Logan Kilpatrick of Google DeepMind discusses the evolution of AI from language models to agentic systems, highlighting the “anti-gravity” harness that enables autonomous, task-performing AI across Google’s products, particularly emphasizing the impact of coding agents in accelerating software development. He also explores the concept of AI models internalizing their supporting infrastructure, the collaborative culture at DeepMind, and future prospects like unified world models and generative media that aim to democratize software creation and enhance content authenticity.
In this insightful conversation, Logan Kilpatrick of Google DeepMind discusses the evolving landscape of AI, particularly focusing on the emergence of agentic AI and its integration across Google’s products. He explains that the new “agent harness,” known as anti-gravity, serves as a foundational layer connecting various Google services, enabling them to act autonomously on behalf of users. This shift marks a transition from simply embedding language models like Gemini into products to building agentic systems that can perform complex tasks, such as coding or managing user interactions, thereby enhancing user outcomes rather than just increasing engagement time.
Kilpatrick highlights the rapid progress in coding agents, describing them as a form of narrow superintelligence that significantly accelerates software development. He notes that while other agentic AI applications are still in early stages, coding agents have demonstrated substantial impact, enabling developers to build complex applications faster and more efficiently. This progress is supported by Google’s internal use of the anti-gravity harness and the Gemini API, which together power a broad ecosystem of AI-driven tools and products. Despite competition and shifting narratives in the AI ecosystem, Kilpatrick expresses confidence in Google’s advancements and the dedicated team driving these innovations.
The discussion also touches on the concept of “the model eating the harness,” where the AI model gradually internalizes functionalities previously handled by external scaffolding or agent frameworks. Kilpatrick explains that as models become more capable, they absorb more of the surrounding infrastructure, reducing the need for separate harnesses. However, he acknowledges that some external tools will remain valuable for flexibility and specialization. This evolution challenges independent companies to find niches where focused expertise and domain knowledge can still create significant value despite the increasing generality of AI models.
Kilpatrick shares insights into Google DeepMind’s culture, emphasizing its unique blend of deep scientific research and applied product development within the vast Google ecosystem. He praises the collaborative environment, the leadership’s visionary approach, and the mission-driven focus on solving real-world problems like disease. This culture fosters innovation while balancing the responsibility of deploying AI technologies safely and effectively at scale. Kilpatrick also reflects on his personal experience navigating Google’s communication protocols while striving to authentically share the company’s AI story with the developer community.
Finally, the conversation explores future directions such as world models and generative media, with Kilpatrick describing Google’s Omni model as a unified AI system capable of handling diverse inputs and outputs, including video editing. He expresses excitement about the potential for AI to enhance content creation without compromising authenticity, as well as the growing trend of developers using AI to build apps and games. Kilpatrick envisions a future where AI-powered tools democratize software development and creative expression, supported by ongoing advances in model capabilities and ecosystem scaffolding.
