AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

DeepMind's Gemini, On-Policy RL, IMO Gold & AI Research in Singapore with Yi Tay

· Latent Space

In this episode of Latent Space, Yi Tay discusses his return to Google DeepMind (GDM) and his involvement in the Gemini and Deep Think projects aimed at advancing reasoning and AGI. He highlights the bold decision to unify specialized systems into a single model, exemplified by the Gemini model achieving gold at the International Mathematical Olympiad (IMO). Yi Tay explains the importance of on-policy reinforcement learning (RL) over imitation learning for robust model training, drawing analogies to human learning. He also shares insights on the evolving role of AI coding assistants in research productivity. The conversation covers the challenges and progress in reasoning benchmarks, including visual and long-horizon planning tasks like Pokémon, and the limitations of current models in novel knowledge discovery. Yi Tay reflects on the architecture landscape, emphasizing the enduring relevance of transformers and attention mechanisms while acknowledging potential future paradigm shifts. He also discusses the growing importance of retrieval-augmented models (DSI) in search and recommendation systems, and the unique challenges of research in ranking and retrieval. Finally, Yi Tay talks about establishing DeepMind’s Singapore lab, the significance of geographic location for AI research, and the qualities sought in new hires, emphasizing demonstrated research taste and exceptional achievements. He closes with personal reflections on health and productivity, underscoring the interplay between physical well-being and intellectual performance.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.