Finite state automata-like feature assemblies emerge from autoregressive training on data patterns, not from co-evolved circuit design

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Finite State Automata Feature Assembliesconcept0.830
  Collections of features that interact via the token stream — one feature increases probability of tokens that activate the next feature — forming FSA-like systems
• Autoregressive language models cannot converge to single stored patterns beyond their context window from local interactions alone.claim0.774
• Autoregressive model unable to converge to a single stored pattern for any finite β (Corollary 2)finding0.757
  Consequence of Theorem 3 and 1D no-order result
• Features are connected by weights forming circuits, and these circuits can be rigorously studied and understood as meaningful algorithms.claim0.755
  Second of three speculative claims asserting that subgraphs of neural networks are tractable and meaningful objects of study
• Evolution learns to generalize beyond default morphologies, producing problem-solving machines.claim0.754
  Argues that evolutionary learning goes beyond specific adaptations.
• Keeping goal state encodings functionally orthogonal from the machinery that executes them (as in bioelectric pattern memories) is a powerful architecture exploited by evolution.claim0.753
  Argues this separation allows reprogramming without hardware change.
• Among 17 chaotic/complex cellular automata rules, 30% show causal emergence, 70% show causal degradation.finding0.752
  Varley (2020) analysis using ordinal partition network on cellular automata.
• By using a variational autoencoder-like architecture for genomic compression, evolution is freed from over-training and pushed to evolve general-purpose problem-solving machines.claim0.752
  Claim linking the indirect genotype-phenotype mapping to robustness and open-endedness.