Building Frontier AI Models: Insights from Google's Tulsee Doshi
Google's product lead discusses the complexities of developing multimodal frontier AI models.
1Defining Frontier AI Development
Tulsee Doshi, head of product for Google’s frontier models, oversees a diverse portfolio including Gemini, Nano Banana, Genie, Veo, and Lyria. She characterizes the development process not as standard engineering, but as a complex, iterative craft similar to a potions class. This approach emphasizes experimentation within high-stakes AI research environments.
2The 60-20-20 Framework
To manage innovation, Doshi utilizes a 60-20-20 framework for balancing product priorities. This strategy allocates resources between features that are currently functional, those nearing completion, and long-term research goals. By segmenting development this way, teams maintain stability while simultaneously pushing the boundaries of what multimodal AI models can achieve.
3Role of Product Managers
Product managers in frontier AI labs face unique challenges that differ from traditional software development. They must bridge the gap between abstract research and tangible user applications. Doshi highlights that the value of skilled product leadership is rising as models become more capable, requiring precise guidance for complex technical deployments.
4Multimodal Model Ecosystem
Google’s strategy involves a full-stack approach to frontier intelligence, integrating various models like Veo for video generation and Lyria for audio. These tools are designed to work in concert, forming a cohesive ecosystem. This architecture allows for more sophisticated interactions, moving beyond simple text-based inputs to richer, multimodal user experiences.
5Evaluating Model Performance
Writing effective evaluations is a critical skill for AI teams. Doshi emphasizes that defining success metrics for frontier models requires rigorous testing protocols. These evaluations ensure that as models evolve, they remain reliable and safe. Establishing these benchmarks early in the development cycle is essential for maintaining high-quality output standards.
6Engineering Intelligence Synergy
The intersection of engineering and product strategy is where frontier AI succeeds. By leveraging tools like Google AI Studio, developers can prototype and iterate faster. This synergy between raw technical power and strategic product direction is what allows Google to scale its AI capabilities across diverse services and critical infrastructure.
7Operationalizing AI Safety
Safety remains a core component of the development lifecycle. Doshi’s work involves integrating safety measures directly into the model architecture. This proactive stance helps mitigate risks associated with high-priority exposures. By treating safety as a fundamental product feature, teams can build more robust and trustworthy AI systems for global users.
8Future of AI Leadership
As frontier models continue to advance, the demand for product leaders who understand both deep learning and user needs will grow. The ability to navigate the ambiguity of AI research while delivering practical value defines the next era of tech leadership. Success depends on balancing rapid innovation with responsible, scalable deployment.