Waymo Rejects End-to-End AI Shortcuts, Defending 15-Year Multi-Model Legacy

Companies like Tesla, Wayve and Waabi are developing AV 2.0 systems with end-to-end neural networks that process raw sensor data and directly output steering commands. [4] Breakthroughs in AI are fueling hopes that these smarter models and more data can provide a shortcut to self-driving vehicles. [4] Srikanth Thirumalai, Waymo's vice president of onboard software, stated in an exclusive interview with Axios that safely deploying autonomous vehicles at scale takes more than just better AI. [4] After more than 15 years and 200 million driverless miles, Waymo says there is no AI shortcut. [4]

Waymo got its start as Google's self-driving car project in 2009, well before the modern deep-learning revolution. [4] In the early days, the company relied on many specialized models, including one for pedestrian detection and another to tell when a light turns green. [4] While Waymo gradually shifted toward fewer, larger foundation models, the company also tried the same end-to-end tools used by competitors. [4] On Wednesday, Thirumalai published a blog post going deeper on 10 AI lessons the company has gleaned from its first 200 million autonomous miles. [4]

Thirumalai, discussing Waymo's goal of demonstrably safe AI, warned that a growing reliance from competitors on single end-to-end AI systems introduces risks. [4] He emphasized that even the best AI models with trillions of parameters still hallucinate. [4] Because physical AI lacks a click reboot, reload, or refresh button, developers have to deal with the consequences of it. [4] The answer to this debate could determine whether the autonomous vehicle race is a long slog that rewards Waymo's head start, or a new track that lets newer rivals hit the road faster. [4]

Waymo’s refusal to adopt purely end-to-end AV systems highlights a fundamental divide in autonomous vehicle engineering. While newer entrants like Wayve and Waabi are betting that modern neural networks can process raw sensor data directly into steering commands, Waymo maintains that hallucination risks make this architecture unsafe for physical deployment. By publicizing lessons from its 200 million driverless miles, Waymo is actively defending its multi-model legacy against the faster development cycles promised by AV 2.0 competitors. What the sources don't address: How the operational costs of Waymo's complex multi-model safety architecture compare to the computational expenses of running trillion-parameter end-to-end models in real time on a vehicle.