Liquid AI's Ramin Hasani on liquid neural networks, AI advancement, the race to AGI & more! | E1928

Episode Summary

Executive Summary: Ramin Hassani, CEO of Liquid AI, explains his company’s biology-inspired alternative to transformers: liquid neural networks modeled on C. elegans that aim to be smaller, more efficient, more robust, and more interpretable. The discussion covers product commercialization, enterprise use cases, funding, AGI timelines, data licensing, open source vs. closed AI, and the societal impact of automation.

Main Topics: Liquid neural networks and biology-inspired AI (Priority: 5/5): Hassani argues that AI should be designed from first principles using insights from biology and physics, especially the worm C. elegans, whose nervous system inspired liquid neural networks. Efficiency, robustness, and interpretability vs. transformers (Priority: 5/5): He contrasts fixed, parameter-heavy transformer models with adaptable liquid neural networks that he claims are smaller, faster, more energy-efficient, and easier to explain. Productization and enterprise commercialization (Priority: 4/5): Liquid AI is moving from research to enterprise products, offering developer tools and platform access through system integrators and direct enterprise relationships. AGI timelines and definition (Priority: 4/5): The conversation explores what counts as AGI, whether it will arrive first at OpenAI or Anthropic, and whether systems can outperform humans using the same resources. Data rights, licensing, and synthetic data (Priority: 3/5): They discuss the growing importance of paying for training data and building a fairer data licensing ecosystem for creators, publishers, and IP owners. Societal impact of AI and job disruption (Priority: 3/5): Hassani presents a largely optimistic view that AI will change jobs rather than eliminate meaning, and that it can help solve major scientific and governance problems. Open source vs. closed model competition (Priority: 3/5): He argues that closed models will likely keep the lead because concentrated capital and compute make it hard for open source to catch up.

Key Arguments: Liquid neural networks are inspired by the biology of C. elegans and aim to model computation in a more understandable, causally tractable way than transformer architectures. Unlike trained transformers whose weights become fixed, liquid neural networks remain adaptable to incoming inputs, making them more dynamic and potentially more robust. The technology is especially strong for time-series and sequential decision-making tasks such as driving, drones, financial prediction, and medical forecasting. Liquid AI claims large efficiency gains: 10x to 20x less training compute and 10x to 1000x lower inference energy footprint than current models. The company’s business strategy is enterprise-first, using system integrators and developer packages rather than a consumer chatbot play. AGI, in Hassani’s view, means a system that can outperform humans on the same resources and ultimately help solve physics, math, energy, and governance problems. He believes data providers should be compensated and that AI training should move toward licensing schemes rather than unchecked scraping. Closed-source AI is likely to outpace open source because top labs have far more concentrated resources, compute, and proprietary data. AI will be disruptive, but he frames it as an augmentation layer that increases human choice and productivity rather than simply replacing workers. Explainability matters because black-box scaling without understanding increases risk and reduces control over powerful AI systems.

Data Points: Liquid AI company age: ~12 months - Hassani says the company launched on March 30 of the prior year and is about a year old. Seed round valuation: $50 million - He says the first seed round was $5 million at a $50 million valuation. Later financing / total raised: $42 million total - He says the company raised about $42 million overall. Later valuation: $300 million - He states the company’s later financing implied roughly a $300 million valuation. Liquid model size in demo: 19 liquid neurons / ~1,000 parameters - The autonomous driving demo replaced the heavy part of the model with 19 liquid neurons and about 1,000 parameters. Comparison model size in demo: 500,000 parameters - The baseline neural network shown for lane-keeping used roughly 500,000 parameters. GPT-4 parameter count mentioned: 1.8 trillion parameters - Used as a reference point for a large fixed transformer model. C. elegans neuron count: 302 neurons - Hassani cites the worm’s nervous system size as a tractable biological inspiration. Human neuron count: 100 billion neurons - He contrasts the worm with human brain complexity. Training efficiency claim: 10x to 20x more efficient - He claims liquid models can be significantly more efficient to train than transformers. Inference efficiency claim: 10x to 1000x more efficient - He says deployment/usage can be dramatically more energy efficient than current models. System integrator partner examples: Capgemini, Itochu CTC, EY, Accenture - He names several global firms involved in commercialization channels. Team size: About 25 people - He says the company has around 25 of the smartest people on Earth on the team. AGI timeline estimate: 2 to 5 years - He predicts major leaps and early AGI-like systems within this window.

Pivotal Quotes: "If you have AGI, as you said, like you can solve the energy problem, you can solve ... politics, basically, like the structure of governments." — Jason Calacanis: Opening framing of the AGI stakes and the scale of impact if the technology is achieved first. "Liquid neural networks ... are systems that you can have, they can stay adaptable to the input, incoming inputs. That's the major kind of difference between the two." — Ramin Hassani: Core explanation of how liquid neural networks differ from fixed transformer models. "We are designing systems that are kind of white boxes that at every step of the go, we have a lot more control into how these AI systems making." — Ramin Hassani: His argument for interpretability and control as a safer path for advanced AI.

Implications: Liquid AI is betting that smaller, interpretable, biologically inspired models can compete with giant transformers. If true, it could shift AI spending, lower compute costs, expand edge deployment, and intensify debates over data rights, regulation, and AGI safety.