en.wikipedia.org/wiki/Recursive_self-improvement
2 corrections found
Claude was writing more than 80% of its own code
Anthropic did not say Claude wrote 80% of Claude’s own code. It said Claude authored more than 80% of the code merged into Anthropic’s codebase.
Full reasoning
Anthropic’s own writeup says: “As of May 2026, more than 80% of the code we merge into Anthropic’s codebase was authored by Claude.” It immediately clarifies that this metric is about code merged into Anthropic’s production codebase, not code that constitutes Claude itself. The footnote further explains that the figure measures “the share of lines merged to production that can be attributed to Claude.”
So the article changes the meaning of the statistic. Anthropic reported a share of code merged into the company’s codebase; it did not report that Claude was writing 80% of “its own code.”
1 source
- When AI builds itself | Anthropic
Claude writes a significant proportion of Anthropic's code. As of May 2026, more than 80% of the code we merge into Anthropic's codebase was authored by Claude... Our >80% figure measures the share of lines merged to production that can be attributed to Claude.
In 2024, researchers proposed the framework "STOP" (Self-Taught OPtimiser), in which a "scaffolding" program recursively improves itself using a fixed LLM.
STOP was introduced in 2023, not first proposed in 2024. The paper’s arXiv entry shows the initial submission was October 3, 2023.
Full reasoning
The STOP paper’s primary record shows it was first submitted to arXiv on October 3, 2023. That page also notes it was later “Published as a conference paper at COLM 2024.” So 2024 is the conference-publication year, but not the year the framework was first proposed.
Because the article says researchers “proposed” STOP in 2024, it shifts the introduction of the framework by about a year. A more accurate phrasing would be that STOP was introduced in 2023 and published as a conference paper in 2024.
1 source
- [2310.02304] Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation
arXiv:2310.02304 ... [Submitted on 3 Oct 2023] ... Comments: Published as a conference paper at COLM 2024.