Recursive
178: 与田渊栋聊 RSI：模型自进化如何到来？ clarifies the company name as Recursive Superintelligence and adds 田渊栋’s founder account of the mission. In this source, Recursive is not framed as a coding-agent company, but as an attempt to build systems that use AI to find new model architectures, training logic, optimization methods, and research workflows for AI itself.
Recursive is the RSI-oriented startup discussed in 171: 【AI季报 26Q2】从 coding 到 RSI，强者愈强的未来？. The source says the team reported early results on NanoChat Auto Research, NanoGPT Speed Run, and GPU Kernel Benchmark, and treats those results as evidence that automated research loops can improve algorithms, training speed, and hardware utilization rather than only write application code.
Recursive’s importance is not only its source-reported benchmark position. The episode uses it to argue that Auto Research and Recursive Self-Improvement are not technically settled, leaving room for startup ideas around research loops, evaluation, and ML Coding even when frontier labs have more compute.
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 external comparison. Based on Recursive/RSI public writing, Liu reads that route as more coding-agent-heavy and “diligent,” while his own AI For AI route aims to be “smarter” by using Physics Of AI, OPHIS Research Workflow, and Meta-Model Training Curve Prediction to understand failures and prioritize fewer better experiments. The source treats the two paths as potentially complementary rather than mutually exclusive.
Connections
- Auto Research — immediate research-automation mechanism.
- Recursive Self-Improvement — broader self-improvement loop the source associates with Recursive’s direction.
- ML Coding, AI Coding Verification, and Research Taste — practical ingredients for useful research automation.
- AI For Science and Discovery Model — adjacent scientific-discovery branch.
- Liu Ziming, AI For AI, Physics Of AI, OPHIS Research Workflow, and Meta-Model Training Curve Prediction — episode 149’s comparison with a structure-first route.
- Tian Yuandong / 田渊栋, AI Research Feedback Compression, Mechanistic Interpretability, and Frontier Model Scaling — episode 178’s founder/operator view of RSI, scaling limits, and discovery infrastructure.