Modular AV stacks had 10 separate models. End-to-end has one: video in, steering out. Inspired by ChatGPT, models like Tesla Hydra, Wayve GAIA-1 and NVIDIA GenAD are the new foundation of driving.
Why End-to-End Wins
- Gets better with more data — 10,000x fleet videos directly improve driving, no hand-label bottleneck
- Learns subtle human behavior: nudging around double-parked car, eye contact nuances
Risks
Explainability drops. Wayve’s LINGO solves by outputting language: “I slowed because school bus stopping.” Regulators love it — EU AI Act requires rationale for safety-critical AI.