Koray Kavukcuoglu on frontier models, coding agents, and building AGI

The pursuit of AGI centers on transitioning from static models to agentic workflows capable of complex software engineering and multi-domain problem-solving. Maintaining a position at the technological frontier requires leveraging Google’s full-stack resources and a long-term commitment to scientific research. Success hinges on a "co-building" strategy, where continuous user feedback loops and real-world application guide the development of increasingly general and reliable systems. This trajectory builds upon foundational breakthroughs in deep learning and reinforcement learning, evolving from early experiments like Atari to current, highly ambitious Gemini pre-training runs. By prioritizing parallel research tracks and integrating diverse technical insights, the focus remains on creating intelligent entities that function as effective, trustworthy partners across the spectrum of human activity.