The concept of AI improving itself through recursive self-improvement is becoming a reality. Anthropic has issued a warning that this trend is advancing much faster than anticipated, based on its internal data.
On June 4, Anthropic released a report titled "When AI Builds Itself," revealing that as of May 2026, over 80% of the code integrated into its codebase is generated by Claude, its AI system. Prior to the launch of Claude's code in February 2025, this figure was in the single digits. The amount of code produced per engineer has increased nearly eightfold compared to the average from 2021 to 2025, and employees using the Mytos Preview reported a fourfold increase in productivity compared to when they worked without AI. Additionally, the length of tasks that AI models can autonomously complete has doubled every seven months, and recently, this acceleration has increased to doubling every four months.
Anthropic stated, "While we cannot definitively say that recursive self-improvement is imminent, if this trend continues, it could become a reality for AI systems to design and build their successors." This marks the first time a leading company in the AI sector has officially declared the onset of an era where "AI creates AI," which is expected to have significant implications for future discussions on AI safety.
OpenAI Enhances ChatGPT Memory Architecture for Free Users
On June 4, OpenAI announced a comprehensive overhaul of ChatGPT's memory architecture based on a system called "Dreaming."
Dreaming automatically analyzes past conversations to integrate and maintain user preferences, constraints, and patterns in the background, even without explicit requests for memory.
A key aspect of this upgrade is reducing the computational requirements for the Dreaming service by about 80%, allowing it to be expanded to free users. At the same time, the memory storage capacity for Plus and Pro subscribers will double. This change will initially apply to Plus and Pro users in the United States, with plans to extend it to additional countries and free Go users in the coming weeks. However, researchers noted that 96% of the existing ChatGPT memory entries were automatically generated by the system without user consent, raising concerns about potential conflicts with EU AI transparency regulations.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.