OpenAI, Large-Scale Model Autonomously Trains Smaller Model

OpenAI's large-scale model 'GPT-5.6 Sol' autonomously performed fine-tuning (additional training) of the smaller model 'Luna'. With only minimal human intervention through rough instructions ("fairly under-specified prompt"), Sol independently trained Luna, according to OpenAI's explanation.

The concept of AI training another smaller AI has previously been discussed at the research level. However, this instance occurred in the context of actual product development and was executed without direct human involvement—a development that represents crossing a new threshold. This is a concrete example demonstrating that AI has begun actively participating in its own improvement processes.

OpenAI uses an evaluation metric internally called the 'RSI benchmark'. RSI stands for 'Recursive Self-Improvement' and measures how effectively AI can improve itself or other AIs. On this benchmark, Sol achieved a score 16.2 points higher than its predecessor model GPT-5.5, according to OpenAI's disclosure.

The background of this initiative lies in OpenAI's positioning of 'automated researcher' as a near-term goal. An automated researcher refers to an entity where AI autonomously performs hypothesis validation, experimentation, and model improvement in place of human researchers. OpenAI states this vision is "within reach," and the Sol-driven Luna training can be viewed as a concrete step toward that goal.

If the process of AI training other AI becomes automated, the pace of research and development could change significantly. Previously, human researchers needed to repeatedly design, evaluate, and improve models, but if AI itself can assume some of these roles, development cycles are expected to shorten. On the other hand, there are views that autonomous modification of models by AI entails challenges such as unintended model degradation and difficulties in quality control.

Going forward, attention should focus on how far such autonomous training will be permitted without human oversight. OpenAI is known for prioritizing AI safety assessment, and how it establishes frameworks to manage and monitor automated research processes will be the next focal point. It is necessary to observe how technological advancement and safety management systems progress in parallel.