Multiple laboratories5 min readSample
Continual learning results fail to replicate at scale
Methods reported to avoid catastrophic forgetting in smaller models are found not to hold at frontier scale.
highAGI relevance: highmodelsreasoning
Agitology analysis
Why it matters
Learning is the dimension where progress most resembles engineering around a limitation rather than removing it. A negative replication is informative precisely because it is rarely published.
Key developments
- Forgetting reappears at frontier scale under the same methods
- Scaffolding-based approaches continue to work and continue not to be learning
- Authors describe the underlying problem as open
Index impact
How this development moved — or failed to move — the dimensions it bears on.
- −0.5Learning