AI Isn't Recursively Self-Improving Yet — But the Feedback Loop Is Tightening

A team led by METR’s Tom Cunningham, writing under the Elasticity Institute, builds a formal economic model of recursive self-improvement (RSI) — the scenario where AI accelerates its own development. Rather than argue over definitions, they focus on a measurable question: does a one-unit gain in model capability produce more than a one-unit gain in the next generation? They render AI progress as directed graphs whose edges are elasticities, and show that a ‘self-sustaining acceleration’ — capabilities rising without any added human labor or compute — occurs only when the combined strength of the feedback loops exceeds one. A key distinction runs through the analysis: AI could accelerate ‘narrow’ capabilities like algorithm optimization and benchmark-tuning while barely moving the ‘broad’ capabilities that drive real economic output, producing fast but jagged and lopsided progress.

Their back-of-the-envelope calibration, anchored to Epoch’s Capabilities Index, finds the tipping point sits at roughly a 15% research-productivity boost per unit of capability gain. Current returns look closer to 9% — derived from the jump between recent Claude models and a self-reported (and likely inflated) 4x productivity claim from Anthropic engineers — so by this model we are not yet in a self-sustaining regime. But the authors stress the number is climbing fast and doesn’t rule out crossing the threshold soon. The loop could stall on bottlenecks instead: shortages of training or experimental compute, scarce high-quality data, diminishing algorithmic returns, or simply ‘running out of GDP’ to fund the next scale-up.

Much of the paper is a plea for better data. The biggest unknowns — how fast lab algorithmic efficiency actually grows, how R&D spending splits across humans, compute and data, and how much of technical progress AI genuinely contributes — are things labs could measure and disclose without giving away secrets. The authors also flag why speed matters beyond the raw capability numbers: a compressed takeoff would strain labor markets and institutions, widen power gaps between firms and states, and shrink the window for safety work. They add that even a technically plausible acceleration could be throttled by regulation and local backlash, citing tens of billions in cancelled US data-center projects.