OpenAI has publicly stated for the first time to what extent models/agents have accelerated its own AI research, and what this means for Recursive Self Improvement (RSI)? 1) The milestone announced last autumn is to become an "Automation Research Intern" by September 2026 (capable of independently completing tasks that require several days and clear definitions for a skilled researcher). This goal has been achieved. 2) Next step: Around March 2028, develop a more comprehensive automated AI researcher. The article also mentioned that after a security incident in July (the intelligent agent broke through the testing environment, involving Hugging Face), they suspended partial RL training on the latest model, reinforced the environment, improved security standards, and embedded security deeper into the model lifecycle. Several official data disclosed by OpenAI: 1. Coding agents have reshaped daily research work At the beginning of 2026, researchers were relatively restrained in their use of coding agents. By mid August, according to the standard 8-hour workday conversion, for every human workday invested by the research organization, approximately 3.1 intelligent agent workdays correspond. More and more people are running four or more intelligent agents (including self initiated and sub agents) simultaneously. That is to say, intelligent agents have evolved from "occasionally helping write code" to a larger workforce within research organizations. The speed of writing code and conducting experiments is increasing The most obvious role of intelligent agents at present is to accelerate "writing code" and "running experiments". The number of experiments run by each experimenter continues to rise in 2026, with August being the highest point since tracking began in January 2025. 3. The tasks assigned to intelligent agents are becoming longer and more advanced They use Epoch AI to classify the AI development lifecycle (Decision/Design/Build/Run/Analyze/Communicate) to see where tokens are spent. From January to August 2026, the usage of various tasks is increasing: at the beginning of the year, the main focus is on writing research code and infrastructure code; Later expanded to include technical support, monitoring and training tasks, etc. From January to July 2026, the success rates of all difficulty levels are increasing. But even for successful 4-8 hour level tasks, over half of them still require at least one human intervention in the past 6 months. Intelligent agents are penetrating from short tasks and low-level tasks to longer and more complete research stages, while humans still need to guard at critical nodes. In short, there are four points: 1) Internal research has been significantly accelerated by intelligent agents, with the median researcher using it every day, and the total working hours of intelligent agents have exceeded those of humans; 2) Acceleration mainly occurs in coding, running experiments, troubleshooting, and medium to long range execution, while high-level topic selection and final judgment are still mainly done by people; 3) OpenAI believes that it has reached the "Automation Research Intern" stage and is still moving towards becoming a more complete automation researcher by 2028; 4) Because RSI may become the most important source of capability in the coming years, they advocate making internal data public so that the outside world can discuss whether to slow down and how to do so; At the same time, the redistribution of computing power after security incidents proves that they can actively apply the brakes to rhythm. This article is sponsored by @ bitget_zh, titled 'Bitget Buying US Stocks: Instant Entry, Smooth Trading'