Captaining IMO Gold, Deep Think, On-Policy RL, Feeling the AGI in Singapore — Yi Tay

Rejoining GDM feels like a saved game

Tay describes returning after 1.5 years as seamless—LDAP and infrastructure unchanged—like resuming a Pokemon save file, though Brain is now part of GDM and many things have evolved.

Leading Reasoning and AGI team

He leads the new Singapore team explicitly named with 'AGI' to signal the north star of developing toward artificial general intelligence, focusing on frontier research close to the model.

Transitioning to RL research

Having spent his career on architectures and pre-training, Tay has shifted focus to reinforcement learning as the primary modeling toolset for modern language models.

On-policy training generates its own path

Modern LM RL uses on-policy learning where models generate their own outputs, receive reward signals, and train on their own trajectories—unlike SFT which mimics other models' outputs.

Imitation has limits for true capability

While imitation learning (watching tutorials) helps initially, both humans and models must eventually transition to on-policy learning through direct environmental feedback to achieve mastery.

Montessori approach to model training

On-policy RL resembles Montessori schooling: providing a safe environment for the model to discover its own path rather than copying predetermined trajectories.

End-to-end text model approach

Unlike previous AlphaGeometry systems, the team pursued a pure text-in/text-out Gemini model, believing that if models cannot solve IMO problems, they cannot achieve AGI.

Real-time competition logistics

The IMO attempt happened live with team members in Australia running inference on fresh problems (P1-P6) released across different days, requiring prepared model checkpoints rather than iterative benchmarking.

One-week intensive training sprint

While the broader effort was long-term, Tay's specific contribution involved an intensive one-week period preparing the final model checkpoint used for the live competition.

Maintain high learning rates

When prior assumptions are violated by new evidence, researchers should update 20-50% rather than 2%, completely invalidating worldviews when counter-examples emerge rather than being Bayesian prisoners.

AI crossing immersion thresholds

Current capabilities like coding and image generation have crossed into practical utility where models can parse spreadsheets from screenshots and generate matplotlib plots, moving beyond toy applications.