A Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI
- Source
- youtube.com
- Author
- Latent Space
- Date
Context
Liquid AI says it designs model architectures for the hardware that will run them. In the interview description, CEO Ramin Hasani explains that the company uses an automated, recursive search framework that designs hybrid architectures (mixes of different building blocks) based on the target hardware, such as CPUs, GPUs, or NPUs. The model's design becomes something to optimize for where it will be deployed. Episode description
That connects its architecture research to its interest in efficient, low-latency models that run on devices or in private deployments. The company presents hardware-specific design as a route to efficient deployment; the description gives no measured speed or energy advantage, so the benefit is a design goal rather than a reported result. Architecture discussion
- Hasani notes C. elegans handles complex motor control with 302 neurons that behave like artificial neurons rather than spiking, which led Liquid toward continuous-time models instead of spiking networks.
- On scaling recurrent models, Hasani warns there is no free lunch: linearizing complicated dynamics to make them tractable loses expressivity.
- Liquid partners with enterprises including Shopify and Mercedes-Benz on on-device or private deployments, and maintains an open library, Leap, for and deploying models in formats like GGUF.
- Hasani estimates about 90% of the market is today and expects the next wave of companies to capitalize on customization tokens.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
- inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…






