Daanish Khazi @bertgodel

interesting tidbit post cursor acq: production data is likely useful for mid-training but tasks are too underspecified and environments too non-stationary for RL bearish for rl-based "continual learning", bullish for synthetic envs & tasks