Waymo Co-CEO Dmitri Dolgov: The Demo Is Only 1% Of The Work

Predictions and Expectations:

• The speaker predicts that the next decade of AI advancement will happen in the physical world, mirroring the last decade in the digital world.
• The speaker expects that the safety benefits of AI in the physical world will grow beyond current levels.
• The speaker believes that earning trust through evidence-grade evaluation and public proof is a business advantage more difficult to replicate than models or algorithms.
• The speaker expects that betting on high-capacity models will yield better scaling laws, allowing for distillation into smaller, efficient models for real-time deployment.
• The speaker expects that structure augmenting end-to-end models will provide strong verifiable feedback signals for training and evaluation.
• The speaker expects that future hardware generations will drastically simplify and radically reduce costs, making designs anchored to current prices obsolete.
• The speaker expects that the Waymo Foundation Model will eventually power different commercial applications like trucking and personally owned vehicles.
• The speaker anticipates that every AI hype cycle produces a wave of spectacular demos but very few real products due to the difficulty of the "long tail."

Timelines and Milestones:

• The speaker notes that the project started in 2009, with the first 90% capability milestone reached in 2010.
• The speaker states it took about 18 months to reach the first demo milestone.
• The speaker states it took about 15 years total to move from the first 90% capability to providing a scalable service.
• The speaker mentions it took about 10 years from the initial 18-month demo to begin providing a service, followed by five more years to scale to half a million trips per week.
• The speaker states it took about eight years to go from the start of initial rider-only operations to serving riders in four cities.
• The speaker notes that it took about seven months to drive the next 100 million miles after the first 100 million.
• The speaker recalls that an initial goal was set to drive 100,000 miles in autonomous mode and ride 10 routes of 100 miles each, accomplished in about a year and a half.
• The speaker references that ConvNets were leveraged around 2013, and Transformers were bet on around 2017.
• The speaker mentions the speaker believes that Waymo is currently preventing a serious injury every eight days based on current scale.

Technology and Product Direction:

• The speaker plans to leverage the Waymo Foundation Model to move complexity upstream, allowing the specialization layer running on the car to be lightweight and speeding up development.
• The speaker expects to build a system that is "maximally learned and minimally constrained" by using structure to boost performance and scaling laws.
• The speaker plans to continue riding waves of technical innovation repeatedly, moving from ConvNets to Transformers to VLMs and frontier world models.
• The speaker intends to build a "structure augmented end-to-end" approach, augmenting learned embeddings with materialized structure representations.
• The speaker plans to utilize a "system one, system two" architecture within the foundation model to handle both split-second safety decisions and complex semantic reasoning.
• The speaker plans to build a highly accurate generative world model that leverages Google DeepMind's Genie 3 for controllable and realistic scenario simulation.
• The speaker expects to expand the fleet to operate in more locations, noting they launched four cities in just one day earlier this year.
• The speaker plans to use the AI ecosystem comprising the agent, the simulator, and the critic to create a flywheel that accelerates progress through shared foundation model capabilities.
• The speaker expects that building a safety and readiness framework is critical to guide development, deployment, and scaling.

Market and Industry Outlook:

• The speaker states that physical AI is currently at the stage where digital AI was a few years ago.
• The speaker expects that the cost of error in physical AI can be measured in human lives rather than tokens.
• The speaker believes that the "long tail" of rare events becomes a daily reality when driving millions of miles per week.
• The speaker anticipates that the opportunity for physical AI is "absolutely massive."
• The speaker notes that currently, someone loses their life on the road every 26 seconds globally.

Company Plans:

• The speaker plans to openly publish safety data and ongoing safety research to build trust with customers, communities, and regulators.
• The speaker expects to continue scaling exponentially, having driven over 200 million fully autonomous miles and served over 20 million fully autonomous trips.
• The speaker plans to focus on "moving fast and shipping safely" rather than "moving fast and breaking things."
• The speaker plans to invest in fully redundant systems and tiered fallback architectures to achieve the necessary reliability.
• The speaker plans to design hardware with future upgrades and commoditization in mind, avoiding anchoring to today's component prices.
• The speaker intends to guide the development flywheel with rigorously defined metrics and evaluations.

Financial Guidance:

• The speaker mentions that with every generation of the Waymo driver, the hardware cost has been drastically simplified and reduced.
• The speaker notes that betting a company on today's hardware prices is a bet on a number with a "fairly short shelf life."
• The speaker suggests that the difficulty of adding reliability follows an exponential ladder, where every additional nine requires about 10 times more effort.

Risks and Caveats:

• The speaker warns that spending on demos rather than the necessary "nines" of reliability will lead to a "rude awakening" and failure to create real products.
• The speaker cautions that common failure modes include picking technology that flattens out before reaching the required performance level.
• The speaker notes that a single leaf or branch obstructing a sensor without redundancy can bring a robot to a full stop.
• The speaker warns that projects can hit a "dead end" if a successful technical effort is not integrated into the broader product system.
• The speaker states that relying solely on a basic vanilla end-to-end model is insufficient for safety-critical, superhuman performance in the physical world.
• The speaker emphasizes that evaluation and metrics are strategic moats, and without them, the company is "flying blind."
• The speaker fears that without a rigorous framework, scaling in the physical world cannot be done in a responsible manner.

Confidence and Disagreement:

• The speaker expresses high confidence that "structure that channels scale always wins" over structure that fights scale.
• The speaker is certain that the "bitter lesson" holds that methods leveraging massive compute and data will always beat handcrafted human knowledge.
• The speaker is confident that the Waymo driver currently has a superhuman safety record, being 17 times better than human drivers in preventing serious injury crashes.
• The speaker believes that "moving fast and shipping safely" is a much more difficult thing to do than moving fast and breaking things.
• The speaker is sure that the best AI moments in the physical world will look like "nothing happened," where the task is done safely and smoothly without notice.