Algorithmic Progress

Algorithmic progress refers to improvements in the performance of an AI system that come from changes to the algorithms, architectures, training procedures, data curation, and implementation details, holding fixed the compute and data available to train the system. The concept is the residual term in AI-progress decompositions: once you have subtracted the contribution of growing training compute and growing data, the remaining improvement in capability is attributed to algorithmic progress. As a measurement target, it is one of the foundational inputs to the economics of AI and to debates about long-run AI progress, because it is what researchers can speed up through better ideas, in contrast to compute scaling which is gated by hardware availability and capital. The term traces in its modern form to OpenAI's 2018 analysis of compute trends in AI, which noted that a fixed amount of compute was producing increasingly capable models over time and which named this trend algorithmic progress. The Epoch AI project subsequently made it a primary measurement, producing the first systematic time series of algorithmic efficiency gains.

The mechanism, as typically formalized, is to choose a target capability level — for example, achieving 80% accuracy on ImageNet or 90% accuracy on a benchmark — and then ask how much compute was required to reach that target as a function of time. The inverse trend line — compute required to reach fixed capability declining over time — is the algorithmic progress curve. Equivalently, one can ask what level of capability a fixed compute budget buys in different years. Measurement is tricky because training compute, datasets, and architectures all change simultaneously, and any decomposition requires assumptions about what is held fixed. Epoch and its collaborators have historically estimated the rate at around 6-12x improvement per year in the period from 2012 through 2024, with substantial variance across tasks and a tendency for the rate to accelerate in the deep-learning era relative to the classical machine-learning era. In the Cunningham et al. RSI paper of 2026, algorithmic progress is one of the four named feedback loops in the directed-graph decomposition of self-sustaining acceleration.

The tradeoffs the concept surfaces are central to forecasting disagreements about AI trajectories. A high rate of algorithmic progress implies that compute can be deployed more efficiently than the hardware curve alone would suggest, that future AI capabilities are not fully gated by hardware availability, and that policy interventions targeting compute — export controls, compute caps, disclosure requirements — are partially erodible by software improvements. A low rate implies the opposite: that capabilities roughly track compute, and that limiting compute is a near-binding constraint on progress. The concept is also critical in separating exogenous progress drivers from endogenous ones; in economic models of AI, algorithmic progress is typically the endogenous variable that AI research itself accelerates, while compute scaling is treated as exogenous, gated by hardware and capital investment. Whether AI can fully automate its own algorithmic-research pipeline — and thus remove the distinction between algorithmic progress driven by humans and algorithmic progress driven by AI — is the central open question for whether algorithmic progress can become self-sustaining.

Open questions include how to make the measurement comparable across radically different tasks and modalities — the natural generalization across language, vision, robotics, and multimodal models is not straightforward, and the headline rate of algorithmic progress depends on which benchmark family is chosen. Measurement also depends on access to compute and architecture details that AI labs do not routinely publish, making the underlying time series more uncertain than the headline numbers suggest. The deeper conceptual question is whether the rate of algorithmic progress itself can be improved by better tools — better experiment tracking, automated hyperparameter search, AI-assisted research — or whether the rate is fundamentally constrained by what human researchers can notice and articulate as ideas. As of mid-2026, the debate is unresolved and the rate estimate is wide enough to support materially different forecasts of when AI capabilities cross various thresholds.