Why One AI Researcher Remains Skeptical of True RSI

In a new post, AI researcher Nathan Lambert lays out why he still hasn’t bought into “true RSI”—recursive self-improvement, the idea that an AI system can improve itself in a runaway loop. Lambert positions himself as an “AI moderate,” suggesting he is neither dismissing the concept outright nor treating it as an imminent development.

Lambert reviews recent events and the trajectory of frontier models, but finds the evidence insufficient to support claims of true RSI. While frontier models are clearly advancing, he appears to distinguish steady progress from the kind of autonomous, compounding self-improvement that RSI implies.

The post is a single, self-contained argument, so there are no other sources to compare or contrast. Lambert’s stance is notable for pushing back against both hype and fatalism, urging a more measured reading of what recent AI milestones actually show.