Author: Zhao Zhiming PhD from Queen's University, Canada (Research direction: code intelligence and world models)

If you are hearing about RSI for the first time, this article can serve as an introductory guide. We will take you through RSI from 0, starting from the earliest attempts of AI to self-improve, then moving on to the Agents today that can modify code, create training data, and participate in model development, exploring how AI learns to improve itself step by step.

In July 2026, OpenAI released GPT-5.6. For the first time, its report card included an unusually new subject: Can AI participate in improving AI itself.

OpenAI classified this ability as self-improvement. GPT-5.6 Sol scored 57.9%, while GPT-5.5 scored 41.7%, a difference of 16.2 percentage points between the two generations of models. What was tested was not how many conventional problems the models could solve but whether they could step into the AI development scene: diagnosing research systems, optimizing GPU kernel programs and training recipes, running machine learning experiments, and even helping to improve another model.

This is not an isolated piece of news. That same summer, OpenAI began referring to agents as "automated research interns." Kuaishou's AgentX has allowed agents to participate in proposing solutions, writing code, and comparing old and new versions online (A/B testing). Google stated that long-running agents contributed to the refining of the next version of the model. Motus2 introduced prediction, evaluation, and strategy updates into real robots.

These cases touch on research execution, industrial experimentation, model training, and physical actions, yet none answer the most critical question: Will the improved system thus be better at completing the next improvement?

This is precisely the question posed by recursive self-improvement (RSI). OpenAI refers to this metric as the RSI Index, used to measure "the ability towards RSI." Since complete problem statements, task weights, and aggregation formulas have not been disclosed, 57.9% can only be viewed as an internal comprehensive indicator, not to be interpreted as a task success rate or the "completion" of RSI.

Even so, these figures mark a change: a question that once mainly existed in thought experiments is beginning to be broken down into measurable developmental capabilities by leading laboratories.

Researchers took forty years to gradually hand over more improvement aspects to machines. This article first distinguishes several easily confused capabilities and then traces five instances of "boundary advancement," observing which bounded loops have started to spin, why they do not equate to an "intelligence explosion," and why how the loops close is more important than whether they close. These five advances merely outline a narrative framework of history and do not represent academically recognized technological generations. They overlap in time and continue to develop in parallel today.

What is RSI: From Self-Improvement to Recursive Gain

These news stories are all discussing “self-improvement”, but do not explain what constitutes a closed loop.

In the traditional sense, recursive self-improvement (RSI) refers to an AI system using its current capabilities to improve the cognitive mechanisms that produced those capabilities. The improved system then participates in the next round of improvements. Here, the easiest confusion comes from: being able to self-improve does not equate to being able to self-accelerate. A system can get better at solving a fixed set of problems without becoming better at designing the next training round. A company can speed up development by heavily utilizing agents but fails to create compound growth driven by the improvements themselves.

This concept is not new. In 1863, Samuel Butler envisioned machines surpassing humans through iterative evolution. In 1950, Turing proposed building a "child machine" that could be educated and modify itself. By 1965, I. J. Good noted that if "designing better machines" also requires intelligence, then a sufficiently smart machine could design stronger successors, which could continue to improve, leading to an "intelligence explosion."

The key point is not whether AI participates in the improvement, but whether the improved system becomes better at improving itself. To clarify this boundary, RSI can be broken down into four observable criteria, ranked from low to high based on difficulty of achievement.

This article adopts the following framework. It is not the only standard definition within academia, but rather a way to separate commonly mixed claims:

• Persistent Improvement: Effective changes can enter weights, memory, tools, harnesses, or training processes, rather than disappearing after a single response. Here, weights also referred to as parameters, are the group of numbers that the model continuously adjusts during training and ultimately saves. Achieving this point makes self-improvement a discussion—this article uses this term generically to indicate that output, memory, parameters, or harnesses have improved.

• Autonomous Closed Loop: The system can discover problems within defined boundaries, propose modifications, conduct experiments, evaluate results, and adopt better versions. New versions can participate in the next round of improvement within this boundary; this article calls it bounded RSI.

• Recursive Gain (also known as "ignition"): New versions score not only higher on tasks but are also better than old versions at generating the next improvement. Only upon crossing this threshold does “the improvement capability itself” begin to compound, which this article calls recursive self-acceleration.

• Robustness and Control: Gains are sustained under fixed resources and hidden evaluations, can generalize to unseen tasks, and do not come at the expense of evaluation pollution, uncontrollable system complexity, alignment degradation, or lack of auditability (alignment refers to making AI's behavior conform to human intentions and boundaries, not just raising a certain score). Only when all four standards are achieved does one get close to open RSI.

Currently, the first two criteria have appeared, the third lacks sufficient public evidence, and the fourth is far from solved. Boun...

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。