How End-to-End Learning Created Autonomous Driving 2.0: Wayve CEO Alex Kendall

Wayve CEO Alex Kendall explains how end-to-end neural nets beat hand-engineered AV stacks and why OEM partnerships beat robo-taxi city rollouts.

• Wayve onboarded Nissan in Tokyo within 4 months of first driving there, on a vehicle it had never touched before.
• 90 million cars are built annually; Tesla sells ~2 million, leaving 88 million units as the addressable OEM market.
• Safety and flow metrics generalize uniformly across countries; utility (road signs, navigation rules) requires hundreds of hours of new data per market.
• Wayve’s GAIA generative world model simulates multi-camera sensor environments for training, enabling synthetic data augmentation without replacing real-world miles.
• Camera-only handles human-level driving; eliminating the long-tail accident residual requires radar and LiDAR — the sub-$2,000 sensor suite now standard on mass-produced vehicles.
• One shared base model serves all OEM fleets; per-customer specialization consumes under 1% of total training cost and time.
• Lingo (2021) was the first vision-language-action model for autonomous driving; next-gen Nvidia Thor compute will enable on-board language reasoning.
• AV 3.0 speculation: vehicle-to-vehicle mesh communication could eliminate traffic lights and reduce onboard sensor requirements entirely.

2025-11-18 · Watch on YouTube