Why PMs are suddenly valuable again | Tulsee Doshi (Head of Product, Gemini)
Tulsee Doshi argues that frontier-model product work means shaping unpredictable models, and that PMs matter again for taste, evals, vibes and adaptability, while products should be built for where models are heading.
Watch on YouTube ↗Charlie's summary
- ▶ 2:29
Model PMs are in Snape's potions class: you set the ingredients, but the magic that comes out is unpredictable
Tulsee says PMs should set goals and a clear view, but the model isn't predictable. Researchers monitor training runs and intervene, yet capabilities still surprise. The PM's job is to look at what comes out, find where the model shines in unexpected ways, and shape that into something users love. Understanding the model's vibes in behavior and persona matters as much as the target evals.
- ▶ 7:04
Google is playing a full-stack game, and Tulsee says it still wants to win the frontier
Tulsee describes Google's strategy as hardware to models to products used by billions, with models spread across a cost-to-quality curve for search users, engineers and cyber defenders. When Lenny asks whether Google is just winning other ways, she says Google does want to win the frontier conversation, and that Gemini 4 is meant to be at the frontier.
- ▶ 11:10
Safety confidence comes from monitoring and from using AI to harden systems
Tulsee says chain-of-thought monitors can spot a model misaligned with its directed actions and trigger intervention, which gives more confidence in deploying. She adds that the debate underplays AI as a defender: Gemini 4 works with cyber defenders and Wiz, whose Scan for Good program found and patched a hospital system vulnerability. Her optimism comes from people being intentional and worried.
- ▶ 17:17
The model itself has a UX, and PMs can shape it through specs, rubrics and reward modeling
Tulsee argues how transparent a model is, how it communicates and what questions it asks back are all UX. Her team works with a model behavior team to define what good looks like for Gemini as a spec. That spec becomes rubrics to evaluate against, then training data and reward modeling.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.