Building a Foundation Model Company from First Principles

How do you turn a fundamental research idea into intelligence that works on the devices people use every day? In this keynote, I’ll share the technical journey from inventing liquid neural networks to building Liquid AI. Instead of insisting on a particular architecture, we built AI that designs AI, allowing competing designs to be evaluated against capability, memory, and latency requirements with the target hardware inside the design process. But architecture is only the beginning. I’ll walk through the interconnected decisions across data, pretraining, post-training, quantization, inference software, and developer tools. Each component shapes the final experience: how quickly a system responds, how reliably it performs, where data lives, and how easily developers can build with it. Our ambition is to bring our models to 50% of the world’s annually produced devices by 2029. I’ll explain what this ambition demands of the technology, the product experience, and the company, and the lessons builders can apply to their own ideas.