stream: Pokémon Red on the left, sensory input and neural activity on the right. The progress tracker is working on the route from Viridian Forest to Pewter City. Cropped still from aflyplayspokemon by Alex Camilo. Pokémon game imagery: Nintendo / Creatures / GAME FREAK.

The tracking/checkpoint code analyzes the save file to determine how much progress was made and gives the fly sugar at milestones.

Additionally, I set up a macro system to connect some neurons to pre-canned actions that are state-dependent, again by watching emulator memory. So, for example, in battle there are neurons hooked up to things like “use an item,” “make an attack,” or “throw a Poké Ball.” I know this is a little against the “fish plays” vibe I was going for, but I think it turns the whole thing into a low-key idle RPG by enabling more focused actions.

I’m still working on it. In my haste, I built it around the wrong connectome: the 2024 female FlyWire brain used by Infinite Sugar, rather than the newer male brain and nerve cord map. I don’t expect swapping it to be too hard, but I haven’t done it yet.

I’ll do a more detailed write-up and a project page at a later point, but if you want to check it out, the stream can be found here.

Rebuilt the site

Thu, 17 Sep 2026 00:00:00 GMT

I rebuilt the site to give my projects a better home and add a log for whatever I’m currently building, mostly hardware with some software in the mix.

The design takes its cues from KiCad: drawing-sheet borders, title blocks, figure numbers, and a project index. The text pages use a light paper background. Project pages take their color from the board’s soldermask, with photos and videos showing the hardware.

It’s built with Astro and generates static HTML. A little JavaScript handles the project previews, color transitions, and photo viewer. Posts and project write-ups are still plain Markdown files with YAML front matter.

Most of the code was written with AI coding assistants.

The deployment workflow is set up for GitHub Actions to build the site and publish it to GitHub Pages. Once connected, publishing a post will be a Markdown file, a commit, and a push to main.