10/15/2026
What happens when an AI driver can imagine what comes next?
That was the focus of our latest Signal, where Dr. Wei Zhan, Co-Director of Berkeley DeepDrive and the BAIR Center for Humanoid Intelligence and Chief Scientist at Applied Intuition, joined us to explore what he calls AV 3.0: training end-to-end in closed loop, where a model learns from the consequences of its own decisions rather than only from human demonstration.
Closed-loop learning needs a world to practice in, which is where simulation and world models come in. We dug into world-action models, which learn to understand the world and act in it together, and why the ability to predict what happens next matters for a driver. We also talked about what makes a simulator good enough to learn in, what changes when models can learn from raw video at scale, and what it takes to trust a world model before training on it.
We also looked beyond the road, at what autonomy's lessons could mean for physical AI more broadly.
Thanks to everyone who joined us at Kodiak AI in Mountain View for the discussion.
Missed this one? Stay tuned. More Signal conversations are on the way.