Frontiers · World models and embodied AI
You can name the three different bets driving the autonomous vehicle frontier — Wayve's end-to-end neural approach, Tesla's vision-only FSD stack, and Waymo's sensor-fused operations model — and explain why their world-model framings differ.
Robotaxi services started 2026 as a curiosity in a few US cities and ended it as a commercial product across the US and China. Waymo is delivering tens of thousands of paid rides per week. Wayve committed to a London robotaxi launch via Uber. Tesla pushed unsupervised consumer FSD to Q4 2026 on its Q1 2026 earnings call and crossed 10 billion FSD miles in May 2026.
Three companies, three different technical bets, three different uses of the phrase "world model."
Wayve trains end-to-end neural networks that take camera input and output steering, throttle, and brake. The "world model" in Wayve's framing is the internal representation that the network learns — a latent of the road, traffic, and predicted behaviors. Wayve's GAIA-style work explicitly trains the network to generate plausible futures of the driving scene as part of the policy itself.
Tesla FSD v13+ also runs end-to-end neural, but on a vision-only stack — no lidar, no radar — using video from the cabin cameras. Tesla rarely uses the phrase "world model" in product communications, but the technical literature on FSD makes clear the system maintains an internal scene model that includes other agents, lanes, and predicted trajectories.
Chapter contains 3 lessons.