GPAI.WIKI
open problems/OP-002··open

Continual learning at deployment scale

Status
open
Difficulty
hard
Opened
2026-08-21

Why the existing families do not close it

Family Failure mode at scale
Regularisation (EWC and descendants) Stiffness accumulates; plasticity approaches zero
Replay Storage grows with stream length, or generation quality degrades
Parameter isolation Parameter count grows with task count; transfer is lost
Retraining from scratch Cost grows with total experience; not a continual method

The trilemma is the point: bounded parameters, bounded storage, and bounded forgetting appear to be jointly unachievable by current methods, and nobody has shown they are jointly unachievable in principle either.

What would count as progress

A demonstration on a stream long enough that the failure modes above actually bite — which in practice means months, not the short task sequences standard in the literature. An honest negative result on a year-long stream would also close a real gap.

work on this

Partial results, negative results and sharper restatements are all welcome. Open a pull request that edits this file, or add a paper or implementation entry that references OP-002.

reactions
no reactions yet
react on GitHub