Ramin Hasani spent a decade studying a 2-millimeter worm before
anyone outside academia knew his name. Then he solved an equation
that had gone unsolved for over 100 years, published it in Nature
Machine Intelligence, and woke up to an inbox full of term sheets
he never asked for. That paper became Liquid AI.
While every other foundation model company was racing to build
bigger transformers and consume more compute, Ramin went the
other direction. Liquid foundation models cut the exponential
energy cost of AI down to linear — same output quality, a
fraction of the hardware. They run on phones, laptops, and cars.
No data center required. Today his models are downloaded 1.4
million times per week on Hugging Face. Liquid AI raised $270M at
a $2.2B valuation with 110 people — while the companies they
compete with raised that in a single week.
In this ICONS episode, Ramin covers the full arc — from growing
up in Iran with an anatomy professor uncle who made him study the
human brain, to the two professors at Politecnico di Milano who
introduced him to brain-inspired computing, to the PhD at Vienna
and MIT where he discovered that a microscopic worm with 302
neurons could control 95 muscles better than any robot ever
built. He explains the mathematics behind liquid neural networks,
why the transformer architecture hits an inevitable energy wall,
and what two years at Vanguard — the $8 trillion asset manager —
taught him about building for real enterprise value. He also
talks about closing Mercedes-Benz in four months, why he thinks
85% of AI token spend is waste, and what he believes the world
looks like in 25 years when humans stop talking to each other
entirely.
Topics Discussed
Growing up in Iran, studying in Italy, completing his PhD
across Vienna and MIT
The C. Elegans worm: 302 neurons, 95 muscles, and four Nobel
Prizes
How modeling worm neuroscience led to the discovery of liquid
neural networks
Solving a 100-year-old differential equation — and why he
thought he'd win a Nobel Prize
The Hacker News moment: one paper, one morning, and an inbox
full of Silicon Valley term sheets
Transformer architecture explained — and why it becomes an
energy crisis at scale
Liquid foundation models: reducing exponential compute to
linear without sacrificing quality
Two years at Vanguard: what a $8 trillion business teaches
you about enterprise and large numbers
Why foundation model companies should be started by
scientists, not operators
Raising $270M while competitors raised $20B — and why that's
a feature, not a bug
1.4 million downloads per week: how a 110-person company
became one third of Nvidia's open source footprint
Mercedes-Benz: closing an enterprise OEM deal in four months
Shopify: foundation models powering product search and
recommendation at scale
Why they charge per request, not per token — and why
contracts run 4 to 10 years
The talent war: winning 3 out of 10 battles against
Anthropic, xAI, Meta, and Cursor
The taste test: why training a machine learning model is an
art, not a recipe
The next frontier: Bezos's Prometheus, AI for science, and
the world model thesis
Why humans can't grasp exponentials — and what the worm's
nervous system has to do with it
25 years from now: neural chips, parallel communication, and
realities indistinguishable from the physical world
The goal: Liquid foundation models running on 50% of all
devices on the planet by 2029
Kommentare (0)
Melde dich an, um einen Kommentar zu schreiben.