Nathan Lambert argues against the fast takeoff / recursive self-improvement (RSI) narrative, proposing instead the concept of 'lossy self-improvement' (LSI). While AI models are genuinely becoming core to the AI development loop — automating engineering tasks, optimizing hyperparameters, and running experiments — three key frictions prevent an intelligence explosion: (1) automatable research is too narrow, with agents excelling at single-metric optimization but failing to navigate the multi-dimensional complexity real researchers manage; (2) parallelizing AI agents hits hard diminishing returns governed by Amdahl's law, with human intuition remaining the bottleneck; and (3) resource allocation and organizational politics keep humans in control of compute budgets. Lambert expects AI progress to look more linear than exponential in retrospect, with each wave (GPT-4, reasoning models, agentic workflows) feeling like a step-change but not triggering runaway acceleration. He acknowledges AGI thresholds may be crossed, but argues the complexity brake will prevent paradigm-shifting breakthroughs from emerging autonomously.
