The obvious reading is that PMs are valuable again because building got cheap. Tulsee Doshi, head of product for Google's frontier models, made that case on Lenny's Podcast, and I don't buy the victory lap.
Doshi said model work is not predictable. She compared it to Snape's potions class, where you add the ingredients and still get magic out. Her hiring test is whether PMs understand the vibes of the models, play with them every day, have an opinion, and can turn surprise capabilities into behavior specs, rubrics, evals and reward data with researchers.
That sounds lovely until you ask what shipped.
Doshi claimed Google does want to win what she called the frontier conversation and pointed to Gemini 4 Argon as proof, backed by a full stack story from hardware to what she estimated as 14 products over a billion users. Yet in the same conversation she conceded Google has not launched the obvious assistant that uses Gmail, Calendar and Docs, saying the team is working on amazing things but there is more work to do. Her 602020 rule only widens the gap, since it tells teams to build features that do not fully work today and trust the next model to catch up.
To be fair, she stated the product problem clearly. Doshi said mass products like AI Mode need simple UX while power tools like Anti-gravity can demand slash commands and terminal skill, and she argued model personality, transparency and follow-up questions are UX too. That part is real, and Google has shipped pieces around it, including Nano Banana 2.1 with a stable API model and lower image-output pricing, per Kingy AI.
Safety was the same pattern, process over proof. Doshi, who led responsible AI product for six years, said confidence comes from monitors that watch chain of thought for misalignment and from using Gemini 4 with Wiz on its Scan for Good work, including a hospital vulnerability patch. No eval numbers, no incident rates, no launch date. If you are a founder, do not hire a vibe curator. Hire someone who will kill your favorite demo.
