I present SNN-Synthesis v10, a comprehensive investigation of stochastic resonance and emergent intelligence in neural networks spanning 2. 8K-parameter Neural Cellular Automata to 7B-parameter LLMs across 100 experimental phases. Building upon v1–v9 (Phases 1–67), v10 adds 33 new experiments (Phases 68–100) that introduce Liquid Neural Cellular Automata (L-NCA) —a radically different paradigm where intelligence emerges from the collective dynamics of minimal local-rule cells rather than from massive parameterization. Five New Findings (v10) (XXVI) Liquid-LIF and Subthreshold Computing — Liquid Leaky Integrate-and-Fire neurons with learnable time constants enable zero-shot time-warping generalization to unseen temporal scales. STDP achieves competitive results without backpropagation. (Phases 68–80) (XXVII) L-NCA: Size-Free Perfect Generalization — L-NCA trained on 8×8 grids achieves 100% pixel accuracy on unseen 12×12 grids with only 2. 8K parameters (22× fewer than CNNs). Local cellular rules are inherently size-invariant. (Phases 81–86) (XXVIII) Liquid Mixture-of-Experts — 5 specialist L-NCAs with zero-shot loss-based routing achieve 100% routing accuracy and 76% solve rate, up from 0% with a single multi-task model. Compositional routing—chaining ExpertB (ExpertA (x) ) —discovers correct rule compositions for novel composite tasks (6% → 100%). (Phases 87–94) (XXIX) Attractor Regularization and Auto-T — Random-T training with L2 state-change penalty maintains 99% accuracy at T=50 (vs. 0% baseline collapse). Auto-T early stopping reduces latency by 7× while improving accuracy. (Phases 95, 99) (XXX) Color-Frequency Invariance — Frequency-based color remapping achieves 98% exact match on color-shifted tasks where the baseline scores 0%. (Phase 97) v20 Ultimate Liquid AGI (Phase 100) The grand finale integrates all innovations into a single agent achieving 88% solve rate, 100% routing accuracy, 338ms average latency (within 500ms budget), and 0% timeout rate on 40 simulated ARC levels—using only ~14K total parameters. v1–v9 Foundations (Phases 1–67) All 25 previous findings remain validated, including Noisy Beam Search, SNN-ExIt, SR-Quantization, the space-time-precision triad, noise source separation, and ARC-AGI-3 thermodynamic coarse-graining. 30 contributions spanning 2. 8K–7B parameters, NCAs to Transformers, 9 task domains, 4 model families, and 22 honest null results. Code and data: https: //github. com/hafufu-stack/SNN-Synthesis Acknowledgments This research was conducted entirely independently, without institutional affiliation or corporate funding. The author currently faces financial constraints that make it increasingly difficult to maintain subscriptions to AI services essential for this line of research. To sustain and improve the quality of future work, the author is actively seeking community sponsorship. Details are available at https: //github. com/sponsors/hafufu-stack.
Hiroto Funasaki (Fri,) studied this question.