Generating Bad Apple autonomously from a single initial state using a tiny recurrent dynamical system (417k params) [P]

A developer trained a tiny recurrent dynamical system with 417k parameters to autonomously generate the Bad Apple video from a single initial state. The training involved a Rollout Horizon Curriculum, progressively doubling the rollout length from K = 2 to K = 512. This method caused characteristic jumps in loss before the transition function adapted. The developer used AI to assist with writing the post and README, and some code, but designed the architecture independently.