Got word from a DeepMind contact recently—their latest breakthrough in reinforcement learning is shaping up to be a game-changer on the path to AGI.

Demis has consistently mentioned they'd need several major breakthroughs to crack this challenge. Looks like they've nailed one.

This progress is actually reflected in their model comparisons. Their newer Flash variant outperforms Pro in quite a few scenarios now. It's the kind of iterative leap that signals real momentum in the field.