A fly brain running a drag racing simulation for 24 hours a day, seven days a week, is exactly the kind of project that makes you stop and reconsider what we mean by intelligence. In Season 11, a second-generation fly model won the latest Grand Prix by 0.02 seconds. That is not a metaphor. That is a real measurement, produced by a 2,400-neuron simulation learning to manage acceleration, grip, and wet-surface conditions through hundreds of thousands of training episodes. The researcher behind FlyPilot has turned this into a live-streamed racing league, complete with Elo ratings and generational dynasties, and it is both fascinating and slightly unsettling to watch.
This project sits at an interesting intersection with recent work in the field. We have seen New AI Research Tools Raise Questions About User Privacy and Consent tackle the ethical boundaries of research agents, while Uncovering Hidden Behaviors Learned During AI Post-Training explores what models pick up after their initial training is complete. FlyPilot offers a tangible, observable example of that same principle in action: these neural models are not just executing instructions. They are discovering emergent strategies. The original creator noted that the flies developed their own form of launch control without any explicit programming for it. That is not a bug. It is the point.
What makes this worth paying attention to is not the novelty of a fly brain playing a racing game. It is what the training setup reveals about the nature of learning itself. The system uses 10 million episodes, distributed across generations where the best performers pass their weights to offspring. The champion fly finished five scenarios in 56.31 seconds total. That is a concrete, measurable outcome from a biological simulation that is far simpler than most large language models. The comparison is instructive. Where large models often obscure their reasoning behind billions of parameters, a 2,400-neuron network is small enough to inspect, evolve, and understand. This is the kind of experimentation that can teach us how small, focused models can develop surprisingly sophisticated behaviors when given the right training regime and enough time.
The practical takeaway here is direct: small models trained with evolutionary strategies can produce behaviors that look emergent, even strategic, without requiring massive scale. For anyone building or using AI tools, this suggests that more of our attention should go toward training design and scenario diversity rather than simply scaling up parameter counts. The FlyPilot project is open source, the code is on GitHub, and the streams are live. You can watch a second-generation fly brain make a 0.02-second decision that secures a championship. That is not a promise of future capability. It is a demonstration happening right now, and it raises a specific question worth tracking: how many other emergent behaviors are waiting to be discovered in models we already have, if we just give them the right conditions to race?