Waymo Prioritizes Rigorous Evaluation Over Model Performance for AI Safety

Waymo, Alphabet's self-driving car company, emphasizes rigorous evaluation as the primary determinant of AI readiness, rather than solely focusing on model performance. This approach is crucial given the high stakes involved in deploying AI for autonomous vehicles, where split-second decisions impact physical safety. Waymo's strategy, termed 'eval-forced development,' integrates continuous evaluation throughout the AI lifecycle, from training to post-deployment monitoring. This methodology ensures that AI systems are not released until their surrounding tests are mature and reliable, a principle applicable across various industries deploying AI agents.

Manasi Joshi, Waymo's director of engineering for systems intelligence and machine learning, highlighted that Waymo has driven over 220 million fully autonomous miles, reporting significantly fewer serious crash injuries compared to human drivers. The company's evaluation process includes extensive testing during model training, after training, and within both open-loop and closed-loop simulations. This continuous evaluation extends beyond initial deployment, adapting to changes in models, business processes, user behavior, and incoming data. Waymo also focuses on testing rare and dangerous scenarios, using specialized data and metrics to ensure safety, especially concerning vulnerable road users and complex environments. Human oversight remains a critical component, with internal safety leaders approving software releases and service-area expansions, underscoring that human lives necessitate a non-automated decision-making process.

Waymo also addresses the challenge of growing resource demands by pursuing efficiency in compute, storage, and data usage, prioritizing 'data efficiency' by selecting optimal training examples. The company leverages transformers and multimodal generative models as part of its foundation-model strategy, dividing technology between onboard vehicle systems and off-board infrastructure. Internally, Waymo uses AI agents as productivity tools for engineers, but these agents are also rigorously evaluated to ensure trustworthy results. The overarching lesson for enterprises is that successful agentic AI deployment requires clear objectives, representative evaluation data, continuous testing, efficient infrastructure, and accountable human decision-makers.