PULSELoCo: 17x less trainer-to-trainer bandwidth in distributed RL post-training | Hacker News Reader