Nathan Lambert: Why I still haven’t bought into true RSI

Summary

Nathan Lambert argues that the belief in short-term recursive self-improvement (RSI) within the AI safety community is a culturally driven anxiety rather than an evidence-based prediction. According to Lambert, due to diminishing returns, resource bottlenecks, and irregularities in AI capabilities, 'lossy self-improvement' is a more realistic scenario than a sharp intelligence explosion.

Why it matters

As leading labs deploy thousands of concurrent agents, the safety community is increasingly leaning toward 'single point of failure' scenarios within the year. Lambert’s objection provides a necessary balancing perspective by distinguishing between predictable inference scaling and truly unpredictable RSI. This framework is significant because it questions the internal culture of labs and the public narrative. This narrative, which conflates rapid agent efficiency with an imminent capability cliff, directly shapes policy urgency and capital allocation.

Highlights

• Lambert's 'lossy self-improvement' thesis: The automatable research domain is very narrow. Parallel agent returns are diminishing. Political and resource bottlenecks limit momentum.
• Schulman, Millidge, and O'Neill present 1-10 year timelines for white-collar job replacement and ASI. Lambert, however, notes that these thresholds are crossed through slow diffusion rather than sharp jumps.
• Lambert warns against conflating massive inference scaling with true RSI. Inference scaling is predictable and subject to economic limits due to public market oversight. True RSI, however, is highly uncertain and shaped by imagination.

Critical take

Lambert’s own position is notably ambiguous. He acknowledges a 'high level of uncertainty.' He also recognizes that labs may possess unpublished breakthroughs that could force him to change his views. This reduces the article to a balanced cautious stance rather than a genuine counter-argument. Additionally, the claim that 'most of science is communication and standard-setting' is presented without evidence. This approach risks underestimating the potential for agent-mediated collaboration to even compress human-driven bottlenecks.