The paper discusses the use of training compute thresholds in regulating general-purpose artificial intelligence (GPAI) models, which are increasingly posing risks of large-scale societal harm. Training compute refers to the total number of computational operations used in the training phase of an AI model. The authors argue that training compute is a suitable metric for identifying GPAI models that require regulatory oversight due to its correlation with model capabilities and risks, ease of measurement, and external verifiability. However, they acknowledge that training compute is an imperfect proxy for risk and should not be used in isolation to determine mitigation measures. Instead, it should be used as an initial filter to identify models that warrant further scrutiny, such as through model evaluations and risk assessments. The paper also highlights the limitations of training compute, including its potential to become less effective as algorithmic efficiency improves and computational resources become more accessible. The US AI Executive Order 14110 and the EU AI Act are discussed as examples of how compute thresholds are currently used in regulatory frameworks. The authors conclude that while compute thresholds are a key tool in GPAI regulation, they should be complemented with other metrics and more precise proxies for risk to ensure effective governance.The paper discusses the use of training compute thresholds in regulating general-purpose artificial intelligence (GPAI) models, which are increasingly posing risks of large-scale societal harm. Training compute refers to the total number of computational operations used in the training phase of an AI model. The authors argue that training compute is a suitable metric for identifying GPAI models that require regulatory oversight due to its correlation with model capabilities and risks, ease of measurement, and external verifiability. However, they acknowledge that training compute is an imperfect proxy for risk and should not be used in isolation to determine mitigation measures. Instead, it should be used as an initial filter to identify models that warrant further scrutiny, such as through model evaluations and risk assessments. The paper also highlights the limitations of training compute, including its potential to become less effective as algorithmic efficiency improves and computational resources become more accessible. The US AI Executive Order 14110 and the EU AI Act are discussed as examples of how compute thresholds are currently used in regulatory frameworks. The authors conclude that while compute thresholds are a key tool in GPAI regulation, they should be complemented with other metrics and more precise proxies for risk to ensure effective governance.