dmitri dolgov

Dolgov describes a flywheel where real-world deployment generates data → grounds simulator → simulator generates hard cases for critic → critic scores and improves agent → smarter agent deploys → more data. The shared foundation model across all three enables this. Crucially, eval and metrics are the strategic moat — 'build your eval before you build your technology' — and Waymo's safety readiness framework (evaluating every component from physical to operational) backed by 220M+ miles of public safety data creates a trust advantage harder to replicate than models or algorithms.

Autonomous vehicles have moved from demo to product at scale — Waymo's 220M+ miles, 17x safety advantage, and 4M weekly miles demonstrate the technology is commercially viable and entering exponential growth phase.

Deploying AI in the physical world faces four structural gaps vs digital AI: irreversible cost of errors (human lives), millisecond latency constraints, no internet-scale pre-labeled data, and need for high confidence before first deployment — creating a high barrier to entry and long development cycles.

Waymo uses a large off-board foundation model specialized into three 'teachers' (Driver, Simulator, Critic) which are then distilled into smaller on-board models for real-time inference — this architecture leverages scaling laws from digital AI (VLMs, transformers) to achieve zero-shot generalization to new cities and weather, dramatically accelerating global deployment.

Dolgov argues that achieving full autonomy requires climbing an exponential ladder of reliability 'nines' where each additional nine takes 10x more effort, and that every AI breakthrough makes demos easier but barely moves the needle on the long tail — explaining why hype cycles produce spectacular demos but few real products.

Dolgov argues that camera-only sensing hits a performance ceiling far below superhuman levels required for full autonomy, and that fusing cameras, lidar, and radar provides complementary physics (lidar works in darkness/glare, radar penetrates weather, cameras give color/resolution) enabling early detection of edge cases like pedestrians in dust storms or at night.

Waymo has achieved superhuman safety (13x fewer serious injury crashes) and is now scaling exponentially from 16 years to 100M miles to 6 months for the next 100M, deploying in 11 US cities plus London and Tokyo, proving autonomous vehicle technology has crossed the chasm to commercial viability.

Dolgov describes Waymo's foundation model as a multimodal (camera/lidar/radar), world model (physics + social semantics), action model (understands agent's effects), language-aligned (unlocks VLM knowledge) system with a fast-path (millisecond geometric reactions) and slow-path (semantic reasoning) — enabling deployment across vehicle platforms and future products (trucking, personal vehicles).