19 Neurons Can Drive a Car. Liquid AI Wants Foundation Models on Every Chip

Nathan Labenz opens with the arc he has watched from afar: Liquid AI's unusual architectures, then a company that has grown more commercial. Ramin Hasani, cofounder and CEO, dates the spinout to three and a half years earlier, out of MIT CSAIL, on work the team had already been doing for about a decade. The objective function at MIT, he says, was always the same: maximize the intelligence you can fit into a smaller algorithm. Efficiency was the cornerstone.

The original setting was robotics. They wanted learning systems that could live on CPUs, NPUs, or small GPUs mounted on physical machines, without millions or billions of parameters, while matching the reliability of much larger models. In the open world—an autonomous car, a flying drone, a fixed-wing vehicle—distribution shift arrives immediately. Humans and other animals handle that comfortably. The field as a whole has been chasing that property; Hasani's group started in brains, from first principles, with how two neurons exchange information.

They started with *C. elegans* for a practical reason. In 2015, when Hasani began the PhD work with cofounder Matthias Lechner, it was the only animal whose entire nervous system was known. About 300 cells produce rich sensory-reactive behavior and dexterous control, smaller than any neural net then used for autonomous systems, and in some ways better than the best robots. The plan was to understand two-neuron exchange in the worm, then build more complex circuits along the evolutionary path of nervous systems.

Because the worm is tiny, its neurons do not spike. They are graded, electrotonic, and therefore differentiable, closer to artificial nets than to spiking models. That made backpropagation legal. Equations already described the dynamics; the more complicated you make them, the more they mimic biology and the harder they are to scale. Standard neural nets throw those differential equations away and keep a sigmoid or a matrix multiply so the system can grow. Hasani wanted to put the elaborate computation back into each cell.

The early results were startling. Twelve neurons could parallel-park a small car. Nineteen could drive. Thirty could navigate a drone. They named the models liquid time-constant networks, *liquid* because the dynamics stay flexible after training: the system can still react to new inputs, and it learns during backpropagation how to be adaptive. Each neuron is more expensive than a standard unit, but you need far fewer of them, and out-of-distribution performance was strong enough for robotics. Boeing hosted Hasani's MIT postdoc; with the U.S. Air Force they showed a handful of these neurons could fly jets.

Each liquid neuron is a differential equation, so more neurons make the forward and backward passes more expensive, up to cubic complexity. You can roll the system out with a numerical solver, taking more steps for more accuracy, but you never reach infinite precision and you cannot scale. The bet was to solve the whole system in closed form. The lineage goes back to 1907, when Louis Lapicque modeled membrane potential; that equation became the basis of ion-channel modeling. Hodgkin and Huxley refined single-neuron dynamics in the early 1950s and won the Nobel Prize in 1963. After that, textbooks treated these equations as having no known closed-form solution. Liquid nets were in that family.

Hasani's group solved it in 2022. The result, closed-form continuous-time systems, ran in *Nature Machine Intelligence* that November. Without a numerical solver, they could put billions of neurons next to each other instead of hundreds, while keeping the nonlinearity. In early February 2023, *Quanta Magazine* profiled Hasani and Lechner. His inbox filled with VCs. Silicon Valley wanted to throw term sheets because the math was a different category from transformers: grounded in biology, not a linearized dynamical system of the kind that made state-space models and scalable convolutions possible.

The company's mission hardened around that moment: efficient general-purpose AI systems at every scale. Hasani claims Liquid is probably the most efficient foundation-model company on the planet, and he offers numbers. Liquid foundation models, or LFMs, were fifth in the U.S. by Hugging Face downloads, behind Google, Meta, Microsoft, and Nvidia, at more than a million downloads a week. They had released more than fifty model instantiations that enterprises were using. The compute behind that, he says, was about 1,000 GPUs in-house.

Labenz notices that the old papers count neurons and everyone else counts parameters. Early liquid nets treated the neuron as the unit of computation: one cell is an equation with about seven parameters and three layers of internal feedback. Today's LFMs use the industry meaning. A 1-billion-parameter liquid foundation model is 1 billion parameters, comparable to any GPT-class model.

The bottleneck on the original paradigm is turning sequential computation into parallel computation. Scalars become vectors, then matrices and tensors. Linear systems tensorize cleanly; nested nonlinear relationships do not. That is why state-space models linearized their dynamics. You can apply a pointwise nonlinearity *after* a linear dynamical system. You cannot, without approximation, tensorize a system whose parameters themselves are governed by nonlinearity. Even the closed form still has to run sequentially at a few hundred thousand to maybe ten million neurons. Sequential scan can speed nonlinear operators without linearizing them; that research is unfinished.

Those tiny nets still fit in 1 to 25 megabytes and run on a Raspberry Pi. Closed-form variants match the original ODEs on speech synthesis and multivariate sequence prediction. They are adaptive, not continual learners: new rain on a camera is low-pass filtered and the dynamics shift, but the weights stay fixed. Dynamics are an extra axis for compressing information. Hasani's rule from large architecture searches: scaling laws define architecture. At tens of trillions of parameters you want unstructured operators—pure matmuls, pure convolutions, pure scans—and no fancy gates. Below roughly 100 billion to a trillion parameters, bias helps: gating, recurrence, convolution. Transformers scale because they are unstructured enough that any matrix can multiply any matrix.

Liquid's customer pitch is no longer "we have one hammer." In-house they built automated foundation model design, AFMD: hardware in the loop, an evolution strategy over about 50 to 100 operators, optimizing memory, latency, and speed with no quality sacrifice. Quality is not perplexity and not a public leaderboard. It is about a hundred benchmarks plus the actual downstream task. The motive is to remove human bias. Hasani says even Anthropic and OpenAI have small groups he calls the Avengers of architecture, post-training, or pre-training, calling shots from personal experience. Recursive self-improvement is starting to fix that, but the real move is to give the decision to a systematic search.

