Ajeya Cotra

We'll be making use of them here.

Crocs found that Kotra and Davidson underestimated annual growth in effective compute.

Here's a table.

It shows different factors and the Kotra, Davidson and Epic slash Crocs estimates.

So for willingness to spend, it's Kotra 1.4, Davidson 1.6 and Epic Crocs 1.7.

Cost per flop, 1.3, 1.1 and 1.4.

Training run length, only Epic and Crocs have provided an estimate, it's 1.5.

Real compute, it's 1.8, 1.8 and 3.6.

And algorithmic progress, 1.3, 1.3 and 3.0.

For total effective compute, Kotra's estimate 2.4, Davidson 2.3, and Epic Crocs 10.7.

Scott captions this, Epic Crocs are current best estimates, and can probably fairly be read as the real answer in quotes, against which Kotra and Davidson's earlier guesses should be judged.

All numbers are yearly multiples, so 1.4 means that willingness to spend grows 1.4 times per year, that is 40%.

Willingness to spend.

How much money are companies willing to spend on AI, in the form of chips and data centers?

Dollars per flop.

How quickly do Moore's Law, economies of scale, and other factors bring down the price of AI compute?

Training run length.

How long are companies spending on AI training runs for frontier models, instead of using those chips for smaller models, experiments, or consumer services?

Real compute.

The product of the three parameters above.