Epoch is a research group trying to forecast the development of transformative Artificial Intelligence. We try to understand how progress in AI happens and what economic impacts we might see from advanced AI.

We want to enable better governance during this economic transition by informing about the timing of new developments and what levers can be used to influence AI progress.

Research agenda

Our research agenda includes both empirical and theoretical work.

Modelling the future of AI

We are developing statistical and economic models to answer crucial questions about the direction and impact of AI in the coming decades.

Studying trends in Machine Learning

Curating and analyzing information about significant ML systems like compute budgets, training dataset sizes, cost and performance.


The Direct Approach


Apr. 25, 2023

Matthew Barnett and Tamay Besiroglu

Empirical scaling laws can help predict the cross-entropy loss associated with training inputs, such as compute and data. However, in order to predict when AI will achieve some subjective level...

Power laws in Speedrunning and Machine Learning


Apr. 21, 2023

Ege Erdil and Jaime Sevilla

We develop a model for predicting record improvements in video game speedrunning and apply it to predicting Machine Learning benchmarks. This model suggests that Machine Learning benchmarks are not close...

Algorithmic progress in computer vision


Dec. 12, 2022

Ege Erdil, Tamay Besiroglu

We use a dataset of over a hundred computer vision models from the last decade to investigate how better algorithms and architectures have enabled researchers to use compute and data...

Will we run out of ML data? Evidence from projecting dataset size trends


Oct. 10, 2022

Pablo Villalobos, Jaime Sevilla, Lennart Heim, Tamay Besiroglu, Marius Hobbhahn and Anson Ho

Based on our previous analysis of trends in dataset size, we project the growth of dataset size in the language and vision domains. We explore the limits of this trend...

The longest training run


Aug. 17, 2022

Jaime Sevilla, Tamay Besiroglu, Owen Dudney and Anson Ho

Training runs of large Machine Learning systems are likely to last less than 14-15 months. This is because longer runs will be outcompeted by runs that start later and therefore...

Machine Learning Model Sizes and the Parameter Gap


Jul. 05, 2022

Pablo Villalobos, Jaime Sevilla, Tamay Besiroglu, Lennart Heim, Anson Ho and Marius Hobbhahn

The model size of notable Machine Learning systems has grown ten times faster than before since 2018. After 2020 growth has not been entirely continuous: there was a jump of...

Compute Trends Across Three Eras of Machine Learning


Feb. 11, 2022

Jaime Sevilla, Tamay Besiroglu, Anson Ho, Lennart Heim, Marius Hobbhahn, and Pablo Villalobos

Compute, data, and algorithmic advances are the three fundamental factors that guide the progress of modern Machine Learning (ML). In this paper we study trends in the most readily quantified...

Estimating Training Compute of Deep Learning Models


Jan. 20, 2022

Jaime Sevilla, Lennart Heim, Marius Hobbhahn, Tamay Besiroglu, Anson Ho and Pablo Villalobos

We describe two approaches for estimating the training compute of Deep Learning systems, by counting operations and looking at GPU time.

Parameters, Compute and Data Trends in Machine Learning


Jaime Sevilla et al.

Public dataset