CONCEPT

Liquid Neural Networks

The biologically inspired class of neural network—whose internal parameters keep adapting after deployment—that a worm with 302 neurons inspired and that proved nineteen of them could steer a car.

In 2020, a team in Daniela Rus’s lab built a neural network with nineteen neurons that could steer a vehicle in its lane—not nineteen layers or nineteen thousand parameters, but nineteen. At a time when the dominant deep-learning systems were ballooning toward hundreds of billions of parameters, this was both a demonstration and an argument: an argument that the dominant strategy of the field, scaling toward ever-larger static models, was not the only path to capable intelligence. Liquid neural networks are the technical foundation of that argument. Where a conventional neural network is frozen at deployment—its parameters fixed by training, its behavior determined by what it learned in the past—a liquid network’s neurons are governed by differential equations, and their internal state continues changing in response to incoming data even after training ends. The network stays fluid, adapting on the fly to conditions it was never explicitly trained on. The inspiration was Caenorhabditis elegans, the millimeter-long roundworm with exactly 302 neurons whose neural connectome has been fully mapped—a creature so simple that scientists can study every connection, and yet capable of robust, adaptive, multisensory behavior. If the worm can do so much with so few neurons, something is being done differently from what we are doing with billions.

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI confronts a technology defined by scale—by the concentration of vast computational resources in a handful of data centers accessible to all but owned by few. Liquid neural networks are a quiet dissent from this model: intelligence compact enough to live inside the body of a machine, running on modest hardware, without access to the cloud. This is not just an engineering preference but a vision of a different relationship between people and intelligent systems—one in which the intelligence is embedded, owned, and local rather than rented, remote, and centralized.

The worm-inspired compactness also speaks to a theme the cycle returns to repeatedly: that bigger is not always smarter, that there may be architectural secrets in biology that brute computational scale cannot substitute for. The nineteen-neuron car-steerer does not outperform a frontier model on tasks requiring general linguistic knowledge; but it navigates a real road with a robustness and interpretability that the frontier model cannot match from the cloud, and it does so from inside the vehicle. Different problems demand different kinds of intelligence.

Origin

The theoretical foundations were developed by Ramin Hasani, Mathias Lechner, and collaborators working with Rus at MIT, drawing on differential-equation neuroscience models of the nematode C. elegans. The key 2020 paper “Neural Circuit Policies Enabling Auditable Autonomy” demonstrated that networks of liquid-time-constant neurons could control autonomous driving with dramatically fewer parameters than conventional approaches—and, crucially, with interpretable decision-making that could be inspected by humans.

The approach attracted significant interest because it solved two problems simultaneously: compactness (enabling edge deployment) and interpretability (enabling trust). In 2023 Rus co-founded Liquid AI to develop the approach into foundation models competitive with mainstream architectures but built from first principles rather than the transformer architecture that dominates the field. The company represents a bet that the worm knows something the giant does not.

Key Ideas

Continuous-Time Dynamics. Where a conventional artificial neuron computes a fixed function of its inputs—multiplying by learned weights and applying a nonlinearity—a liquid neuron is governed by a differential equation whose solution continuously evolves in response to the input stream. This makes the network’s internal state a dynamical system rather than a static lookup table, giving it the ability to adapt its own parameters in real time without retraining.

Causal Structure over Surface Statistics. Rus argues that liquid networks extract the causal structure of a task rather than its statistical co-occurrences. A conventional network trained to find an object in a summer forest may learn to rely on the greenness—and fail in winter when the leaves have turned. The liquid network learns what actually matters: the object’s shape and behavior, the features that persist across visual conditions. This robustness to distributional shift is the analogue of the worm’s ability to find food across varying environments.

Compactness and Interpretability. Small size is not only efficient; it is epistemically valuable. With nineteen neurons, a researcher can actually inspect what the network is attending to and trace its decision process—achieving the kind of interpretability that frontier models with billions of parameters deny. The same property that makes liquid networks deployable on edge hardware makes them auditable by humans.

The Worm as Proof of Concept. C. elegans has 302 neurons, a fully mapped connectome, and the capacity for robust, adaptive, multisensory behavior. It is an existence proof that physical intelligence can be achieved at radically small scale. Liquid neural networks are an attempt to extract the architectural principles behind this achievement and apply them to artificial systems—not to mimic the worm, but to learn from what the worm demonstrates is possible.

Debates & Critiques

The central debate is whether liquid neural networks’s advantages—compactness, interpretability, robustness to distributional shift—are genuine architectural innovations or engineering achievements that sufficiently scaled conventional networks will eventually match. Critics note that the nineteen-neuron steering result operates in a constrained domain (lane-keeping on a controlled road) that may not generalize to the complexity of real-world driving. Proponents argue that the compactness and interpretability are not merely engineering conveniences but reflect a fundamentally different learning principle (causal structure extraction vs. statistical co-occurrence learning) that will remain advantageous as tasks become more complex and as deployment requirements push toward edge hardware. A second debate concerns whether the biological inspiration is scientifically load-bearing or metaphorically useful: does the worm connectome actually teach us something about the differential-equation architecture, or does it provide a compelling story for an approach that would have been developed anyway?

In the [YOU] on AI Field Guide

Origin

Key Ideas

Debates & Critiques

Related Entries

Further Reading