Chat history + files aren’t enough to reliably restore coding agents. Wisely timed sandbox snapshots preserve the full environment agents depend on. The sandbox is increasingly part of the agent itself. More below!
https://t.co/WT9o76aU8y
ran into this problem when trying to run swebench verified back in the day. that was the first time I crashed my mbp. much easier to do in the cloud w sandboxes and composable compute.
This is excellent read about Reinforcement Learning for a few different reasons
> helped me understand TITO (Token In Token Out) and why it is important for RL
> got to see what a large scale RL training run looks like with great details, including inference to train GPU ratio. (12:8)
> mentions @daytonaio
https://t.co/fAyD0aLNwm