Async algorithm speeds reinforcement learning training without quality loss

Original: Running RL asynchronously is the key to faster and cheaper training runs. We've been doing A LOT of research here to make the most performant RL training stack for open weight models.

Source: x.com ↗