There comes a moment when a technology begins to help build itself. Anthropic has just published data suggesting this moment has already begun. In March 2024, Claude could autonomously complete tasks lasting about 4 minutes. > One year later: 1.5 hours. > Today: 12 hours. If this trend continues, tasks taking days may be within reach this year; tasks taking weeks might already be achievable by 2027. 80% of the code entering Anthropic’s codebase today was written by Claude. In 2024, that number was in the single digits. The typical engineer at the company now merges 8x more code per day than before. Not because they’re working harder. > Because Claude writes, tests, debugs, and fixes while the human reviews. Anthropic calls this recursive self-improvement: an AI helping to build its own next generation. > They say they haven’t arrived there yet. But the data published now shows the path is already being traveled. > The loop is beginning to close. > This is not a projection for the future. These are internal metrics showing what is already happening. The question that remains: If AI is already helping to build the next AI, what exactly do humans control in this cycle?