Unpredictability of model development

Why PMs are suddenly valuable again | Tulsee Doshi (Head of Product, Gemini)

Topic 2 of 18

Published Nov 8, 2026 · Clip 2:03–5:30 (3:26)

Tulsi explains what surprises people about how AI models are made. From a PM perspective, you set goals and evals, but models are not predictable—she compares it to Snape's potions class: you put in the ingredients but must expect a bit of magic. The PM's job is not just shaping inputs but discovering where the model shines and turning that into amazing user experiences. She adds that researchers monitor runs and track metrics to make it as much of a science as possible, but PMs must also understand the 'vibes' of the model—how it communicates, its persona, and how capabilities shine through downstream in products.