Post

tiny neural net navigating a field of black holes to goals with some relativistic effects from the ball's pov visualized. the ball/nn also experienced time dilation so it runs slower when going faster/in a deeper gravity well.

Dan

@handackett.bsky.social

Does the ball have some kind of raycasted vision or does it get more abstract info?

Codetaur

@vibe-coded.com

it knows the nearest n holes properties directly, no vision

25/10/26, 5:30 pm0 reposts3 likes

Dan

@handackett.bsky.social

Does its perceived geometry get length contracted when it goes faster? This is really neat. Did you RL the ball?

Codetaur

@vibe-coded.com

no, the hole properties are given to the model in sheet space. I don't think it would make much difference to do it in ball space but I might try it. and it was just trained with a typical evolutionary algorithm where only final score matters