Google’s Cloud AI lead on the three frontiers of model capability

As the Vice President of Product at Google Cloud, Michael Gerstenhaber leads efforts on Vertex, the company's comprehensive platform for deploying enterprise-level AI solutions. His position grants him a unique perspective on how organizations deploy AI models and the challenges they face in fully realizing the potential of agentic AI. In a recent conversation, Gerstenhaber introduced a compelling framework: AI models are advancing across three critical frontiers. The first is raw intelligence, where models strive for the highest capability. The second frontier is response time, emphasizing the importance of quick answers in practical applications. Finally, the third aspect concerns cost-effectiveness, particularly whether models can be deployed affordably at a massive scale. Reflecting on his journey, Gerstenhaber shared that he has been involved in the AI field for about two years, with significant experience at Anthropic before joining Google. His focus is on empowering engineers to create applications utilizing advanced AI patterns, while the actual application development is left to clients like Shopify and Thomson Reuters. Gerstenhaber was drawn to Google for its unique vertical integration. He noted that the company possesses everything from data center infrastructure to proprietary chips and models, allowing for a seamless integration of AI capabilities. He explained that the ability to manage everything from hardware to APIs provides a competitive edge. When discussing the competitive landscape among major AI laboratories, he identified three boundaries that impact the development of AI models. For instance, models like Gemini Pro excel in producing high-quality outputs, even if it takes longer to generate them. Conversely, for real-time applications, such as customer support, latency becomes critical. Users need prompt responses, as prolonged waits can lead to frustration, regardless of the model's intelligence. The third boundary pertains to scalability. Companies like Reddit and Meta face challenges in moderating vast amounts of content while managing costs. They require highly intelligent models that can operate within tight budget constraints and adapt to unpredictable volumes of data. Gerstenhaber also pondered why agentic AI systems have not yet seen widespread adoption, despite promising demonstrations. He explained that the technology is still relatively new, and the necessary infrastructure for auditing and authorizing agent actions is lacking. This gap means that production often lags behind what the technology can achieve. He pointed out that while the software engineering sector has rapidly integrated AI due to its compatibility with existing workflows, similar advancements are needed in other fields. The implementation of human oversight in the coding process at Google, for example, minimizes risk, but such practices need to be established across various professions to facilitate broader adoption.