Jeff Dean's First Talk Post-Google: How to Pick Research Problems Worth 5 Years
FinanceYF5 · x · 2026-08-10
In his first public appearance since leaving Google, Jeff Dean sat down with Dawn Song at Stanford to share his research methodology rather than vision:
- The '5+2' Rule for Problems: The perfect problem shape consists of 5 things the community has fuzzy work on, plus 2 things you have no idea how to solve but can imagine cracking.
- Breadth-First Reading: He advises students to skim 100 paper abstracts rather than deep-diving, aiming to connect important ideas that haven't been linked yet.
- TensorFlow's Wins and Losses: The core win was decoupling model logic from machine management. The mistake was letting external contributions pollute the core via 'contrib', causing community chaos.
- The Intuition Behind MoE: Aim for massive capacity while activating only the most useful parts, much like the human brain. Internal experiments once saw a 10x improvement in training compute/quality ratio.
Related event: Jeff Dean Shares '5+2' Research Methodology Post-Google(2 posts)→
More from Companies & People
- AI-Generated Books Capture 40% of Self-Published Sales; Kimi K3 Sees Global Surge — FinanceYF5 · 2026-08-10
- Pinterest Earnings: Open Models Cost Under 8% of Closed Alternatives — juliey4 · 2026-08-10
- Testing Alibaba's Qwen Open Platform: Practical Agents for Housing, Rentals, and Delivery — 葬AI · 2026-08-10
- Chery's Humanoid Robot Mornine Reaches 2,000 Deliveries Across 60+ Countries — CyberRobooo · 2026-08-10
- Jeff Dean Announces Discovery Loop to Automate AI, Science, and Engineering — JeffDean · 2026-08-10
- Apple Accelerates China AI with Alibaba Qwen and Baidu, Tests CXMT Memory — 量子位 · 2026-08-10