What it actually takes to build frontier models | Tulsee Doshi (Head of Product for Gemini)

Product management for frontier AI models requires shifting from rigid planning to embracing the inherent unpredictability of model behavior. PMs must focus on understanding model "vibes" and conducting iterative stress tests, as these systems are grown rather than traditionally built. A 60-20-20 framework—balancing functional MVPs with near-term and experimental features—enables teams to adapt as model capabilities rapidly evolve. As technical barriers to building lower, the PM’s primary value shifts toward exercising taste, driving consensus, and identifying high-impact problems. Responsible AI development necessitates proactive monitoring and leveraging models to harden infrastructure against cyber threats. Success in this field depends on cultivating a culture of continuous evaluation, where PMs actively integrate AI into their workflows to automate routine tasks, thereby creating space for creative, high-level problem solving and strategic decision-making.