Lambert argues for 'lossy self-improvement' over true RSI

Nathan Lambert doesn't buy the recursive self-improvement story. In a post on Interconnects, he lays out why he still expects "lossy self-improvement" rather than true RSI to describe the trajectory of frontier model progress.

His argument turns on three limits. Automatable research is too narrow, he writes, given that "all of our scaling laws show that you need exponential compute and resources to make linear improvements in intelligence." Diminishing returns from running more AI agents in parallel are real. Resource bottlenecks and politics still gate how strong LLMs get built, and AI can do little to loosen them.

Lambert cites Anthropic's own recent framing as evidence for the softer read. The company's system card says: "internal usage of recent AI models has been a key factor in maintaining the current rate of progress, but we do not yet see clear signs of dramatic acceleration beyond that rate." Post-training, he adds via John Schulman, resists full automation because on any post-training team "there are a lot of different areas where you have to figure out how the model should behave."

He is sharper on the surrounding discourse. "The step from this anxiety...to extinction risks feels very religious," he writes. Richard Ngo, he notes, has argued that today's superintelligence-soon claims will turn out "directionally correct...but factually wrong. Specifically, we won't have superintelligence within the next 8 years."

Two researchers we track shared the link the same day.