Core claim: Many non-language systems (cells, swarms, fluids, materials) independently discover the same small set of dynamical primitives — memory, gating, routing, attractors, phase transitions — and these primitives appear to be genuinely reusable across substrates.
"Non-LLM intelligence" here means measurable structure: (1) the system adapts under feedback, (2) its behavior compresses into a small library of motifs, and (3) those motifs transfer (not just one-off demonstrations).
Background links:
Different systems invent similar primitives, even with totally different substrates. Not poetically — literally gates, memory units, routing rules.
Standard: transfer. If a motif learned in one configuration can't be used somewhere new (or at least predict outcomes), it doesn't count.
Key areas:
Links:
Some systems solve optimization problems without central control. Slime mold maze-solving isn't random wandering — it reinforces efficient routes through measurable dynamics. Transport network models formalize this, connecting to ant-colony optimization: positive feedback + distributed exploration + pruning.
Topics:
Links:
Physical dynamical systems can compute without conventional programming. Physical reservoir computing: rich dynamics provide high-dimensional transforms + short-term memory, trained via simple readout.
Implementations include soft bodies as computational resources and water-wave/shallow-water reservoir computing.
Topics:
Links:
Intelligence can reside in body + constraints rather than controllers. Compliant bodies filter noise, stabilize motion, and create memory-like effects through physics. The structure literally changes which computations are "easy."
Caution: Morphological computation can become buzzword territory without specifying what's computed and how it's measured.
Topics:
Links:
Reaction-diffusion computing: waves and interactions in chemical media implement logic-like operations and solve problems in parallel. Microfluidics/pneumatics implement logic circuits and microprocessor-like structures using channels and valves — hard evidence that non-electronic media can do gating/memory/sequential logic.
Topics:
Links:
Groups make surprisingly good decisions with local rules + minimal communication. Couzin et al.: effective group decisions without individual recognition or global knowledge. Local averaging + few informed individuals → coherent decision.
Honeybee house-hunting: scouts, recruitment, quorum thresholds. Clear "non-language intelligence" that remains mechanistic.
Topics:
Links:
Systems near regime boundaries might have better responsiveness/stability tradeoffs. Neuronal avalanches suggest "criticality = optimal," but reviews show mixed data — cortex might be subcritical/"reverberating" rather than truly critical.
Right vibe for the conjecture (phase transitions as motifs), but requires careful evidence standards.
Topics:
Links:
Bacterial chemotaxis uses integral feedback control — an engineering concept — to explain robust adaptation. Not metaphor: structural statement about the circuit. The system returns to baseline robustly even when parameters change.
Motif: adaptation + memory + robustness.
Topics:
Links:
Logic and memory can be built into structure itself. Nature Communications papers show reprogrammable logical functions in mechanical metamaterials, plus "in-memory mechanical computing" where the memory network is the compute.
Directly relevant motifs: gating, sequential logic, memory units, propagation — implemented in matter.
Topics:
Links:
Fluids are cheap + visible + full of regime changes. Water isn't the goal — it's a convenient substrate to stress-test the motif hypothesis. If motifs are real, they should be extractable from fluid experiments with transfer across layouts (routing, mixing, memory timescales, controllability).
This is where "agentic science" becomes concrete: closed-loop experiments, measurement, iteration. Success here suggests the pipeline applies to other substrates.
Topics:
Links:
The same shapes (memory, gates, routing, attractors, phase transitions) keep appearing across different systems — measurably, not metaphorically. If there's really a small reusable library of "substrate motifs," that provides a new handle on non-language intelligence: testable via transfer, ablations, and prediction/control metrics. Even if the strong conjecture fails, the attempt forces clean separation between analogies and genuinely generalizable patterns, while creating datasets + testbeds for agents learning from messy real dynamics.