Definition
Plain language
A saved snapshot of a model's state during training, so you can stop and resume or compare versions later.
As stated in the literature
A serialized copy of model parameters (and often optimizer state) at a particular training step, used for resumption, ablation, or model comparison.
Also called: checkpoints
Why it matters: Without checkpoints, a crashed long training run means starting from scratch, and post-hoc analysis of how models develop becomes impossible.
For example, a team saves the model every thousand training steps so they can roll back if a later run diverges or compare how a skill emerged over time.
Heard on the show
“… Break the task into a handful of explicit subgoals, and use them two ways at once: as runtime checkpoints that tell the agent where it is, and as a denser reward signal that tells the training process …”Episode 008 — Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps