What a 302-Neuron Worm Knows That a 100-Billion-Parameter Model Doesn't

What a 302-Neuron Worm Knows That a 100-Billion-Parameter Model Doesn't

October 1, 2026
liquid-neural-networks spiking-neural-networks c-elegans continuous-time-computing neuromorphic-hardware

What a 302-Neuron Worm Knows That a 100-Billion-Parameter Model Doesn’t

C. elegans has 302 neurons. That’s not a typo, and it’s not a metaphor for something bigger — it’s the whole nervous system, wired the same way in every individual, mapped down to the synapse since the 1980s. And yet this worm navigates chemical gradients, avoids danger, remembers, and adapts to novel environments well enough that researchers have spent forty years failing to fully replicate what it does with software. Meanwhile we’ve built driving models with parameter counts a million times higher that still get confused by a shadow on the road. If scale were really the answer, this shouldn’t be a fair fight.

I don’t think it’s a fair fight, but not for the reason people usually assume. The worm isn’t smarter because evolution found some efficient set of weights we haven’t discovered yet. It’s smarter because it computes in a completely different medium. Its neurons don’t take turns. There’s no clock. Information isn’t a number that gets updated once per step — it’s encoded in when something happens, at a resolution that our discrete architectures don’t even have a slot for.

Rate vs. timing

Every transformer, every standard deep net, represents a signal’s strength as a magnitude — a float, a scalar, an activation. This is called rate coding, and it’s borrowed loosely from an old, incomplete model of how neurons work: fire more often, mean it more. But real neurons also do temporal coding, where the timing of a spike relative to other spikes is itself the message. The human auditory system can resolve interaural time differences of about 10 microseconds — well below the roughly 1-millisecond refractory period of a single neuron. That precision is structurally impossible to get from rate coding alone. It only exists because biological systems treat time as a first-class computational substrate, not a sequence index.

Transformers do have a notion of time, technically — positional encoding. But it’s a label stapled onto a token, an ordinal tag saying “you are third.” It’s not time as physics; it’s time as metadata. The worm’s nervous system, and its computational descendants like Liquid Neural Networks, do something structurally different: each unit is governed by a differential equation, with a time constant that adapts continuously to the incoming signal. There’s no step. There’s a flow.

Why 19 neurons can out-drive a black box

This isn’t just elegant on paper — it’s the actual mechanism behind results that look absurd on first read. Liquid Time-Constant networks built by Ramin Hasani, Daniela Rus, and their MIT collaborators have handled autonomous driving lane-keeping tasks with as few as 19 neurons, benchmarked against DNNs with orders of magnitude more parameters. The liquid net doesn’t win because it’s cleverly compressed; it wins because it’s solving a fundamentally lower-dimensional problem. A driving scene is a continuous stream, and a network built to be continuous doesn’t have to work around the mismatch between its architecture and its input the way a discrete model does. It adapts its own time constants on the fly, in response to distribution shift, without retraining.

There’s a bigger structural story underneath this, too. Biological neural tissue seems to organize itself toward a specific dynamical regime — “criticality,” the edge between order and chaos — where a system maximizes information transmission without either dying out or seizing into runaway noise. It shows up in neuronal avalanches, in the loss of consciousness under anesthesia, in early biomarkers of Alzheimer’s before any structural symptoms appear. Intelligence, in this frame, isn’t a property of having enough units. It’s a property of a system tuned to a very specific, very narrow operating point in continuous dynamics — a point that’s meaningless to even ask about in a network that only exists at discrete integer time steps.

Convenience, not a lack of value

None of this happened because continuous time was a bad idea. It happened because backpropagation through discrete layers was tractable in 2012 and backpropagation through spike timing wasn’t. We built what we could differentiate, not what best matched the phenomenon we were trying to reproduce — and then spent a decade mistaking the scale of what we built for evidence that we’d chosen the right foundation. Closed-form continuous-time networks and diffusive memristors that move actual silver ions instead of simulating voltage are early signs that the field is starting to correct course, but they’re still a rounding error next to the capital pouring into bigger clocked models.

I keep coming back to the fact that a fixed-wing jet doesn’t flap, and it still flies better than any bird. Maybe intelligence really is like flight — a functional target reachable by unrelated mechanisms, and biological fidelity is a red herring. But flight is a mechanical problem with a clean physical objective. I’m not convinced intelligence is that kind of problem. If it’s closer to consciousness — something that might only exist as a flow, not as a sequence of snapshots — then no amount of clock cycles gets you there, no matter how many of them you buy.

Sources

Comments

comments powered by Disqus