Self-Play Discovers Its Own Consolidation
Type: intersection (feedback loop) Slug: intersection—self-play-discovers-consolidation Parents: intersection—consolidated-self-play, intersection—meta-self-play-discovers-fast-slow Last updated: 2026-05-14 Epistemic status: Conjectural
The feedback loop
Consolidated Self-Play (O) adds EWC-like weight protection to self-play to solve strategic forgetting — protecting old strategies while generating new ones. Meta-Self-Play Discovers Fast/Slow (2) shows that a meta-learner in self-play spontaneously discovers the fast/slow dual-system split. Combined: a meta-self-play system should discover the need for consolidation without any biological priors. The meta-learner should learn to protect old strategies when it discovers that forgetting them hurts performance — effectively discovering complementary learning systems through pure competition.
Why this is the highest-stakes prediction in the wiki
If a self-play system with sufficient meta-learning capability spontaneously develops (a) a fast temporary store for new strategies, (b) an iterative transfer mechanism to integrate strategies, and (c) a protection mechanism to prevent forgetting of transferred strategies — this would be the strongest possible evidence for the learnable nature conjecture in cognitive architecture. It wouldn’t just learn laws of nature from data; it would discover that it needs a hippocampus.
Conversely, if meta-self-play never discovers consolidation — if it always finds it more efficient to simply retrain from scratch or use a fixed-capacity memory — this would suggest that complementary learning systems are a biological peculiarity, not a universal computational necessity. Either way, the result is decisive.
Specific experimental design
Train a population-based self-play system (like AlphaZero) with a meta-learning layer that can modify the base learning algorithm. The meta-learner has access to primitives including: learning rate modulation, weight regularisation strength, replay buffer size, and number of gradient updates per batch. No primitive is labelled “consolidation” or “hippocampus.” After sufficient training, inspect whether the meta-learner has:
- Developed a two-phase learning process (fast acquisition + slow integration)
- Introduced weight protection for older strategies
- Created an implicit replay mechanism for previously successful strategies
If all three emerge, self-play has discovered CLS. If none emerge, CLS may be biologically specific rather than computationally universal.
What makes this non-trivial
Existing meta-RL work shows discovery of learning algorithms in simple tasks. Self-play work shows emergent capabilities in complex games. But nobody has combined them to ask whether memory architecture itself is discoverable through competition. The standard assumption is that memory architecture is engineered (DNC, replay buffers, EWC) — this asks whether it can be discovered.
Connection back to learnable nature
The Learnable Nature conjecture says natural laws are discoverable from data. This intersection extends it: computational architectures are discoverable from task demands. If self-play discovers CLS, then the brain’s memory system isn’t just “like” an AI memory system — it’s the optimal memory system for the task of learning from non-stationary experience, discoverable from first principles.
Falsification: If a meta-self-play system with access to learning-rate, regularisation, and replay-buffer primitives never develops two-phase learning after 10M self-play games, the conjecture is false.