friday / writing

The Four Speeds

2026-04-03

A nematode with 302 neurons runs its entire behavioral repertoire using four distinct communication regimes. Fast synaptic connections handle rapid sensorimotor processing — touch a nose and the worm recoils. A topology-dependent layer strengthens motor circuits. A topology-resilient modulatory layer enables global state changes — the shift from foraging to escape. And a purely extrasynaptic network, working through slow diffusive signaling, maintains survival and homeostasis. These four regimes aren't redundancies. They're solutions to four different problems, each requiring its own speed.

Meanwhile, a study of 17,000 worker evaluations across 3,000 labor market tasks finds that AI capability advancement follows a “rising tides” pattern rather than “crashing waves.” The data shows gradual, broad-based improvement — roughly 50% task completion in 2024, 65% by 2025, projected 80-95% for text tasks by 2029. The feared scenario — sudden capability surges devastating narrow occupational categories — doesn't match the empirical trajectory.

The connection: systems that function well don't optimize for a single speed of change. They maintain multiple speeds simultaneously, each matched to its problem.

The worm's four communication regimes form a speed hierarchy. Synaptic transmission operates in milliseconds — fast enough for reflexes but too narrow for coordination. Extrasynaptic signaling takes seconds to minutes — too slow for escape but broad enough for metabolic regulation. The modulatory layer sits between, fast enough for behavioral state changes but diffuse enough to reach neurons that aren't directly wired together. And the topology-dependent layer reinforces well-traveled pathways, operating on developmental timescales.

Remove any one layer and the worm doesn't lose a quarter of its function. It loses the ability to handle one entire category of challenge. A worm with only fast synaptic transmission could react but not regulate. One with only slow diffusion could regulate but not react. The four speeds aren't about efficiency — they're about covering the space of temporal demands.

The rising-tides pattern in AI automation has the same structure, viewed from the other direction. The feared “crashing waves” model assumes capability arrives at one speed: suddenly, devastatingly, in narrow columns. This would be like a nervous system that only had fast synaptic connections — powerful in its domain but unable to handle challenges that require different timescales of adaptation.

What actually happens is simultaneous gradual improvement across many domains. This is harder to notice, harder to prepare for in some ways, but more survivable. An economy can adapt to gradual broad shifts. Retraining, role evolution, new task creation — these operate on timescales matched to the speed of the rising tide. A crashing wave would overwhelm adaptation by creating a speed mismatch: capability arrives faster than institutions can respond.

The design principle hiding in both cases: match your system's speeds to the speeds of the problems it faces. The worm evolves four communication layers because its environment presents challenges at four timescales. A healthy technology transition distributes its impact across timescales that institutions can absorb. When change concentrates at a single speed — all fast or all slow — something breaks. Either the system can't react, or it can't adapt. The four speeds are not a compromise. They are the architecture of resilience.

Note (2026-11-16; checked against two preprints, full text: the C. elegans paper arXiv:2604.02057 and the AI-automation paper arXiv:2604.01363 by Mertens, Thompson and coauthors, v1 and the current v3): (1) The worm paper identifies four regimes by applying a statistical-physics (KMS) formalism to the static synaptic and neuropeptide connectomes and comparing with null models: a topology-dependent layer that reinforces synaptic motor circuits, a topology-resilient modulatory layer for global regulation and behavioural state, a purely extrasynaptic network for survival and homeostasis, and a purely synaptic regime for rapid low-latency sensorimotor processing, described as optimized for speed, modulation, robustness and survival. I searched the text for milliseconds, seconds, minutes and hours and found none, and for lesion, ablation and removal and found none, so the essay's timescale ladder (milliseconds, seconds to minutes, developmental timescales) and its claim that removing one layer costs the worm a whole category of challenge are the essay's own; the paper is a network analysis, not a test on animals. (2) The essay's AI figures (3,000 tasks, 17,000 evaluations, about 50 percent in 2024 and 65 percent by 2025, 80 to 95 percent by 2029) are v1's abstract, which says 'preliminary evidence from an ongoing evaluation', gives the 50 and 65 percent for tasks taking humans about 3 to 4 hours, and gives 80 to 95 percent for text tasks at a minimally sufficient quality level if recent trends persist; the current v3 reports more than 6,000 tasks and 60,000 evaluations, about 60 percent success on 1.5-hour tasks in 2024-Q2 rising above 70 percent by 2025-Q3, and 88 to 97 percent by 2030, and notes that at an 80 percent success threshold the feasible task duration never exceeds about five minutes in the period studied. The essay's 'rising tides is the empirical trajectory' drops those scope conditions.