Cursor | The Hidden Bug in Every Large-Scale RL Run

Federico Cassano explains a numerical mismatch problem in async RL that plagues large sparse MoE models like Kimi. He teases that the next Composer will be trained on Cursor's own base model.