---
title: Self-Sustaining Acceleration
type: vocabulary
url: "https://www.envisioning.com/vocab/self-sustaining-acceleration"
summary: AI progress that grows without exogenous inputs by feeding on its own feedback loops.
year: 2026
generality: 0.55
---

# Self-Sustaining Acceleration

AI progress that grows without exogenous inputs by feeding on its own feedback loops.
Self-sustaining acceleration in AI capabilities refers to a regime in which advances in AI continue to compound without requiring any growth in exogenous inputs — additional human labor, training compute, capital, energy, or other externally supplied resources beyond what is already deployed. The term was formalized in the economics of AI literature to describe a specific, measurable outcome: AI progress that maintains or accelerates itself purely by routing its outputs back as inputs to its own production process. The formulation matters because it is the closest operational test for whether the recursive-self-improvement hypothesis is occurring: a self-sustaining acceleration is what you would actually observe in the data if RSI were underway, as distinct from the many looser claims that have been attached to the RSI label over the years. The term appears explicitly in a July 2026 paper by Tom Cunningham and colleagues at METR, Stanford, Virginia, CMU, MIT, Columbia, Epoch AI, and Yale as part of a careful taxonomy separating feedback loops, R&D automatability, self-sustaining acceleration, and the intelligence explosion.

The mechanism, per the Cunningham et al. framework, can be analyzed by decomposing the AI development process into directed graphs of feedback loops and asking whether the elasticity of each loop with respect to improvements in AI capability is large enough to compound. Net acceleration depends on the product of elasticities across every loop: if any loop has an elasticity below one, the cascade decays, and even a single bottleneck can prevent self-sustaining acceleration regardless of how strong the other links are. The result is that strong AI progress requires strong feedback not just in compute or in algorithmic innovation but across the whole chain — automated researchers using AI to improve AI, automated infrastructure, automated evaluation and benchmarking, and automated data generation. The model also distinguishes narrow from broad capability: an AI system that improves at AI R&D benchmarks without improving at economically useful tasks accelerates only its own R&D, not the broader economy, and self-sustaining acceleration of capabilities does not by itself imply economic take-off.

The tradeoffs the concept surfaces are significant for both measurement and policy. Empirical work to date suggests that, as of mid-2026, the feedback loops are not yet strong enough to generate self-sustaining acceleration: the calibration presented by Cunningham et al. reports that aggregate AI progress continues to require exogenous compute and human researcher input to grow, even as those feedback loops are strengthening. This is a different empirical claim from the dramatic intelligence-explosion scenarios, which require not just self-sustaining acceleration but unbounded acceleration in finite time, and from generic worries about RSI that have been part of the AI safety conversation since Good (1965) and Yudkowsky (2001, 2008). For policymakers and forecasting organizations, the value of the self-sustaining-acceleration framing is that it identifies measurable, intermediate signals — trends in the share of AI-generated code in AI labs, trends in automated R&D budgets, trends in evaluation automation — that can be tracked without committing to strong claims about what an intelligence explosion would look like.

Several open questions remain. The elasticity estimates required by the framework are not directly observable for many of the loops in the cascade; they have to be inferred from indirect evidence about how AI is used inside AI labs, how much of frontier capability progress is now attributable to AI-assisted research, and how the latency between scientific insight and AI capability has evolved over the past several years. The paper explicitly identifies a wish list of empirical objects — measurements that AI companies could share publicly without revealing proprietary information — that would tighten these estimates substantially. Whether the loops ever cross the threshold from strengthening to self-sustaining is, on this framework, a quantitative rather than qualitative question, and resolving it depends on data the field does not yet collect systematically. The framing also leaves open whether self-sustaining acceleration, once initiated, would actually prove stable or whether it would induce countervailing dynamics — regulation, diminishing returns to AI ideas as the easier improvements get found, social pushback — that prevent runaway persistence.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/self-sustaining-acceleration)