Thinking Deeper in Gemini
summary
Identifies a test‑time compute bottleneck and adds an explicit thinking stage trained with reinforcement learning so the model can emit iterative tokens and loop thousands to tens of thousands of extra steps before committing to an answer. Implements thinking budgets across Gemini 2.5 Pro and Flash, plus a high‑cost Deep Think mode that raises performance on hard benchmarks (USA Math Olympiad from near‑zero to about 50th percentile, then to ~65th with Deep Think). Shows emergent self‑testing, hypothesis rejection, tool use and parallel chains of thought.
AI-written summary · about
chapters (6)
- 0:00Introduction to thinking in Gemini
- 2:32Bottlenecks in intelligence progress
- 7:25Mechanics of thinking in models
- 10:16Developer benefits of thinking
- 13:25Future of thinking and deep research
- 17:59Closing and Q&A
related talks
56mBuilding Multimodal Apps with Gemini 2 0 and Gemini SDK
19mGemini 2.0: Beyond the Basics
21mBuilding at the Speed of Thought
18mIntroduction to Gemini APIs and AI Studio
01h 16mI'm in my Gemini Era!
02mUsing Gemini AI workflows for productivity
01h 01mBuilding Multimodal Apps with Gemini 2.0 and Gemini SDK
15mBuilding with the Gemini API and AI Studio