用果蝇连接组在 Astra 上做城市交通智能体

Marko @Marko_Poly原推发布于 2026.09.16

在 X 上打开原推

I gave a fly a city. Not a metaphor - a build. I took the fruit-fly connectome (the first brain ever fully mapped: ~140,000 neurons, ~50 million synapses) and rebuilt it as an agent on GPT-6 Astra. Then I gave it one job: be the traffic conductor for a simulated New York. I called it FLY/01. Why a fly. A fly doesn't plan a route across Manhattan - it reacts, locally and instantly, thousands of times a second, on almost no energy. That's the opposite of how we build traffic AI, which tries to solve the whole grid at once. I wanted a controller that's reflex, not central planning. How it runs. FLY/01 sits above the grid as one long-running agent. Every intersection is an input, every light an output. It doesn't re-plan the city each tick — it holds a small state of what's true right now (which corridors are loaded, where the next wave is forming) and patches it. That's the only reason it survives hours without falling over. Why that's the whole point of this article. An agent that runs a city for 8 hours is exactly the thing that dies of a quadratic token bill and a poisoned context by hour two - unless you build the boring layer under it: a state file instead of a transcript, a spending ceiling, evals on its own calls. The fly solved this in biology first. It keeps state, not a log of every wingbeat. What the sim showed. 250 vehicles in motion, 0 stopped. Midtown, Central Park, downtown - every corridor in sync. One green wave. Time lost to traffic: 00.0. Less waiting, more city. It's a simulation, not the MTA. But it's a clean test of one idea: the best controller might be the smallest one that never stops paying attention. The full 8 steps for running an agent this long without burning out are in the article. Would you let a fly run your city?

这条是引用着另一条写的：x.com/marko_poly/status/2097659882275049837