AI Now Builds Itself: The Race Accelerates

Summary

Artificial intelligence is now literally building itself, a development that Anthropic warns society is unprepared for and suggests a slowdown in development. The trend shows AI systems increasingly developing other AI systems, pointing towards autonomous design of successors. While recursive self-improvement isn't inevitable, a crucial ingredient is still missing. Evidence shows AI's capability in completing tasks is accelerating rapidly. For example, Claude Opus three could complete tasks taking humans about four minutes in March of two thousand twenty-four. By April of two thousand twenty-five, Sonnet three point seven handled tasks that took humans about one and a half hours. And recently, Opus four point six managed twelve-hour tasks. If this continues, AI could handle week-long human tasks by two thousand twenty-seven. AI models now also successfully reproduce AI research paper novel ideas nearly one hundred percent of the time. Within Anthropic, over eighty percent of code merged into their codebase as of May two thousand twenty-six was authored by Claude, a significant jump from low single digits before its launch. However, this raises questions about code quality, with some believing Claude's code is less valuable than human-written code, though still contributing to overall productivity. Humans remain crucial for setting research directions, judging results, and verifying outcomes. The potential futures range from stalling progress to AI systems achieving full recursive self-improvement, which could be determined solely by computational power and energy availability, leading to a permanent societal underclass. Anthropic suggests slowing development would be beneficial, but acknowledges the difficulty and competitive pressures involved, especially since detecting AI development is harder than tracking nuclear proliferation.