Google has allegedly achieved recursive self-improvement, researcher says
The author distinguishes weak and strong RSI and notes that OpenAI's Tibo said Astra is already helping ship the next generation of models.
Google has (allegedly) achieved RSI. It’s a bigger deal than another model release. RSI = recursive self-improvement. The idea is an AI system helping build, train, or improve the next version of itself, and that loop repeating so each generation accelerates the one after it. There is a difference between the weak version and the wild version. Weak RSI is already normal: AI writing code for AI infra, automated architecture search, Google’s own AlphaEvolve finding faster algorithms. Humans still review everything. A stronger version is also happening at the big labs right now, just with less public detail: Anthropic has said Claude writes a large chunk of its own new code, Tibo said Astra helps shipping next things at OpenAI. Still human-gated at important decisions. If this is true, model cycles that used to take years could take weeks. Google will be so back!