It's a race: every car starts together and the first to finish 3 full laps wins. There's no speed limit β cars accelerate as hard as they dare.
The track changes every generation, so a driver can't just memorise one circuit β it has to learn to drive any track. That's what lets a driver you train work when you share it to another device.
Generalist = how the all-time best driver does on a fixed set of benchmark tracks it never trained on β higher is better. While it's still learning to finish, this shows average laps (out of 3). Once it finishes every benchmark track, laps would just sit at 3, so it switches to a 0β100 performance score based on how fast it finishes them β that keeps climbing as the driver gets quicker. from gen is the generation that produced that record holder (β Watch best Β· Share). This is the driver that gets saved and shared, so what you share is your best all-round driver, not a one-track fluke.
This gen = best of the current generation on today's track; Leader now = laps done by the car you're watching.
The skill: with no speed cap, flooring it is easy β the hard part is not flying off the track in the corners. The network has to learn when to brake and how to carry speed through bends on tracks it has never seen.
A car dies if it leaves the track or stops making progress. The generation ends as soon as the leader finishes the race (the rest are ranked by how far they got) β so training speeds up once a winner emerges. Each generation also adds fresh-random and bold-mutated cars, so you'll always see variety.
No driving β and no training data. Each car has a tiny neural network; a genetic algorithm breeds the best every generation. Watch them go from crashing instantly to mastering the track.