AGI Experiment — World Rules
Can composable causal primitives emerge from unsupervised dictionary learning? Training on single-rule physics, testing on multi-rule compositions.
The challenge
Can a model trained only on simple, single-rule physics events learn representations that generalize to novel multi-rule interactions it has never seen?
Approach
- Built a custom 5×5 grid world with typed objects governed by 5 deterministic physics rules: gravity, containment, contact, bounce, and breakage — rules compose naturally.
- Events encoded as 18-dimensional vectors from raw fields. A sparse dictionary (ISTA) is trained on single-rule events using Hebbian learning — no backpropagation.
- Compared 4 architectures: ISTA baseline (3/9 composition tests), ProductOfExperts (1–2/9), ContrastiveProductOfExperts (0–2/9), and ContrastiveDictionary (9/9).
- Evaluated via 9 composition tests measuring reconstruction quality and atom-activation overlap (Jaccard similarity) to verify true decomposition of multi-rule events.
Key takeaway
Sparsity alone is insufficient for compositional generalization. Contrastive specialization pressure — penalizing atoms for activating on multiple rules — is the mechanism by which causal structure emerges. Not a regularizer, but the core mechanism.