A Cheeky Pint• Mar 24, 2026• 1:02:32Interview

The 20-year journey to fully autonomous cars with Dmitri Dolgov of Waymo

Dmitri Dolgov•co-CEO, Waymo

Executive Summary

Waymo is operating at significant scale, providing approximately 500,000 fully autonomous rides per week across 11 U.S.
The company's AI strategy involves a sophisticated distillation process, training smaller, efficient on-vehicle "student" models from large, off-board "teacher" models (Driver, Simulator, Critic) that originate from a single foundation model.
Waymo is on the cusp of deploying its sixth-generation system, featuring a custom-designed vehicle and a sensor/compute hardware suite that costs a fraction of the previous generation, enabling more scalable economics.
The transition from the fourth to the fifth-generation driver marked a pivotal, discontinuous jump by making AI the core of the system, which enabled a more generalizable platform and rapid geographic expansion.

12 quotes

Concerns Raised

The immense difficulty and long-term commitment required to achieve the final 'nines' of reliability for full autonomy.
Adapting the Waymo Driver for new international cities with different driving rules and environments (e.g., London) is a non-trivial task.

Opportunities Identified

Rapid geographic expansion into new domestic and international markets like London and Tokyo.
Deployment of the cost-effective sixth-generation hardware and custom vehicle to improve unit economics and passenger experience.
Long-term potential to reshape urban landscapes by reducing the need for parking infrastructure.

Key Themes

AI Architecture: Foundation Models and Distillation

Waymo's AI development starts with a large, off-board foundation model that is specialized into three 'teacher' models: the Driver, the Simulator, and the Critic. These large models are then used in a distillation process to train smaller, more efficient 'student' models that run on the vehicle's local hardware for real-time inference.

This teacher-student architecture allows Waymo to leverage the power of massive, cloud-based models for development and training while deploying highly optimized, fast-executing models on the edge, solving a key challenge in real-world AI application.

Generational Hardware and Software Evolution

The company has progressed through distinct generations of its 'Driver' system. The jump from Gen 4 to Gen 5 was a software-led, 'discontinuous' leap to an AI-first backbone, enabling scalability. The upcoming Gen 6 introduces a custom vehicle platform and significantly lower hardware costs, comparable to a high-end laptop.

This demonstrates a strategic, iterative approach to de-risking the technology, focusing first on software generalizability (Gen 5) before investing heavily in custom, cost-optimized hardware (Gen 6), which is critical for achieving profitability at scale.

Scalability and Geographic Expansion

After proving the end-to-end service in a limited domain with its fourth-generation system, Waymo's fifth-generation AI-centric platform enabled rapid expansion. The company now operates in 11 U.S. cities and plans to launch in international markets like London and Tokyo, though this will require specific adaptation and validation.

Waymo's ability to move from a single, constrained market to multiple complex urban environments validates the generalizability of its core technology and signals a shift from R&D to a phase of aggressive commercial scaling.

The Path to Full Autonomy (L4/L5)

Co-CEO Dmitri Dolgov asserts that achieving Level 4/5 autonomy is a fundamentally different problem than enhancing Level 2/3 driver-assist systems. He argues it cannot be reached through incremental improvements from L2/L3, emphasizing the immense difficulty of achieving the final 'nines' of reliability required for a fully driverless system.

This perspective frames the autonomous vehicle landscape as having two distinct strategic paths, positioning Waymo's 'full autonomy first' approach as necessary for true driverless applications, in contrast to competitors focused on evolving consumer-grade driver-assist features.

Sensor Fusion vs. Vision-Only

Waymo maintains a multi-modal sensor strategy, utilizing cameras, LiDAR, and radar to provide 360-degree coverage. Dolgov highlights the complementary nature of these sensors, with radar being particularly crucial for robust performance in adverse weather conditions where other sensors might be degraded.

This reinforces the industry debate on sensor suites, with Waymo's stance suggesting that redundancy and diverse sensor modalities are essential for achieving the safety and reliability required for a commercial, fully autonomous service.

Get started free

Topics

Autonomous Vehicles Waymo AI Strategy Machine Learning Foundation Models Model Distillation Sensor Fusion LiDAR Radar Ride-Hailing Robotics Hardware Generations Scalability On-device Inference L4 Autonomy

Processed May 10, 2026 yt-dlp + mlx-whisper + Gemini

You're reading a preview

Get started free →