Waymo CEO on AI's Real-World Challenges

Waymo Co-CEO Dmitri Dolgov recently shared valuable insights into the complexities of building and deploying AI in the physical world, drawing from Waymo's nearly two decades of experience with its autonomous driving technology. Speaking at a Startup School event, Dolgov outlined seven key lessons learned, emphasizing the stark differences between digital and physical AI applications.

The 1% Demo vs. the 99% Product

Dolgov began by highlighting the vast gap between a working AI demo and a truly shippable product. He stressed that "the demo is the first 1%, the nines are the product." He illustrated this point with Waymo's own journey, noting that achieving a capability-complete autonomous driving system in 2010 took about 18 months, but scaling it to a reliable public service took another 15 years. This extended period was necessary to achieve the "many nines of performance and reliability" required for public deployment.

The full discussion can be found on YC's YouTube channel.

He explained that each additional "nine" of reliability demands exponentially more effort, necessitating fundamentally different approaches like fully redundant systems and tiered fallback architectures. Dolgov cautioned against overspending on spectacular demos, advising founders to "count your nines before you count your demo views."

The Four Gaps of Physical AI

A core theme of Dolgov's presentation was the identification of four key gaps that differentiate building AI for the physical world compared to the digital realm:

• Cost of Error Gap: Mistakes in digital AI might cost a retry, but in the physical world, they can be measured in human lives. There's no "undo" button.
• Latency Gap: Milliseconds matter for physical agents like cars traveling at high speeds, demanding on-board compute that makes split-second decisions.
• Data Gap: Unlike digital AI, which benefits from the internet's vast cache of human knowledge, the physical world lacks a readily digitized, pre-labeled equivalent.
• Validation Gap: Physical AI requires a very high level of safety and confidence from day one, before any units are deployed, unlike digital AI where iteration can happen post-launch.

Navigating the Hype Cycles

Dolgov also touched upon the cyclical nature of AI hype, noting that each technological breakthrough, from deep learning to transformers, makes initial demos easier but doesn't necessarily solve the hard problems inherent in scaling to a real product. He advised founders to remain honest about their product's demands and avoid cutting corners, as this can lead to a "pretty rude awakening later."

The Waymo Driver's Scale and Safety

He shared Waymo's current operational scale: approximately 500,000 trips per week, driving over 4 million fully autonomous miles weekly across 15 cities in the United States. This translates to over 300 years of an average American driver's annual mileage each week. Crucially, the Waymo Driver achieves this with a "superhuman safety record," preventing a serious injury every eight days by preventing potential crashes.

The Importance of Architecture and Iteration

Dolgov emphasized that the required "number of nines" dictates the architecture and technical approach. He used the example of sensing for autonomous driving, explaining that while cameras alone might suffice for assistance products, full autonomy demands multiple sensing modalities like lidar and radar to overcome limitations in various conditions. He also highlighted the importance of hardware evolution, noting that companies should design for future upgrades rather than anchoring to current component costs.

Furthermore, Dolgov stressed the need to "ride those tech waves and do that repeatedly," rebuilding the Waymo driver around AI breakthroughs while focusing on unification and simplification. The company has successfully integrated technologies like convolutional neural networks (CNNs) and transformers, and is now leveraging the latest in vision-language models (VLMs) and frontier world models. He noted that driving, with its social and sequential aspects, shares similarities with language modeling.

The Waymo Foundation Model and the Future of Physical AI

The presentation culminated in a discussion of Waymo's foundational model, a multimodal world action language model. This model processes camera, lidar, and radar data, understands world physics and social dynamics, and can predict the effects of its own actions. Dolgov stated that this approach allows for a more agile development process by moving complexity upstream to a large, shared foundation.

Ultimately, Dolgov's message was one of pragmatic optimism: while the magic of early-stage development is crucial, the real challenge and reward lie in building reliable, safe, and scalable AI solutions for the physical world. He concluded by stating that physical AI is where digital AI was years ago, and the next decade will likely see its major advancements in the physical realm.