Ramin Hasani
Co-founder and CEO of Liquid AI
aka Ramin
“Our mission has always been building efficient general purpose AI at every scale. That explores the computational graphs of intelligence beyond transformer... that brings the same level of intelligence than a frontier model into, let's say on a CPU.”
“we brought one of our multimodal foundation models that is only less than 1 gigabyte in size and it can go inside the car's chip... The chip could be as cheap as $60.”
“it takes us 350 years to really fine tune a 2 billion parameter model with this framework... Anthropic has been having a lead on all of these things because they thought about this before everybody else.”
“the model layer is not anymore the place where you can actually extract value, especially if you're not hitting the maximum frontiers... you need to leave some room for fine tuning these models.”
Ramin Hasani
One-line summary: Co-founder/CEO of Liquid AI; pioneer of Liquid Neural Networks and post-transformer small language models for on-device deployment.
What they're known for
Co-founder and CEO of liquid-ai. Did his PhD under Daniela Rus at MIT CSAIL (after starting in Vienna under Radu Grosu), where he and co-founder Matthias Lechner derived Liquid Neural Networks from the 302-neuron nervous system of the C. elegans worm — a continuous-time, recurrent, "post-transformer" architecture. Liquid AI is, per dave-blundin (an early investor), "the only foundation model company that I know of that reached unicorn status coming out of MIT."
Why they matter to artificial-intelligence
Ramin is the thread's primary voice on post-transformer architecture and on-device small language models. His claims matter on two axes: (1) that efficient, specialized sub-100B models can run below the operating system on cheap edge chips (a <1GB multimodal model on a ~$60 automotive chip, shipping in Mercedes-Benz North America), and (2) a grounded skepticism about "recursive self-improvement" hype — he argues real RSI means weight/architecture change, is computationally intractable at the framework demonstrated by WECO, and that Anthropic leads because it started earliest. See post-transformer-architectures, edge-inference-shift. For stock-market, his on-device thesis routes demand to edge silicon (Qualcomm, Samsung, AMD) and names auto/PC OEM deployment channels.
Said
Speaker-attributed claims extracted from diarized sources. Each bullet mirrors one entry in quotes: frontmatter — keep them in sync.
-
On post-transformer-architectures, edge-inference-shift:
"Our mission has always been building efficient general purpose AI at every scale. That explores the computational graphs of intelligence beyond transformer... that brings the same level of intelligence than a frontier model into, let's say on a CPU." — 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on (2026-07-17)
-
"we brought one of our multimodal foundation models that is only less than 1 gigabyte in size and it can go inside the car's chip... The chip could be as cheap as $60." — 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on (2026-07-17)
-
On autoresearch-recursive-self-improvement:
"it takes us 350 years to really fine tune a 2 billion parameter model with this framework... Anthropic has been having a lead on all of these things because they thought about this before everybody else." — 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on (2026-07-17)
-
On open-vs-closed-source-model-economics:
"the model layer is not anymore the place where you can actually extract value, especially if you're not hitting the maximum frontiers... you need to leave some room for fine tuning these models." — 2026-07-17-podcast-moonshots-mira-murati-s-975b-open-model-ramin-hasani-on (2026-07-17)