The details
Today’s models are mostly frozen once deployed. Updating them with new data tends to cause “catastrophic forgetting”, where new learning wipes out old skills, while anything picked up in a conversation is lost when the session ends. During sleep, the model adds new capacity rather than overwriting existing parameters, then moves short-term memories into it. In the dreaming stage it mixes knowledge across domains and keeps only the most useful practice examples. In translation tests on unseen languages such as Manchu and Kalamang, it showed almost no drop over several stages, while standard methods forgot heavily.
Why it matters
A model that keeps learning after release could stay current without costly retraining and build new skills from use. The idea borrows from how human sleep consolidates memory. These are research results on the authors’ chosen benchmarks, and the post doesn’t say what sleep costs in computing power or when it might reach products.