Epoch AI

What it is

Epoch AI is a data-first research nonprofit that investigates the future of artificial intelligence ^[1]. "Data-first" describes its working method: instead of starting from speculation about where AI might be heading, the organization begins by assembling measurable records — how much computation went into training a model, what hardware exists, where investment flows — and reasons from those numbers.

Origins

Epoch AI began in 2021 as a group of volunteers collecting and analyzing data on AI and curating information from hundreds of machine learning models ^[1]. After the success of its 2022 paper analyzing training compute trends, the group founded Epoch AI in order to scale up its efforts in researching trends in artificial intelligence ^[1]. The path from a volunteer data-collection effort to a formal organization is unusual but telling: the founders' initial activity was measurement, and the organization was built around continuing it at a larger scale.

Databases and measurement

Epoch AI develops and hosts databases that track machine learning hardware, estimates of the training compute used by AI models, investment, GPU clusters and AI data centers ^[1].

Training compute generally refers to the amount of calculation spent on training a model, often expressed in floating-point operations; because it is comparatively easier to measure than many other properties of a model, it is used as a rough proxy for scale. Epoch AI was among the first to systematically monitor trends in AI training compute ^[1]. It also identified crucial shifts in how computation is used, such as the role of inference compute scaling ^[1]. Inference is the stage at which a trained model is run to answer requests, so inference compute scaling means putting more calculation into that stage rather than into training alone — a change in where the industry's computing effort goes.

Evaluations and benchmarks

Epoch AI regularly runs independent evaluations of new AI models across domains ^[1]. An evaluation typically means testing a model on a standardized set of tasks with a fixed scoring method; the point of an independent evaluator is that the results are not produced solely by the model's developer.

Epoch AI created FrontierMath, which it describes as the state-of-the-art benchmark for testing frontier math capabilities ^[1]. A benchmark is a shared set of problems that lets different models be compared against the same yardstick.

The economics of AI

Epoch AI has conducted highly cited research on the economics of AI ^[1]. This includes analyses of the consequences of automating remote work and quantitative models of AI's impact on the economy ^[1]. Such work connects the measurement side of the organization — compute, hardware, investment — to questions about what those trends mean for labor and economic output.

Partnerships and public visibility

Epoch AI's past and ongoing partnerships include Stanford's AI Index and the UK's Department for Science, Innovation, & Technology ^[1]. Time magazine profiled Epoch AI's Jaime Sevilla in a piece about how Epoch came about ^[1].

Where it sits in the field

Taken together, Epoch AI's work is mostly infrastructure for other people's thinking: databases that others can draw on, benchmarks that make capability comparisons possible, evaluations that do not come from the model builders, and economic models that turn trends into scenarios. The organization's stated focus is the future of artificial intelligence, approached through the accumulation of concrete, checkable data rather than forecasts alone ^[1].