Lenny's Podcast · Nov 8, 2026 · 58m · 47m saved
With Tulsee Doshi, Head of Product for Google Gemini
Google has 14 products with more than a billion users each, by Tulsee Doshi's count. Tulsee, head of product for Google's frontier models including Gemini, told Lenny's Podcast that this reach, from chips to consumer apps, is the company's real advantage in AI.
Most of the industry measures the AI race by who has the smartest model. Tulsee says Google is also building models at every point on the cost-and-quality curve, for users from search to engineers to the people defending critical infrastructure.
"The model isn't predictable. It's kind of like Snape's potions class — you can put all the ingredients in, and you should be thoughtful about what you expect, but you also have to expect that there will be a little bit of magic, and that magic is going to be surprising." — Tulsee Doshi
Tulsee spent six years leading product for Google's responsible AI team before taking over the models that now include Gemini, the Nano Banana, Lyria and Veo media models, the open Gemma suite and a set of audio models.
The full interview is covered here so you can skip it. 58 minutes of audio, 11 minutes of reading.
Here are the 11 principles that matter.
Key Takeaways
A model is not a predictable feature: the PM's job includes finding the surprising upside
Google's edge is the full stack, from hardware to 14 products with more than a billion users
The 60-20-20 rule: 60% of a feature works today, 20% works well enough, 20% is a bet on the model
Tulsee doubts anyone can predict model capability three months out
When anyone can build anything, taste is the scarce skill that makes PMs valuable again
A first eval can start with 5 prompts, and a 20-prompt mini eval "already is so powerful"
Burnout comes from churn and work people do not enjoy, not from hours alone
AI should supply the data and people the insight, or internal documents fill with slop
1. Models Are Not Predictable
Tulsee opened by pushing back on the idea that building a model is ordinary product management.
Tulsee Doshi says a PM on a frontier model has to find and shape surprises, not just define inputs
PMs set goals and target evals, the tests a model is scored on, the way they would for any feature. The model itself does not behave predictably, which is the point of her potions-class comparison above. The research side is rigorous: researchers track training runs and step in on metrics in real time. A PM also has to judge a model's "vibes," meaning how a new capability actually feels to use, and not only whether it scored well.
