The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement

The Last AI Built by Humans. Toward Genuine Recursive Self-Improvement. By Yi Duan and thirty-two others, from Shanghai Jiao Tong University, Theseus Labs, and eight other institutions. Published on the tenth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Recursive self-improvement, or RSI, is the ability of an AI system to turn experience and feedback into persistent changes. Those changes improve two things at once: what the system can do, and the process by which it will improve next time. The paper opens by using a measure called the Headroom-Closed Index, or HCI, to expose problems in existing LLMs, as the motivation for everything that follows.

The organising device is a development roadmap in stages of autonomy. It begins with improvement-execution autonomy, then improvement-strategy autonomy, then experience-acquisition autonomy, then environment-adaptation autonomy, and it ends at recursive meta-improvement. The paper then examines RSI across scenarios, with scientific discovery, embodied intelligence and software engineering among the examples, and highlights that each has distinct requirements and develops at a different speed.

Drawing on a range of industry practices and on preliminary empirical evidence, the authors connect RSI research with practical systems, and identify the key challenges that stand between current systems and what they call genuine RSI. The scenarios matter to the argument because the roadmap is not climbed at one pace everywhere. A field with different requirements may sit at a different rung.

What the roadmap supplies is a ladder. Treating autonomy as a sequence of stages, from executing an improvement to improving the way improvement happens, gives a way to say where a given system sits and what it would need next.