Ramin Hasani

The CompanyWhat Ramin Hasani Actually Does

Ramin Hasani runs Liquid AI, the MIT CSAIL spin-off he co-founded in 2023 with Mathias Lechner, Alexander Amini, and MIT's Daniela Rus - who was, until Liquid, his postdoctoral advisor. Founders and their PhD supervisors don't usually end up on the same cap table. In this case, they built the same architecture together for four years, so the arrangement is more like a joint patent than an unusual startup.

The architecture is called a liquid neural network. It descends from Hasani's doctoral thesis at Vienna University of Technology, which he defended with distinction in May 2020. Liquid networks are continuous-time systems. The parameters shift while the model is running, in response to the input. If you have used a transformer, the mental model to hold is that transformers are frozen at inference and liquid networks are not.

Liquid AI's product is a family of Liquid Foundation Models - LFMs - that compete with transformer-based systems on general-purpose tasks but claim large advantages on memory, latency, and edge deployment. The company is selling this as a substrate for foundation-model intelligence that runs on the device rather than in the data center. Automotive, defense, consumer electronics, and enterprise infrastructure are the stated verticals.

In December 2024, AMD Ventures led a $250 million Series A into Liquid AI at a $2.35 billion valuation, with Automattic and OSS Capital participating. AMD is a chipmaker with an obvious interest in a foundation-model architecture that runs efficiently on hardware it sells. Hasani, as CEO, is the person who has to convert that architectural bet into a business.

The OriginThe Worm on the Cap Table

Caenorhabditis elegans is a nematode, roughly one millimeter long, transparent, and living in soil across most of the planet. Its nervous system is completely mapped. There are 302 neurons, wired in a specific and known way. The connectome is small enough to fit in a diagram and large enough to produce the animal's foraging, mating, and avoidance behaviors.

Hasani spent years working with that connectome. The intuition he took from it: complexity in the world does not require complexity in the network. If you get the dynamics right, small can suffice. A drone that had never seen weather could stay upright in weather. A car in an unfamiliar city could still make the correct turn. The empirical result at MIT CSAIL was that liquid networks generalized where fixed-weight networks brittled.

This is a bigger claim than it sounds. The mainstream position in AI over the last decade has been that intelligence is what you get when you scale up transformers on more data and more compute. Hasani's counter-argument is that a different math - continuous-time dynamics rather than discrete attention layers - can produce better behavior with fewer parameters. He now has $537 million in cumulative funding and a chipmaker's endorsement to prove it.

The PathVienna, Vanguard, Cambridge

Hasani, who is Persian-Austrian, did his doctorate at TU Wien. The dissertation was on liquid neural networks and was nominated for the TÜV Austria Dissertation Award in 2020. He then moved to MIT CSAIL as a postdoctoral associate with Daniela Rus, where the LTC (Liquid Time-Constant) paper - now a foundational citation for the field - was published in 2021.

A less obvious line in his CV is the joint appointment as Principal AI and Machine Learning Scientist at Vanguard, the index fund manager. Quantitative finance is one of the more demanding testbeds for machine-learning systems, because the return on getting the prediction right is measurable in dollars and the return on getting it wrong is measurable in dollars. That period sits between the academic work and Liquid AI.

He collected an HPC Innovation Excellence Award in 2022 and, with collaborator Djordje Zikelic, an Outstanding Scientific Achievement Award at IST Austria in 2023, for work on proving safety properties in stochastic ML systems. Safety proofs and dynamical systems are the two threads that keep showing up.

What's InterestingThe Contrarian Bet

The interesting question at a company like Liquid AI is not whether the models work - the benchmarks exist and the customers are running them - but whether Hasani's architectural bet holds against the scale bet. Transformer people believe more parameters and more data will keep producing better outputs for the foreseeable future. Liquid people believe that continuous-time dynamics and adaptive parameters are a categorically better substrate, one that can produce comparable performance with an order of magnitude less compute.

Those views are not fully compatible. If Hasani is right, the industry's compute intensity is a specific engineering choice rather than a fundamental requirement. If he is wrong, Liquid AI is a very well-funded niche vendor of edge-optimized models. AMD Ventures is betting the former; time is the referee.

What is notable about Hasani specifically is the durability of the through-line. TEDx Vienna in 2018: worm brains. TU Wien dissertation, 2020: liquid networks. MIT CSAIL, 2021: LTC paper. Liquid AI, 2023-2024: LFMs and $250 million. There is no pivot in the story. It is one idea, developed patiently, that met a market.