Is AI already speeding up AI research?
Follow-up to Tuesday Nov 3: AI that edits AI: Astra rewires Qwen, and Anthropic says self-building AI is closeRemember Tuesday's token, where Anthropic said AI that helps build AI is getting close? Now more voices say it is already happening. A post on Planned Obsolescence, titled "AI is dramatically accelerating AI research", got 305 points and 253 comments on Hacker News. The AI Daily Brief podcast discussed a METR report on AI speeding up AI research. Transformer calls a new RSI report "alarming". RSI means recursive self-improvement: AI improving the next AI, again and again.
Not everybody is convinced. A LessWrong comment titled "I struggle to believe that AI R&D is ingenious" got 198 points on Hacker News. Our sources do not give the exact size of the speed-up, so treat it as an open debate, not a measured fact.
This connects to Friday's story: OpenAI's own team said GPT-6 Astra quietly does worse on safety research. If AI does more of the research, and less well on the safety part, the gap can grow.
If AI speeds up its own making, rules and safety checks fall behind even faster than today. This is the speed problem in its purest form. It is the best reason yet for independent testing before new models ship, not after.
