Routes How fast is the computing power used to train frontier AI growing?
How fast is the computing power used to train frontier AI growing?
A spreadsheet is enough. No programming, account or background in AI needed.
kosplora-route/0.2 · by Bentley Moon
What would change this route
If the dated file stopped matching Epoch AI's own records for the same models, or Epoch changed how it picks frontier models, the route would move to a new dated file and the figures it checks against would change with it, with the date of the change.
The whole route, before you start
- 1 · orient · 20mRead the summary of Epoch AI's report.
- 2 · source · 45mRead the paper's summary and its section on why computing power is easier to govern than data or algorithms.
- 3 · practice · 50mOpen the dated file in a spreadsheet.
- 4 · counterpoint · 30mRead the summary of a paper on how much less compute the same language-model result needs each year.
- 5 · make · 30mWrite a one-page note: the file and window you used, your four figures, whether your growth multiple falls inside Epoch's 4 to 5 times a year, and what the counterpoint changes about reading the trend.
- 6 · person · 15mGive the note and the file to one other person and ask them to recompute one figure.
- 7 · reflect · 10mRecord what a fitted trend does and does not tell you about the next model.
Optional study preferences. These stay in this browser and do not change the route’s steps.
Station 1 · orient · 20m
Read the summary of Epoch AI's report. Write down the growth figure it recommends, and how it decides which models count as frontier models: those in the top 10 by training compute when they were released.
Training compute of frontier AI models grows by 4-5x per year · Jaime Sevilla and Edu Roldán, Epoch AI · 2024-05-28
Station 2 · source · 45m
Read the paper's summary and its section on why computing power is easier to govern than data or algorithms. Write down the properties it names and one rule that already uses a compute threshold.
Computing Power and the Governance of Artificial Intelligence · Girish Sastry, Lennart Heim, Haydn Belfield, Markus Anderljung, Miles Brundage, Julian Hazell and others · 2024-02-13
Station 3 · practice · 50m
Open the dated file in a spreadsheet. Keep the models published from 1 January 2010 to 28 May 2024, the window of Epoch's report. Add a column of years since the start, =(date - DATE(2010,1,1))/365.25, and a column of =LOG10(compute). Take =SLOPE(log column, years column): that is orders of magnitude per year. The growth multiple is 10 to the power of the slope, and the doubling time in months is 12*LOG10(2)/slope. Then count the models in the whole file above 10^25 FLOP.
Frontier AI models: model, organization, publication date and training compute, as published by Epoch AI on 21 October 2026 (127 models, CC BY 4.0) · Epoch AI · 2026-10-21
Alternative: Epoch's live data, with every column and its documentation, is at epoch.ai/data/ai-models. Its figures will have moved since this file was taken, so check against this file first.
Station 4 · counterpoint · 30m
Read the summary of a paper on how much less compute the same language-model result needs each year. Write down what it means for a rule that is triggered by a compute threshold.
Algorithmic progress in language models · Anson Ho, Tamay Besiroglu, Ege Erdil, David Owen, Robi Rahman, Zifan Carl Guo and others · 2024-03-09
Station 5 · make · 30m
Write a one-page note: the file and window you used, your four figures, whether your growth multiple falls inside Epoch's 4 to 5 times a year, and what the counterpoint changes about reading the trend.
Station 6 · person · 15m
Give the note and the file to one other person and ask them to recompute one figure. Note whether they got the same number.
Station 7 · reflect · 10m
Record what a fitted trend does and does not tell you about the next model. Epoch itself reports a slowdown at the frontier around 2018 and says the figure is sensitive to its choices. If you want to go further, the free course below covers how compute could be checked by others.
Compute Verification 1 · Lens Academy · undated
What you leave with
A one-page note reproducing Epoch AI's growth rate for frontier training compute from its own data, with a second person's recomputation, the counterpoint, and the standard it was checked against.
The standard it is checked against
ACM Artifact Review and Badging, version 1.1: Results Reproduced
Association for Computing Machinery · 2020-08-24
Someone other than the authors obtains the central result, using at least in part the authors' own artifacts. Here the result is Epoch AI's growth rate for frontier training compute, and the artifact is Epoch AI's own data.
Who checks it: You, against Epoch AI's published range of 4 to 5 times a year, and then one other person who recomputes one figure from the same file.
Kosplora does not issue a certificate. A partner who recognises this standard would assess the report; no partner has agreed to yet.
When you are done
- study preferences, completed steps and notes in one file
- Download the route file edit a copy in any text editor
- Read the route file format
- Or close the page.