Hotdog.
A bun. A sausage. A moment of clarity.
Big ambition. A fly-inspired neural network.
An entirely unnecessary way to find a hotdog.

One questionable career choice.
All thinking about lunch.
The fly has the wheel.
We like to keep things focused.
Some apps try to do everything.
We have a much more specific vision.
The fly explores. The verdict follows.
A bun. A sausage. A moment of clarity.
Pizza. Burgers. Donuts. Tomatoes. We checked the scope.
Season four of HBO's Silicon Valley, 2017.
A food app with one opinion.
We are the sequel nobody greenlit.
Jian-Yang has an octopus-recipe app. In the room, Erlich rebrands it on the spot as a camera that names any dish, calls it SeeFood, and leaves with $200,000 for software that does not exist.
Point it at a hotdog and it says hotdog. Point it at pizza, a shoe, or a co-worker and it says not hotdog. Nobody could be bothered to train it on every food in existence.
A working Not Hotdog app went up on the App Store the night the episode aired, built by an engineer on the show with a genuine deep-learning classifier. Reviewers failed to fool it with pistachios.
Same one-bit question, no camera. A neural network wired like a real fruit-fly connectome walks a picnic table, smells its way to hotdogs, and pays for every pizza. Pizza is very much on the menu.
Built on a real fruit-fly connectome, our trained neural network turns sensory signals into movement. It decides where to go, what to approach, and which way to turn.
You get the best job: watch it figure things out. Open the demo to see its neural activity, antenna signals, and decisions as they happen.
Meet your new favorite overthinkerNo downloads or account. The game and its brain load right in your browser.
It senses food, follows its neural policy, and rebounds off the table edges.
Hotdogs earn points. Everything else costs them. Pick a camera and enjoy the show.
We've anticipated the board's concerns.
A trained neural network built on real fly-brain connections controls the game. It’s a computational model, not a living fly. Every forward move and turn comes from that model’s policy.
You’re here to watch the demo. You can change the camera, pause, restart, adjust the pace, and enable sound. Steering is entirely the fly’s department.
Because of Silicon Valley. In season four, Jian-Yang's SeeFood app is pitched as "Shazam for food" and turns out to recognise exactly one dish: hotdog, or not hotdog. Two food groups felt like a manageable business plan. Then we added an autonomous fly, a 3D picnic, and a frankly unnecessary amount of neuroscience.
No. That was a real phone app HBO released the night the episode aired, built by an engineer who worked with the show, and it classified photos. This is an unaffiliated fan project with the same one-bit ambition and a different brain. Nothing here is endorsed by HBO.
No installation, account, or API key. The game and its exported neural network run locally in your browser after loading. A browser with WebGL support is required. Sound is optional.
The fly’s internal neural state changes as it senses and moves, but its trained weights stay fixed during the demo. Each new round resets the arena and the network’s memory.
One fly. One picnic. No human steering.