Anthropic’s recent warning about self-improvement is another marketing stunt, meant to pre-emptively create an exculpatory narrative to justify their slower rate of model releases heading into IPO. They are no closer to RSI than OpenAI — who admitted openly in the GPT-5.5 system card that they see no signs of RSI on their own internal AI research benchmark tasts.

RSI is the core narrative propping up their entire valuations right now. They need to maintain the belief that the future will very quickly deliver exponential cashflow growth. “2-3 years” — always 2-3 years! Show the actual gears-level mechanical analysis whereby we will see evidence of RSI. I tried doing so in this essay and while there’s evidence of marginal 1-off tasks where AI can genuinely improve training cycles, there’s no evidence for improvements which would recur run to run, or yield improvements in the most critical area: AI R&D.

We won’t see automated AI R&D until we see automated scientific R&D. Automating parts of the pipeline doesn’t count here — you need to automate the production of fruitful new ideas. And LLMs just don’t do that, because that stuff is almost always way out of distribution.