Recursive self-improvement คืออะไร ทำไม AI สร้าง AI ถึงน่ากังวลที่สุด

Recursive self-improvement (RSI) refers to an AI system's ability to enhance its own capabilities, a concept that first emerged in discussions by Dario Amodei and Jakub Pachocki in a 2026 paper. While many assume readers understand the term's implications, this article seeks to clarify what RSI truly entails and why AI creating subsequent AI models is a matter of significant concern.

The main distinction between RSI and AI aiding programmers in writing code lies in the level of autonomy and independence. Initially, AI like Claude (2021-2023) functions akin to a typical tech company - programmers write and submit code, and AI assists in refining it. Subsequently, coding agents (2025-2026) begin writing and editing code independently, handling entire files.

At this stage, multiple hours of work can be delegated to other agents. The ultimate goal (not yet achieved) involves AI being skilled enough to generate and train its own models (Phase 5).

Significant evidence supporting the notion that this stage is already underway includes internal company data and publicly available statistics. For instance, Anthropic reports that as of May 2026, over 80% of their codebase was written by Claude, a figure that was still just a unit in early 2025. Furthermore, the rate at which the general population writes code has reportedly doubled every four months, with an additional 8 times more code written per day by the second quarter of 2026.

Anthropic explicitly notes that while lines of code are a poor metric for measuring quality, the rapid increase in productivity is a clear indicator of RSI.

Interestingly, both Anthropic and OpenAI have published data related to RSI. Anthropic reveals that 80% of their code is written by Claude, while OpenAI's Jakub Pachocki hints at the potential for recursive self-improvement, expressing concern about the lack of preparedness for the rapid increase in AI intelligence. The article emphasizes that these figures, while indicative, are company-specific and should be viewed with caution, as they may not accurately reflect the broader AI landscape.