Astral Codex Ten Podcast What Happened With Bio Anchors?
Mar 10, 2026
A deep look at Ajeya Cotra's BioAnchors forecasting method and where its assumptions held up or slipped. Discussion of compute, flops, and effective compute growth rates. Examination of algorithmic progress, revised timelines, and why some parameters were misestimated. Reflections on sensitivity, uncertainty, and what changing data means for AI forecasting.
AI Snips
Chapters
Transcript
Episode notes
How BioAnchors Translated Compute Into Timelines
- BioAnchors predicted AGI by estimating compute growth and compute-needed anchors and dividing the two to get a date.
- Ajeya Cotra used five biological anchors and measured effective flops growth from planned data center buildouts and algorithmic efficiency to produce timelines.
Small Parameter Changes Shift AGI Dates Wildly
- Re-estimating growth parameters (compute availability and algorithmic progress) moved median AGI dates from the 2050s into the 2030s.
- Epoch/Crox found total effective compute grew ~10.7x/year versus Cotra's ~2.4x, largely due to much faster algorithmic gains.
Algorithmic Progress Was The Dominant Error
- Cotra underweighted algorithmic progress because she based it mainly on Hernandez and Brown's AlexNet halving results and adjusted them downward.
- Later data showed frontier AI tasks experienced far faster algorithmic improvement than AlexNet proxies predicted.