Koray Kavukcuoglu, VP of Research at Google DeepMind, discusses the journey toward Artificial General Intelligence (AGI) through the development of Gemini models. He argues that AGI is not a single benchmark but a gradual process of building increasingly useful and trustworthy systems, with products serving as essential stepping stones. The conversation covers the evolution from specialized projects to a unified model approach, the importance of human data, and the critical role of safety in deploying AI to billions of users.
Summarized by Podsumo
Koray argues that AGI will not be achieved through a single benchmark but by building systems that people find increasingly useful and trustworthy in their daily lives.
Products are framed as essential stepping stones toward AGI, providing real-world feedback that guides development beyond what any benchmark can offer.
Gemini is described as a constantly evolving architecture, not a static model, with each generation being 'custom built by hand' and incorporating new research from a continuous pipeline.
The discussion covers the importance of chain-of-thought reasoning in natural language for safety, allowing both human monitoring and inter-agent communication to be transparent and auditable.
Koray expresses confidence that recursive self-improvement (RSI) is possible, though it is not yet a fully automated loop but a growing set of capabilities where models assist in their own development.
Safety is treated as a foundational requirement, with Koray emphasizing that safety and development must go hand in hand, and that releasing frontier models now involves extensive testing with governments and external experts.
"AGI is something that if people find it useful, if you trust it, if you can work with it, the more you trust it to make decisions for you or to make decisions with you, then it will start feeling like AGI."
— Koray Kavukcuoglu
"If I can solve all these hard problems, I gotta be able to book a restaurant. Surely."
— Koray Kavukcuoglu
"We don't have a test for AGI. We'll never have a test for AGI. It's not like we are going to have a benchmark, and then you solve that benchmark, and then you say 'this is AGI or not'."
— Koray Kavukcuoglu