They had reason to need a referee. Founding scientists include Jimmy Smith, inventor of S5, who brought parallel scan into the SSM world, and Stefano Massaroli and Michael Poli, designers of the Hyena Hierarchy convolution path. Those are linear-system people in a room with liquid's nonlinear, control-theoretic nets. The search ran scaling laws from 10 million parameters to 72 billion. When human gating fashions—Mamba-style gates, gated DeltaNet, hand-tweaked linear attention—were stripped away, the winner for general-purpose CPUs was a double-gated convolution. In LFM2, 70 to 80 percent of layers are those unstructured 1D gated convolutions. They replace attention, cut compute and memory, raise speed, and stay competitive with transformer alternatives on quality. NPUs inside AMD, Qualcomm, and Intel AI PCs get their own variants from the same loop.

Labenz restates the headline: attention is still something you need, but the other layers can be extremely simple if they are gated. Hasani agrees, then puts architecture in its place. Input dependence was the biological discovery; Liquid S4, a year and a half before Mamba, introduced input-dependent SSMs. In the backward pass, an input-dependent operator encodes its own dynamics into the gradient, so the network learns a second axis beyond parameter count. Gating helps discrete sequences too, including language. At 100 trillion parameters, maybe matrix multiplication is enough. Until then, architecture research's most important job is efficiency, because demand is exponential and the large labs are wiping compute off the planet. The next generation of brain-like, 20-watt intelligence will not come from architecture alone. It needs new learning algorithms, data representation, and objectives. Current theory assumes i.i.d. data. "We are building averaging machines," which is why the writing is bland and the models are sycophantic. Multi-agent setups and test-time compute are workarounds.

Different data shapes want different winners. DNA has a tiny vocabulary and sequences Hasani puts in the range of 1 to 100 billion elements; quadratic attention explodes, so you want pure convolutions, SSMs, recurrent nets, or linear attention. Pure audio—voice-to-voice, noise-to-clean—still favors RNNs, especially in the low-data regime, where feedback fills the gap. Physics-informed nets fit factory digital twins. The original open-source liquid repo is still used for sensor and physical time series because the math is continuous-time. Video is unsettled between diffusion and autoregression. At small scale, transformers lose to these biased systems; at the largest scale, transformers win again.

The original differential-equation nets still ship where latency has to be microseconds, or where you are simulating a chemical reaction. Commercially, Hasani names the category device foundation models: every processor outside a data center. Inside data centers they chase ultra-low latency, extremely long sequences, and tiny memory. Shopify is in production on recommendations, search, and product-catalog understanding, with internal click-through metrics up. The landmark contract is Mercedes-Benz: in-car audio and visual intelligence. Talk to the car and the voice comes from a liquid foundation model, 600 megabytes, small enough for the smallest processor in the vehicle.

The mobile market alone is about $500 billion a year, comparable to data centers. Add laptops and Labenz's "dark compute" is on the order of a trillion dollars of processors in pockets and on desks. Hasani's reply is blunt: you have to use it, because there is not enough energy to host all intelligence in the cloud. Data extraction should not call a frontier model. Heterogeneous silicon means AFMD plus post-hoc kernel work, quantization, and even architecture search over matrix-multiply kernels. For undisclosed NPU IP, you need another search; for kernels the open-source world has seen, agents can already automate a lot.

Labenz asks what hardware companies should prioritize. Hasani thinks the CUDA moat is getting shallower because agents can now write kernels. The stack to climb is the intelligence layer, which is what Nvidia is doing with Nemotron—he estimates a couple of billion dollars went into it—and with NIM: buy the hardware, get a usable model, fine-tune with Megatron. Other chip firms still optimize for open-source architectures, which is necessary but not differentiating. Software can close a bandwidth or memory gap. Nvidia is entering the CPU-device game; Google is pushing Android laptops (Hasani thinks the name is Aluminum OS); Meta will bring devices too. Glasses, he suspects, become a hardware arena that might replace laptops. The new base of software is a foundation model on the chip. If the default is fast, tunable, and already orchestrating several instances, why switch—so long as the hardware still runs the open-source ecosystem?

On local coworkers, Labenz describes five years of email, Slack, and transcripts in a local database, with desktop Claude calling local tools and sending results to the cloud. Hasani's blueprint is a local orchestrator: hard tasks go to the cloud; PII gets stripped by a small on-device model before anything leaves; extraction and classification stay on the edge. LFM2 24B A2B—24 billion total parameters, 2 billion active—is built for tool-calling agents on consumer hardware, but no local model is there off the shelf. You have to fine-tune. That, he says, costs tens to low thousands of dollars, not tens of thousands, on compute you already have. Liquid will announce a platform in the coming months that you can hook to a channel; it can even synthesize data so the model never sees your private set.

Asked whether biology already sits on the Pareto frontier of intelligence per watt, Hasani is skeptical that current algorithms can approach the brain's 20 watts. Humans arrive with evolutionary pretraining. Today's in-context learning is a mushy least-squares, a vague gradient descent that emerged from next-token prediction. Brains also run reinforcement learning, internal simulation, and Bayesian inference. Intelligence is an emergent property; to miniaturize it you have to plant more seed algorithms at design time so curiosity-driven learning can grow at low energy. Next-token prediction is the first seed. The second has not been found. He closes as a techno-optimist: science is curiosity plus understanding the world, agents have dropped the barrier to frontier research, and the direction is already clear—get intelligence off the data center and onto every device that should have it.