Descriptions:
Yi Tay, a leading researcher at Google DeepMind who heads the Reasoning & AGI team in Singapore, joins the Latent Space podcast for a wide-ranging conversation on frontier AI research. Having returned to GDM after 18 months at Reka, Tay describes how he transitioned from years of architecture and pretraining work into reinforcement learning — now the dominant modeling paradigm for capable reasoning systems like Deep Think, GDM’s model that achieved gold at the International Mathematical Olympiad.
The discussion covers why on-policy RL has become the central lever for improving reasoning, and Tay’s observation that fundamental research skills transfer surprisingly well across paradigms. He also shares how his own use of AI coding tools has evolved dramatically — from occasionally generating matplotlib plots, to now delegating debugging entirely without even reading error logs, a mode he describes as ‘vibe ML’ or ‘vibe training.’
The episode covers the naming of GDM’s Singapore team, the cultural moment of Amazon, Google, and Meta all signaling AGI intent through team branding, and what it felt like to return to Google infrastructure after an extended absence. Tay offers candid insider perspective on how a top AI lab approaches the reasoning frontier and why RL has become the lingua franca of model development.
📺 Source: Latent Space · Published January 23, 2026
🏷️ Format: Interview
