Tian Yuandong / 田渊栋

Tian Yuandong / 田渊栋 is the AI researcher and [[Recursive|Recursive Superintelligence]] co-founder interviewed in 178: 与田渊栋聊 RSI：模型自进化如何到来？. In the episode, he frames Recursive Self-Improvement as using AI to improve AI research itself, not merely as building a stronger coding agent.

His source-specific position is that AI can already compress parts of the research loop by writing code, running experiments, and returning results quickly, but that open-ended AI research still depends on Research Taste, abstraction, direction judgment, and AI Verification. He is also skeptical of treating Frontier Model Scaling as the whole story: scaling may continue to work, but resource limits and platform-like pauses make new methods and Mechanistic Interpretability important.

The episode uses Tian’s background in Google X autonomous-driving work and FAIR systems work to explain his preference for hands-on research judgment. He argues that people who only set high-level directions can underestimate implementation difficulty, while the strongest researchers often understand both what to do and how to make it work.

Connections

• [[Recursive|Recursive Superintelligence]] — company he co-founded and the main RSI case in the source.
• Recursive Self-Improvement, AI For AI, Auto Research, and AI Research Feedback Compression — technical route discussed in the interview.
• Research Taste, Problem Definition In Research, and Human Taste as AI Training Signal / 人的品味作为AI训练信号 — judgment bottlenecks he emphasizes.
• Mechanistic Interpretability, AI For Science, and Discovery Model — longer-term path from AI-for-AI toward broader discovery systems.
• Google, FAIR, and AlphaGo — background institutions and systems context referenced in the episode.
• AI Organization Design — small, hands-on, high-feedback team structure he argues is important for model research.