The Sequence Knowledge - Issue #959: Learning RSI: How Models Learn to Direct Their Own Updates
SEAL and the architecture of persistent adaptation
TL;DR
- Models can exhibit in-context learning but may not retain knowledge after a session ends.
- Persistent learning requires a decision on how experience should alter the model.
- SEAL allows models to generate and apply self-edits, leading to improved question-answering accuracy.
- The article distinguishes between remembering and learning, and discusses computational choices in training examples.
- Nested Learning and Titans explore adaptation across different timescales.
- Catastrophic forgetting is a key problem for persistent learners.