Data◆Audio · 14:37 · 2m
Reinforcement-learning rollout workloads are so extremely bursty that they can require spinning up on the order of a hundred thousand sandboxes at once, far more bursty than typical agent workloads.
Akshat Bubna · Latent Space