Gestern • 58 Min.

Mathias Lechner did his research in Vienna on the nervous system of C. elegans, a worm with just 302 neurons. That work became liquid neural networks, a self-driving demo that steered a car with 19 neurons, and eventually Liquid AI, the MIT spinoff he co-founded and leads as CTO. AMD led Liquid AI's $250 million round, and its models power the in-car assistant in Mercedes' new CLA. In this episode, Tobi and Mathias dig into what it really takes to run AI without the cloud. Mathias explains why Liquid AI starts with the hardware instead of model size, why memory bandwidth and memory size matter as much as FLOPS on a phone or in a car, and why he believes there's no long-term technical moat in AI, only a race where silicon partnerships give you the edge. He also shares his own coding agent setup, how the team judges training data, and what he'd tell his 2017 self. CTOs will come away with a clear picture of which workloads will move on-device and which stay in the cloud, what limits local inference today, and how to start testing small open-weight models themselves. - From a worm's nervous system to liquid neural networks and Liquid AI - Why a 19-neuron network could steer a car while the industry scaled up - No long-term moat: silicon partnerships and access to pre-release chips - Speed is non-negotiable: designing models around the hardware - Memory bandwidth, memory size and FLOPS: the limits of on-device AI - Cars, phones and the Mercedes in-car assistant - Data quality, evals and his YOLO-mode coding agent setup