Dean makes an obvious-when-you-consider it point here that "strong" RSI--fully end-to-end automated self-improving AI development by AI--would basically require letting models run the company, including making critical capital allocation decisions, without human approval, which is unlikely to happen for myriad practical and legal reasons.
Recursive self-improvement is a critical issue, but the rhetoric on the subject is becoming more poorly calibrated over time. As we approach new stages of AI development—and RSI is one of them—we need more precise frameworks and terminology. I fear the opposite is happening.View quoted post on X
