Small, adaptive models are underrated

My doctoral research applied liquid neural networks to intrusion detection on IoT networks. If you have not met them, liquid networks are small, continuous-time models inspired by the nervous system of a roundworm with 302 neurons. They adapt their dynamics to the signal as it arrives, which makes them remarkably good at time-series problems, and they do it with a parameter count that rounds to zero next to a modern transformer.

Working with them taught me a bias I have kept: reach for the smallest model that can carry the mission, and spend the savings on data quality and evaluation.

That bias is unfashionable. The default motion in the industry is to route everything through the largest model the budget allows, and for open-ended language work that is often right. But a large share of real operational problems are not open-ended language work. They are narrow, structured, and continuous: is this network traffic hostile, is this reading drifting, is this track behaving like a threat. For problems shaped like that, a small model trained on the right data is frequently more accurate, radically cheaper, auditable in ways a frontier model is not, and able to run where the mission actually happens, on a sensor, on a vessel, at the edge of a disconnected network.

There is a second advantage that gets less attention: small models are easier to evaluate honestly. You can characterize their behavior across the whole input envelope, stress them adversarially, and write down what they do, in language a certification authority can act on. Try writing that document for a hundred-billion-parameter model that changed versions twice during your test window.

The field is quietly rediscovering this. Edge deployment, on-device inference, and distilled task-specific models are all the same idea wearing different names: fit the model to the mission, not the mission to the model.

Big models are a wonderful tool. They are not the only tool, and in the places where AI has to earn certification, run disconnected, or justify every decision it makes, they are often the wrong one. The roundworm got there first.