Podcast Atlas
concept Updated 2026-08-09 Topics: Technology

Auto Research

从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青 adds 孟繁青’s data-and-environment route into Auto Research. He says coding remains the most important current AI-lab direction because it is a foundation for model work, but as coding pipelines stabilize, frontier labs are moving toward Auto Research, AI For Science, and RSI Data-like traces where models help build or improve future training loops.

178: 与田渊栋聊 RSI：模型自进化如何到来？ adds 田渊栋’s boundary between research acceleration and full self-improvement. In this source, AI can already compress the loop from idea to experiment through AI Research Feedback Compression, but this is still not enough for open-ended Recursive Self-Improvement unless the system can select worthwhile research directions, verify gains, and turn them into future model capability.

Auto Research is the research-automation layer defined in 171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？ as AI acting like a researcher: reading papers, forming hypotheses, writing code, running experiments, and analyzing results. It overlaps with ML Coding, Deep Research, and Discovery Model, but the episode uses it specifically as the step before Recursive Self-Improvement.

The distinction matters. Auto Research can make human researchers faster without proving that the system improves itself. Recursive Self-Improvement requires the research loop to improve the next round of model, data, training recipe, benchmark, verifier, or agent system.

149. 亲历中美 New Labs 资本狂潮，和清华刘子鸣聊：AI for AI、机制可解释性和 Max Tegmark adds Liu Ziming’s AI For AI route. He agrees that AI for coding points toward AI for research, but argues that research lacks code’s ready-made structure and public corpus. His answer is to structure research through OPHIS Research Workflow, collect process-level reasoning data, and use Physics Of AI plus Meta-Model Training Curve Prediction to make research automation more selective rather than only more tireless.

Key Claims

  • Auto Research needs long-horizon planning, paper reading, experiment coding, execution, analysis, and iteration.
  • Code is the first strong substrate because experiments, data pipelines, benchmark harnesses, and evaluation scripts are executable and reviewable.
  • The bottlenecks shift toward AI Coding Verification, AI Verification, Research Taste, compute allocation, and whether the task is worth optimizing.
  • Auto Research can contribute to AI For Science even before it becomes full RSI.
  • Startup and frontier-lab claims in this area should be treated as loop-quality claims, not just benchmark-score claims.
  • Research automation needs structured observations, hypotheses, interventions, failures, and next actions because final papers omit much of the reasoning a model would need to learn.
  • Tian’s source adds that execution speed can improve before research taste does, so Auto Research claims should separate faster experiments from better scientific direction.
  • Meng’s source adds that Auto Research demand can appear as data demand: model labs may need long-running traces where research-like agents generate data, train, evaluate, and revise.

Connections