AI that edits AI: Astra rewires Qwen, and Anthropic says self-building AI is close
The top story on Hacker News today has 2,630 points and 802 comments. A blog post shows GPT-6 Astra, an OpenAI model, reading and changing the weights of a Qwen model directly. Weights are the billions of numbers that make up a model, a bit like its brain.
According to the post, Astra's edits lifted the Qwen model's score on MATH-500, a math test, from 7.9% to 48.1%. Normally you improve a model with long training. Here, one AI rewrote another. Not everyone is convinced: Sean Goedecke argues that LLMs are not superhuman at reading and writing weights.
On the same day, TechCrunch reports that Anthropic warns it is very close to RSI capability. RSI means recursive self-improvement: AI so good at building AI that progress speeds up by itself. Last week, people built a computer and a Nintendo inside Qwen's weights with AI help. Now the AI does the editing.
AI improving AI faster than humans can check it is the scenario many safety researchers fear most. When a builder raises this alarm itself, we should ask what they see that we don't, and who outside the lab can verify it.