Intel / article

Where I stand on RSI

Source: interconnects.ai
Author: Nathan Lambert
Date: Published Sep 19, 2026

From the Daily Brief

AI researcher Nathan Lambert argues that frontier labs' internal anxiety over recursive self-improvement, RSI, is doing more to shape public AI-risk narratives than the underlying evidence supports. He points to thousands of concurrent coding and research agents now running inside labs like OpenAI and Anthropic as the proximate cause of that anxiety: watching agent swarms operate at that scale internally, Lambert argues, naturally produces alarm among lab staff that then filters into public statements and safety framing, independent of whether current systems are actually closer to self-improving capability. The piece is a contrarian, insider-adjacent read that doesn't dismiss RSI as a long-term concern but pushes back on treating today's lab anxiety as strong evidence that recursive self-improvement is imminent. Useful context for anyone trying to calibrate how much weight to put on safety alarms coming out of frontier labs versus the specific evidence behind them.

Read the 2026-09-21 Brief →

Context

Nathan Lambert expects AI-assisted research to make models cheaper and research operations faster, while leaving substantial obstacles to discovering fundamentally stronger systems. The essay distinguishes automating measurable work, such as running experiments or improving serving efficiency, from choosing valuable hypotheses and developing scientific understanding.

That distinction explains the essay's skepticism about recursive self-improvement becoming a runaway cycle of increasingly capable AI building its successor. More agents can tackle more well-defined problems, Lambert argues, but computing resources, diminishing returns and human understanding still constrain the overall process. The forecast allows substantial economic change through cheaper, more widely usable AI without assuming that research breakthroughs accelerate at the same rate.

Key takeaways · AI-distilled

Nathan Lambert's alternative to recursive self-improvement, 'lossy self-improvement,' argues automatable research is too narrow for runaway acceleration: scaling laws keep exponential costs, parallel- returns diminish, and resource/political bottlenecks resist automation.
John Schulman, Beren Millidge, and Charlie O'Neill gave different timelines for a 10x AI-researcher productivity boost: about 2 years (Schulman; Millidge calls it plausible) versus 5-10 years (O'Neill), bottlenecked on deciding which experiment to run next.
OpenAI's John Schulman argues post-training can't be fully automated because someone still has to decide how the model should behave in each specific area — and mistakes there are 'really easy' to make in ways that never show up in benchmarks.
Anthropic's Claude Fable 5.1 & Mythos 5.1 system card states that internal AI usage has been key to maintaining the labs' current pace of progress, but the company does not yet see 'clear signs of dramatic acceleration' beyond that existing rate.
Lambert's read of where labs' automation gains are actually showing up so far is narrow and routine — software engineering, log monitoring, and managing planned experiments — not the jump in peak intelligence that 'recursive self-improvement' implies.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source www.interconnects.ai