Use the rule of thumb “training compute ≈ 6 × parameters × training tokens” to estimate the compute behind a model, then compare your answer with Epoch AI’s database.
Compute thresholds appear in real regulation. Being able to estimate compute, and know how uncertain the estimate is, is a core governance skill.
Output: A one-page estimate with your assumptions laid out.
Choose a model whose parameter count and training tokens have been published (many open models report both).
Multiply 6 × parameters × tokens to get training FLOP. Write the answer in scientific notation, such as 3 × 10²⁴.
Look the model up in Epoch AI’s data. How close were you? If you are off, find out why.
Estimate how long that would take on a cluster of chips and roughly what it would cost. Say which assumption you are least sure about.
Find one law or policy that uses a compute threshold. Is your model above or below it?
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Next: How much compute did the key AI innovations need? Next steps
Could algorithmic progress continue under a compute halt? Catalogue the innovations in a major model paper and estimate the compute used in the experiments behind each, extending published MIRI TGT work.
Export controls target data-centre chips. But consumer hardware keeps getting better, and frontier-level performance reaches it within about a year. Update the analysis with the latest hardware.
Download Epoch AI’s free dataset of notable AI models, chart training compute over time in Python, estimate how fast it grows, and count how many models cross the EU AI Act’s compute threshold.
New methods let models train across many loosely connected computers. Use Epoch AI’s simulator and recent papers to analyse whether this undermines governance based on big data centres.
Why do so many governance proposals focus on chips? Read one key paper and write a jargon-free explainer for someone outside tech.
