Product lessons from Waymo | Shweta Shrivastava (Waymo, Amazon, Cisco)

The episode opens with a cold-open remark from Shweta Shrivastava about challenging your own assumptions: a seasoned product leader who is not doing enough of this may not be listening well. The point is returned to later in the interview, where it is developed as a listening skill.

If there's no conflict, if there's no contention, then something is missing.

Lenny's introduction presents Shrivastava as senior director of product management at Waymo, whose self-driving cars are already on the streets of San Francisco, LA and Phoenix. It lists her earlier roles: chief product officer at Nauto, head of product management for database and analytics services at Amazon Web Services, and earlier work at Cisco. It previews the themes of the conversation: how PM work at Waymo differs from and resembles software-only products, how progress and goals are tracked, how subtle cues build trust for riders and other road users, and lessons about building products and teams. The introduction also contains two sponsor messages, for Vanta (security compliance) and public.com (treasury accounts), which are part of the broadcast but not part of the interview's substance.

Asked what she is responsible for, Shrivastava names three areas. The first is a large part of the software that runs on board the fully autonomous vehicle and determines its behavior and trajectory. The second is the simulation tools and technologies needed to validate the performance of the system. The third is a team focused on commercially scaling the ride-hailing business, which she describes as one of the key go-to-market applications of the technology.

Lenny describes a ride he took in San Francisco on a rainy day, with no driver in the front. He compares the app to Google Maps, except that a car arrives and takes you there, and you can change course along the way. He notes how quickly the experience came to feel normal. Shrivastava replies that the rainy-day ride was a bonus, because the technology has been performing very well, which has been heartening for the team to see. This is a single rider's account and an executive's response, not a measured result.

Lenny asks about something he noticed during the ride: while trying to move into a lane with moving traffic, the car inched out, signaling an intent to merge without any eye contact. He asks how Waymo thinks about building this kind of body language into the car, and says he does not know whether she works on that piece.

Shrivastava explains that Waymo uses a lot of human driving data to train its deep models, so the car does not seem robotic or unnatural. Because the system was built to be fully autonomous from the start, she says it is important to build familiarity and trust so riders are not daunted by the technology. The target is a balance: safer than human driving, without feeling unnatural.

It has to feel very human-like, but in a good way. Making it safer than human driving, but then not making it feel unnatural.

She describes models that understand the intent of other road users, using cues such as which way a pedestrian is looking, their body orientation, road signs, and gestures by someone trying to stop the vehicle. The training data is filtered, so that mimicry is limited to good driving.

We discard the bad human driving data.