Twelve neurons are enough to park a car, and that fact should reframe how we think about neural network scale

The number Ramin Hasani puts on it is the kind that stops a reader mid-sentence. With 12 neurons, he says, a liquid neural network can parallel park a car autonomously. Not 12 million. Not 12 billion. Twelve.

That claim is not unverified. A report published via EurekAlert describes exactly this result: a car-parking task solved autonomously through compact, liquid-style neural networks using a handful of simulated neurons. The finding was on the record years before the idea entered broader circulation. Hasani’s framing matches the published benchmark closely enough that the two reinforce each other rather than merely rhyming.

Liquid neural networks are architecturally distinct from the large, dense models that currently dominate AI development. Their defining property is that the behavior of each neuron is governed by a differential equation, meaning the network’s internal dynamics change continuously in response to input rather than staying fixed after training. That flexibility allows a very small number of neurons to represent and respond to complex, time-varying environments, which is precisely why a task as geometrically demanding as parallel parking can fall within reach of a 12-neuron system.

With 12 neurons with with actually 12 neurons you could you could parallel park autonomously like a car Ramin Hasani

The parking result is a benchmark, not a general claim about everything neural networks must do. Autonomous parking is a bounded, sequential task: the car occupies a known physical space, the goal state is well-defined, and the required sensing is limited. A liquid network’s temporal adaptability fits that problem structure well. The same neuron count does not imply that arbitrary perception or language tasks collapse to comparable sizes. What the result does establish is that the instinct to add parameters as the default response to a hard problem is not always the right one.

That instinct runs deep in contemporary AI development. The scaling hypothesis, roughly the idea that more compute and more parameters reliably produce better models, has driven the infrastructure investment logic of the last several years. Hasani’s result does not refute scaling where it applies. It does suggest that scaling has been applied to some problems where a different architecture would have produced a more efficient solution. Autonomous driving is one of those problems. The domain’s real-time, sensor-dense, temporally continuous demands map onto liquid network strengths in ways that standard feedforward or transformer architectures do not naturally exploit.

Hackaday and other engineering-focused outlets have noted the “do more with less” efficiency theme in liquid neural network research, particularly in the context of autonomous vehicles. That framing has sometimes read as aspirational. The parking benchmark reported by EurekAlert gives it a concrete floor: 12 neurons, one completed parking maneuver, no asterisk on the result.

The broader question the evidence raises is about how efficiency is counted in neural network research. Parameter count is a tractable metric, easy to report and easy to compare. It is not the same as computational necessity. A system that solves a real-world driving task in 12 neurons is not a curiosity or a demo. It is evidence that the design space most practitioners are exploring occupies a small and possibly non-optimal corner of what is available. Whether liquid architectures scale to harder problems, or whether their efficiency advantage is specific to the temporal and continuous domains where their dynamics shine brightest, is the live research question. The parking result does not answer it. It makes the question harder to dismiss.