Waymo explains why single-model self-driving AI won't work alone

Waymo's August 19 announcement represents a deliberate rejection of the single-model paradigm that has dominated recent AI progress. While large language models and vision transformers have shown remarkable generalization on broad tasks, Waymo argues that autonomous driving—where a single error can cause injury or death—cannot rely on a monolithic AI system. The company's reasoning centers on two failure modes the body explicitly identifies: the long tail of unpredictable edge cases that no training dataset can fully capture, and the millisecond-level safety decisions that require reasoning unavailable from a single forward pass.

Waymo Foundation Model, introduced in 2023, is the vessel through which Waymo operationalizes this modular philosophy. Rather than replacing Waymo Driver's existing architecture, the model enhances specific components: a driver module that steers and accelerates, a simulator for testing scenarios, a validator to assess safety, and a critic to evaluate decisions in real time. This design reflects lessons from Waymo's history—from the Google Self-Driving Car Project (started 2009) through commercial deployment in 2022 and adoption of transformer and VLM technology in recent years. The modular stack has enabled Waymo to serve autonomous ride-hailing services (with riders in US cities) while maintaining the validation and safety infrastructure that single-model approaches cannot guarantee. By structuring AI as a tool within a larger safety-critical system rather than the system itself, Waymo positions its approach as orthogonal to the industry's push toward unified reasoning models.