Read the original writing here.

In this note I’ll summarize the bio-anchors report, describe my initial reactions to it, and take a closer look at two disagreements that I have with background assumptions used by (readers of) the report. (Thanks to Steph Lin for comments on a draft of this review.)

Summary of the report This report attempts to forecast the year when the amount of compute required to train a transformative AI (TAI) model will first become available, as the year when a forecast for the amount of compute required to train TAI in a given year will intersect a forecast for the amount of compute that will be available for a training run of a single project in a given year.

The report estimates the former by first estimating the amount of compute needed to train TAI in 2020 assuming that 2020 algorithms anchored to the sizes of various biological processes can scale to TAI, then correcting for how this requirement may fall over time due to incremental algorithmic progress. It estimates the latter by multiplying an estimate for compute available per dollar in a given year with an estimate for dollars available per training run in a given year.

Most of the report focuses on generating the 2020 training compute requirements distribution.

---

narrator_time: 3h20m
editing_time:
narrator: pw
editor: pw
qa: mds
client: EA Forum
project_id: EA Forum Red Team Prize

---