Thinking Deeper in Gemini - Test-Time Compute and Model Architecture

Explore how Google DeepMind approaches fundamental intelligence bottlenecks in AI models through this 18-minute conference talk from the AI Engineer World's Fair. Learn about the development of "Thinking in Gemini" as a solution to test-time compute limitations, examining how this approach represents progress toward general intelligence by improving model architecture and training objectives. Discover recent advances in Thinking capabilities that enhance both model performance and steerability, while gaining insights into the future direction of Gemini models. Understand the technical perspective on identifying and solving core intelligence challenges from Jack Rae, who leads Gemini Thinking and co-leads Gemini Pre-training at Google DeepMind.