By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 5:04
Riding Fully Autonomously in San Francisco
Steph Smith and Waymo's Chief Product Officer Saswat Panigrahi preview the ride and episode topics. They then get into the car and discuss how smooth it feels and how special it is even for someone who has worked on it since 2016.
- •Teaser covers levels of autonomy, the LiDAR vs video debate, regulation, user experience and AI
- •Riders can play music, ask the car to pull over, or call support
- •Smoothness of the drive builds instant trust
- •Saswat first rode with a non-controlling person in front in 2017 and with truly nobody in 2019
- •Even after years of preparation, being inside it is still special to him
- 5:04 – 7:10
The 5 Levels of Autonomy
Saswat explains that Waymo operates at level 4, with no human expected to take over, within a defined scope. He contrasts this with levels 2 and 3 and level 5, which means anywhere, anytime.
- •Level 4 means nobody in the front seat and no expectation of human takeover
- •Waymo wouldn't operate in heavy snow, so its scope is limited
- •Level 5 is defined as anywhere, anytime
- •Saswat says the gap between level 3 and level 4 is like driving versus flying
- •If a rider interferes with the car, it pulls over
- 7:10 – 9:41
Barriers to Level 4 and Building the Full Stack
Asked what delayed level 4, Saswat points to the clarity of the societal prize and says the largest challenge was technology. That meant building the driver, measuring it, and doing it at scale across very different conditions, which required Waymo to build hardware, software and simulation itself.
- •Waymo was convinced a fully autonomous driver would be safer because humans get drowsy and have physical limits
- •Building and measuring the driver at scale is the hard part
- •Phoenix's faster, wider streets pose different challenges than dense SF streets
- •Waymo builds lasers, cameras, radars, software and simulation to iterate quickly
- 9:41 – 14:39
How the Car Perceives, Predicts and Decides
Saswat breaks the system into three parts: perceiving the surroundings, anticipating what others will do, and deciding what to do. Live examples during the ride illustrate pedestrians, jaywalkers, and on-screen feedback for riders.
- •Lasers, cameras and radars each contribute strengths, and the car sees three football fields in 360 degrees
- •Anticipating pedestrians involves reading gait and movement, which takes heavy machine learning
- •The challenge was being confident rather than overly conservative
- •The screen shows riders what the car sees and why it stops
- 14:39 – 19:02
LiDAR vs Video and Build vs Buy
Asked about the rivalry between LiDAR and camera-based approaches, Saswat calls the debate ideological and says sensors are tools with different strengths. He argues cost is a practical question that is falling, and explains why Waymo built its own hardware.
- •Each sensor type has strengths: LiDAR at night, cameras for red versus green lights, radar sensing objects coming in
- •Combining sensors positions the system better than any single one
- •LiDAR costs have come down sharply in the last two years, which Saswat describes as proof of an earlier belief
- •The best off-the-shelf LiDAR and radar weren't optimized for autonomous driving
- •Build versus buy is re-evaluated each hardware generation
- 19:02 – 21:45
Where the Moat Is and Where Value Accrues
The host asks what differentiates companies once autonomy is widespread. Saswat describes a vast market of a trillion miles and a knowledge flywheel from real-world and simulated miles.
- •The addressable space covers all cars, trucks and transportation
- •There are fewer autonomy players than three years ago
- •20 million-plus testing miles and billions of simulated miles reveal problems others haven't found
- •Hardware design has long lead times, and Waymo's own laser decision came after 10 million miles
- 21:45 – 26:58
Technological Unlocks and the Role of AI
Saswat points to a driver that generalizes across cities and better simulation as key unlocks. He also describes the mix of general and specialized deep models used across the stack, and Google's infrastructure as another moat.
- •The driver worked well from the start in Los Angeles and Scottsdale opened within two months
- •Good-weather cities are a mix of SF-style and Phoenix-style driving
- •Simulation could be extended to rain and combinatorial variations
- •Many models are used, general and specialized, with AI at every layer
- •Google's infrastructure and early machine learning investment are described as a moat
- 26:58 – 34:06
How Waymo Thinks About Safety
The host asks how regulators, technologists and riders each view safety. Saswat lays out layers: quantitative collision data, risky behavior, whether riders feel safe, and not overpromising capabilities. He cites Waymo's results, including no injury collisions in 1 million autonomous miles.
- •Safety includes crash statistics, avoiding risky behavior, rider feeling of safety and honesty about capabilities
- •Microphones let the car detect and locate sirens
- •In East Valley Phoenix, Waymo re-simulated every fatal crash and showed it could avoid them
- •In 1 million fully autonomous miles, there was no collision with an injury
- •Alerts like cyclist approaching help protect riders opening doors
- 34:06 – 39:52
Working with Regulators and Transparency
The host asks about pushback and pushing regulation forward. Saswat says Waymo's mission of safer driving aligns with regulators' goals, and that transparently sharing data builds constructive conversations.
- •Conversations with government, firefighters and law enforcement start from shared safety goals
- •Waymo publishes accident reconstruction and shares every contact event
- •Saswat says about to cross 2 million autonomous miles, equal to over 160 years of human driving
- •Internal debates come from different angles but share the safety mission
- 39:52 – 46:59
Designing the Rider Experience
Saswat explains how Waymo tuned the in-car screen after experimentation and how rider feedback shapes pickup and drop-off behavior. The conversation also touches on the steering wheel, defensive driving, and reassuring assertiveness.
- •The screen shows less detail than early versions and explains stops, with a dark mode at night
- •Feedback from many riders is a flywheel, with nearly 100 thousand rides a month
- •Pickup side depends on context, such as a narrow residential street versus a busy one
- •Steering wheel removal is blocked by regulation, and the next vehicle was designed with partners
- •The car anticipates reckless behavior like a red-light runner and drives defensively
- 46:59 – 51:19
Choosing Markets and Early Testing
Saswat explains how Waymo picked cities, from testing in 20 diverse cities to committing to launch markets like SF, LA and Phoenix. He also shares how it felt to first test the cars and how regulators have responded.
- •Early testing in 20 cities, including Miami, Death Valley and Tahoe, built a generalizable driver
- •Launch cities were chosen for commercial scale, such as top ride-hailing markets and a busy airport
- •Saswat compares the first ride to exam day after six years of preparation
- •Phoenix East Valley shows support from passengers to the state level
- 51:19 – 56:24
Societal Impact of Autonomy
The host asks about wider effects on insurance, trucking and city design. Saswat discusses safer roads, mobility for people who can't drive, reclaiming parking space and pollution, and trucking, with live examples like construction zones.
- •Cities and life are designed around driving, including large amounts of parking
- •Saswat says 1.35 million people are killed on streets and calls it a pandemic
- •Autonomy offers mobility for people over 65 and people with visual impairments, and economic opportunity
- •Long-haul trucking faces driver shortages and mandatory fatigue breaks, and truck driving could become a local job
- •Construction zones are unique and require substantial intelligence
- 56:24 – 1:00:21
Will Humans Stop Driving? Closing and Reflections
The host asks whether human driving might become illegal, and Saswat says the aim is to offer a safe, easy option and that choice is good. They wrap with what excites him, the rider's reaction to the ride, and Steph's personal note on waiting for self-driving.
- •Saswat says it's good to have the choice
- •He is most excited by riders' first and fifth rides and what they say
- •Steph noticed the lack of a driver for only about two minutes
- •Steph got her license at 29 because she was waiting for self-driving
- •She invites comments and points to an 8-part a16z series
