Jeff Dean admits TensorFlow mistakes, details new venture DiscoveryLoop
量子位 · wechat · 2026-08-20
In his first interview since leaving Google, Jeff Dean discussed his 27-year tenure and new venture. Key points: 1. Gemini: Successful multimodal strategy but initially underprioritized coding capabilities; 2. TensorFlow: Admitted mistakes in delaying Eager Execution and the confusing 'contrib' directory; 3. Leaving Google: Small teams allow extreme focus; using cloud infra to automate scientific discovery; 4. DiscoveryLoop: Aims for recursive self-improvement, enabling AI to accelerate scientific experiments like having 20 PhDs; 5. Research Philosophy: Skim many paper abstracts to find feasible problems rather than deep-reading one.
Related event: Jeff Dean Leaves Google After 27 Years to Start a New Venture(4 posts)→
More from Companies & People
- 'Claude is an incredible model' — users say Anthropic is pulling ahead of OpenAI — iruletheworldmo · 2026-10-16
- Accenture caps employee Claude spending at $29/month — LinkedInNews · 2026-10-16
- Ex-Apple engineer recounts 6 years on Siri, from QA tooling to offline Siri — jdluk87 · 2026-10-16
- Chile's C³ index: 65% of firms train staff on AI but only 8% use it to innovate — robertonappe · 2026-10-16
- Musk teases "crazy stuff" at Tesla event, needs audience to prove it's not AI — XFreeze · 2026-10-16
- Synthwave is tech's current aesthetic, from Tesla Roadster to TBPN intros — granawkins · 2026-10-16