1247 ET - Among the takeaways from Anthropic CEO Dario Amodei's calls this weekend to slow AI development was confirmation that Recursive Self Improvement, self-improving AI models, are being deployed across the industry, Melius Research analysts say in a research note. Amodei warned that the acceleration of RSI introduces a spate of new risks. But the admission is also a huge catalyst for compute demand, the analysts say. RSI runs on inference with thousands of agents constantly writing code, launching experiments and reading results, they say. Slowing upfront model training doesn't stop that loop, and arguably just redirects compute budgets elsewhere, the analysts say. "The models RSI produces are better and cheaper per token and every price cut in the last two months has been met with more consumption," they say.