Summary
Chelsea Finn, co-founder and CEO of Physical Intelligence (pi.ai), delivers a keynote at Y Combinator tracing the state of the art in physical AI as of mid-2026. Drawing on two years of progress since founding the company, Finn walks through demonstrations of robots performing complex household tasks — washing greasy pans, peeling carrots, making grilled cheese sandwiches, slicing vegetables — and frames the technical gap that still separates robotics from the general-purpose AI breakthroughs seen in language and image models.
Finn structures the talk around two frontier goals: building out-of-the-box robot models that require no per-task fine-tuning, and achieving compositional generalization — the ability to combine learned concepts to handle novel task combinations without exhaustive coverage in training data. She draws the analogy to DALL-E’s 2021 leap in image generation, arguing robotics is approaching a comparable inflection point. A key distinction she draws is that physical AI systems face far stricter reliability requirements than software AI: a bad product recommendation is recoverable, but a robot acting in the physical world without human intervention is not.
The talk also surveys the historical arc of AI deployment — from early ad ranking and recommendation systems through deep learning to ChatGPT and coding agents — to show how generalist models are progressively displacing task-specific systems, and why robotics has lagged behind. Finn’s framing positions Physical Intelligence’s work as the next rung of that progression.
📺 Source: Y Combinator · Published August 12, 2026
🏷️ Format: Keynote Launch
