Lossy Self-Improvement — Nathan Lambert
URL: https://www.interconnects.ai/p/lossy-self-improvement
Author: Nathan Lambert (Interconnects AI)
Published: March 22, 2026
Summary
Lambert argues against fast-takeoff recursive self-improvement (RSI). AI models are already tools in the development loop, but the loop is “lossy” — friction from complexity, organizational dynamics, and the gap between automatable and hard research tasks makes exponential self-acceleration implausible. Expect more linear than exponential progress.
Key Claims
- RSI requires three conditions: (1) closed loop — models improve future models; (2) self-amplifying — each iteration yields bigger gains; (3) low friction — the loop doesn’t lose efficiency. Lambert believes condition 3 fails.
- Autoresearch analogy: Karpathy’s autoresearch works for narrow targets (lower val loss on one benchmark). The hard part is the gap between “on-paper more accurate model” and a production-ready, aligned system.
- Paul Allen’s complexity brake: More progress → more complexity → harder to make additional progress. Patent rates peaked 1850-1900 and have been declining. Complexity eventually becomes self-limiting.
- Organizational friction: Building leading models requires deep intuitions, tacit knowledge, and organizational trust that doesn’t transfer cleanly across model generations.
- Multi-agent diminishing returns: Many AI agents working together hit limits because they share similar skills and still require human oversight for the hard decisions.
- Prediction: Progress will feel dramatic socially, but the trend line will look more linear than exponential in retrospect.
Counterpoint to Consider
The 2026 landscape (Claude Code writing production code, AI research co-pilots) does show the loop is partially closed. Lambert may be right about the slope without being right about the destination being linear.