Definition
Plain language
A periodic deeper revision pass that captures long-term lessons.
As stated in the literature
In SkillOpt, an epoch-boundary consolidation operation comparing this epoch's skill to last epoch's across replayed training tasks and writing persistent lessons into a protected region of the skill document.
Why it matters: Frequent small edits handle local mistakes, but a periodic deeper consolidation is what captures the broader patterns that wouldn't fit in any single update.
For example, at the end of each training epoch, the system reviews this epoch's skill document against last epoch's across many tasks and writes down lasting lessons in a protected section.
Heard on the show
“SkillOpt has an epoch boundary, and at the boundary it does what they call a slow update.”Episode 078 — Training a Markdown File: When LLM Self-Improvement Borrows the Discipline of Neural Net Training