Physical AI requires moving fast while shipping safelyDmitri Dolgov says that in physical AI, breaking things is not acceptable, so the principle is to move fast and ship safely rather than move fast and break things.Dmitri DolgovFour Gaps in Physical AISource at 2:21Save viewpoint
Physical AI faces life-cost errors, extreme latency, no digitized internet, and pre-deployment safety validation.Dmitri Dolgov describes four gaps between physical and digital AI: the cost of errors is measured in lives, latency requirements are extreme, there is no digitized internet for the physical world, and safety must be validated before deployment.Dmitri DolgovFour Gaps in Physical AISource at 3:29Save viewpoint
An 18-month demo versus a 15-year scalable service with over 200 million miles.Dmitri Dolgov notes the massive difference between a one-time demo and a scalable service; Waymo's demo took 18 months, while the product took about 15 years, but now operates over 200 million autonomous miles with exponential scaling.Dmitri DolgovFrom Demo to ProductSource at 10:13Save viewpoint
Reliability follows an exponential ladder of nines; at fleet scale rare events become daily.Dmitri Dolgov explains that reliability and performance follow an exponential ladder of nines: each additional nine takes about ten times more effort, and at fleet scale, rare events become daily realities.Dmitri DolgovFrom Demo to ProductSource at 11:33Save viewpoint
Complementary camera, LiDAR, and radar with active sensors provide redundant, superhuman safety.Dmitri Dolgov says Waymo uses complementary camera, LiDAR, and radar sensing modalities, with active sensors providing redundancy and enabling detection in darkness, dust, and glare for superhuman safety.Dmitri DolgovSensing and HardwareSource at 16:04Save viewpoint
Do not anchor to today's component costsDmitri Dolgov warns founders not to anchor to today's component prices; Waymo's hardware suite has repeatedly simplified and radically reduced costs across six generations.Dmitri DolgovSensing and HardwareSource at 20:17Save viewpoint
Waymo rebuilt its stack around AI breakthroughs; production without regressions is harder than prototyping.Dmitri Dolgov describes how Waymo rebuilt its stack around successive AI breakthroughs—ConvNets, Transformers, and now VLMs—and says carrying new research into production without regressions is harder than prototyping.Dmitri DolgovRiding Technology WavesSource at 21:05Save viewpoint
Waymo's foundation model is a multimodal world-action language model using encoder-decoder and fast/slow reasoning.The Waymo Foundation model is a multimodal, world-action language model with an encoder-decoder architecture and fast and slow reasoning paths to handle both millisecond safety decisions and complex semantic understanding.Dmitri DolgovWaymo Foundation ModelSource at 24:41Save viewpoint
Structured approaches should channel scale, not fight itDmitri Dolgov invokes the bitter lesson: methods that scale with compute and data win over handcrafted knowledge, and structure in models should channel scale rather than fight it.Dmitri DolgovStructure and Scaling in AI ModelsSource at 30:00Save viewpoint
Waymo's structure-augmented models improve safety validationWaymo's structure-augmented end-to-end approach adds materialized structured representations to learned embeddings, enabling real-time safety validation and more efficient large-scale training and evaluation.Dmitri DolgovStructure and Scaling in AI ModelsSource at 34:33Save viewpoint
Waymo's simulator is a large AI model for closed-loop evaluation and training.Dmitri Dolgov says a real simulator is itself a large AI model, and Waymo has built behavioral and sensing world models to support closed-loop evaluation and training.Dmitri DolgovSimulation and Closed-Loop EvaluationSource at 37:50Save viewpoint
Generative models enable training on rare edge-case scenariosUsing generative world models, Waymo can train and evaluate in rare synthetic scenarios never seen in the real world, such as an airplane landing on a freeway or an elephant in an intersection.Dmitri DolgovSimulation and Closed-Loop EvaluationSource at 39:56Save viewpoint
Agent, simulator, and critic AIs share a foundation world model and data flywheel.Waymo builds three AIs—agent, simulator, and critic—all based on the same foundation world model; real-world data powers a flywheel that generates harder edge cases and improves the agent.Dmitri DolgovAgent-Simulator-Critic FlywheelSource at 41:21Save viewpoint
Metrics and evaluation build the strategic moatDmitri Dolgov argues that evaluation and metrics are the strategic moat, and founders should build their eval and metrics before building technology or product.Dmitri DolgovEvaluation and MetricsSource at 43:09Save viewpoint
Autonomy trust is earned through safety frameworks and dataDmitri Dolgov says trust in physical AI is earned through a safety and readiness framework and by publishing safety data; hundreds of millions of autonomous miles are more difficult to replicate than models or algorithms.Dmitri DolgovSafety and TrustSource at 43:57Save viewpoint
Physical AI mirrors digital AI's earlier stage; next decade unfolds in physical world.Dmitri Dolgov believes physical AI is where digital AI was a few years ago and predicts the next decade of AI will unfold in the physical world.Dmitri DolgovFuture of Physical AISource at 47:44Save viewpoint
