Waymo Co-CEO Dmitri Dolgov: "Move Fast And Ship Safely"

Full Title

Waymo Co-CEO Dmitri Dolgov: "Move Fast And Ship Safely"

Summary

This episode features Waymo Co-CEO Dmitri Dolgov discussing the challenges and lessons learned in building and safely deploying AI in the physical world, contrasting it with digital AI.

Dolgov outlines seven key lessons, emphasizing the critical difference between a demo and a real product, and the importance of rigorous safety, robust architecture, and a comprehensive ecosystem for scaling physical AI.

Key Points

• Physical AI development requires a "move fast and ship safely" approach, prioritizing robustness and safety from the outset, unlike the "move fast and break things" mantra of digital AI.
• Building AI for the physical world faces unique challenges: a higher cost of error (human lives vs. tokens), the need for real-time low-latency decisions, the absence of a readily available digitized internet equivalent for data, and the strict validation requirements before deployment.
• The transition from a successful AI demo to a scalable, reliable product is a significant hurdle, often taking years and requiring the pursuit of many "nines" of performance and reliability, which becomes exponentially harder to achieve.
• Selecting the right technology and architecture is crucial, as early performance gains from one technology might plateau before reaching the required product performance, necessitating an honest assessment of long-term scaling potential.
• Waymo's strategy involves using multiple, complementary sensing modalities (cameras, lidar, radar) to achieve robust perception, rather than relying on a single sensor type, which is essential for safety and reliability.
• Hardware costs and capabilities evolve rapidly; companies should not anchor their strategy to current component prices but design for future improvements and commoditization.
• Embracing technological breakthroughs (like deep learning, transformers, VLMs) is essential, but the real challenge lies in integrating these innovations into a production, safety-critical system without regressions and while maintaining scaling momentum.
• A "structure-augmented end-to-end" approach in AI models, combining learned representations with materialized structured elements, is key for boosting performance, simplifying validation, and improving scaling laws.
• Highly realistic, large-scale simulation is critical for training and evaluating physical AI agents in closed-loop scenarios, allowing for testing of rare and dangerous situations that are infeasible to replicate in the real world.
• Building a complete AI ecosystem, including the agent, a robust simulator, and a critic for evaluation, is necessary to create a powerful flywheel for continuous improvement, driven by real-world data and guided by rigorous metrics.
• Evaluation and metrics are more critical than the AI model itself for steering technological development and ensuring a product's safety and effectiveness in the physical world.
• Trust is paramount in physical AI deployment, and it is earned through relentless, real-world validation and transparent safety data, not just through impressive technology or demos.
• The next decade of AI advancement is expected to occur in the physical world, offering massive opportunities for innovation and impact.

Conclusion

Developing AI for the physical world requires a fundamental shift from "move fast and break things" to "move fast and ship safely," prioritizing robustness and reliability from the start.

The path from a compelling AI demo to a production-ready, safe, and scalable product is arduous, demanding a deep understanding of technical challenges, architectural choices, and the relentless pursuit of quality.

The future of AI lies in the physical world, with significant opportunities for companies that can master the complexities of building safe, trustworthy, and impactful AI systems that genuinely improve people's lives.

Discussion Topics

• What are the most significant differences between developing AI for digital applications versus the physical world, and what are the key challenges in bridging that gap?
• How can companies ensure they are not just building impressive AI demos but are genuinely creating scalable and safe products for real-world deployment?
• Considering the rapid advancements in AI, what is the role of simulation versus real-world testing in validating the safety and performance of autonomous systems?

Key Terms

VLMs

Vision-Language Models, AI models that can process and understand both visual and textual information.

Transformers

A type of neural network architecture that excels at processing sequential data, widely used in natural language processing and increasingly in computer vision.

Continual Learning

A machine learning paradigm where a model learns from a continuous stream of data, retaining knowledge from previous tasks while learning new ones.

Foundation Model

A large AI model trained on a vast dataset that can be adapted to a wide range of downstream tasks.

Lidar

Light Detection and Ranging, a remote sensing method that uses light in the form of a pulsed laser to measure ranges (variable distances) to the Earth.

Radar

Radio Detection and Ranging, a system that uses radio waves to determine the range, angle, or velocity of objects.

Doppler Effect

The change in frequency of a wave in relation to an observer who is moving relative to the wave source. Used in radar to measure velocity.

VLM

Vision-Language Model, an AI model capable of processing and relating visual and textual data.

Reinforcement Learning

A type of machine learning where an agent learns to make a sequence of decisions by trying them out in an environment and receiving rewards or penalties.

World Models

AI models that aim to represent and understand the dynamics and physics of the environment, enabling agents to predict outcomes and plan actions.

Timeline