- Anthropic is seeking to pause AI development, citing that Claude AI is improving at a rate that may surpass human oversight.
- The rapid enhancement of Claude AI has raised concerns about the potential for recursive self-improvement, allowing it to design, build, and train its own successors with minimal human intervention.
Anthropic has highlighted the alarming pace at which Claude AI is advancing, stating that its continuous improvement could eventually lead to recursive self-improvement, a phase where AI autonomously designs and trains successors. According to Anthropic, the efficiency of AI systems is doubling every four months, a significant increase from earlier trends of seven months.123
In March 2024, Claude Opus 3 managed software engineering tasks that typically took humans about four minutes to complete. By 2025, Claude Sonnet 3.7 handled tasks averaging 90 minutes, while projections indicate Claude Opus 4.6 could tackle activities requiring approximately 12 hours of human effort by 2026. This rapid evolution has enabled engineers using Claude to produce eight times more code daily than in 2024, with the success rate on complex coding tasks soaring to 76% by May 2026.
Anthropic urges a coordinated effort among governments and AI organizations to implement a temporary slowdown in developing cutting-edge technologies, with the current key challenge emphasizing not merely the creation of more advanced AIs but the limitations of human oversight, review, and validation.
"We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for," a representative warned, reflecting a need for caution amidst rapid AI progress.
“Anthropic is advocating for a slowdown in AI development due to Claude's accelerating improvement, potentially reducing the need for human involvement. This shift highlights a shift from technology development to addressing oversight challenges.”
