What Happened With Bio Anchors?

·25 min·2 clips

Khotra expected AI to get 3.6 times better yearly, but it actually improved 10.7 times—here's how one parameter wrecked her forecast.

The episode opens by introducing Ajaya Khotra's Biological Anchors report from the early 2020s, which was prescient in nailing the scaling hypothesis and predicting the AI boom but whose headline AGI timeline around the 2050s now seems implausible. It then provides a refresher on how Bio-Anchors worked: estimating the rate of compute growth in FLOPs and comparing it to biologically anchored guesses for how much compute AGI would require, yielding a forecast through simple division. The model has held up surprisingly well in hindsight, as AI progress since 2020 has indeed been driven by compute and measurable in effective FLOPs. The central puzzle is why, if its premises were correct, its conclusion was 20 years too late. The episode details Tom Davidson's 2023 update adding recursive self-improvement, which shifted the median to 2043, but notes this still feels late. It then delves into John Crocs's 2025 report card, which used real data from 2020-2025 to show that Khotra and Davidson underestimated annual growth in effective compute, especially algorithmic progress. A table compares estimates: Khotra and Davidson had total effective compute growth at 2.4-2.3 times yearly, while Epoch-Crocs estimates 10.7 times, largely due to algorithmic progress at 3.5 times instead of 1.3. Khotra's section on algorithmic progress is quoted, revealing she based it on one paper about ImageNet classifiers and admitted spending little time on it. The episode explains that this task was an easy one, whereas frontier AI progress has been faster, leading to the error. With corrected parameters, the model predicts AGI around 2030, matching current vibes. The analysis then reviews contemporaneous critiques, including objections to the biological anchors and Eliezer Yudkowsky's argument that a paradigm shift would obviate the calculations. It notes that no such shift has occurred, with scaling laws holding and only minor kinks in 2010 and 2024. Nostalgia Braced's critique that the model hinged on Moore's law is discussed, but the episode argues Moore's law hasn't broken yet. It highlights that algorithmic progress was the crucial parameter, as Nostalgia Braced hinted. The episode distinguishes between good and bad critiques, noting that bad ones often default to epistemic nihilism. It concludes that correcting just one or two parameters would have made the forecast prescient, making Bio-Anchors a white pill for forecasting. The takeaway emphasizes that uncertainty can mean danger, encouraging the use of fallible forecasts as small updates. The episode ends with a speculative image about AI in 2027, illustrating ongoing progress.

As heard by us

A forecast autopsy that shows why the 2050s call aged badly.

The episode frames Ajaya Khotra's Biological Anchors report as an early-2020s milestone in AI timelines: sharp on scaling, useful for putting horizons in view, and still worth revisiting because the 2050s endpoint now feels late.

Read the full review in PlayNext →

Why you'd press play

If you want the math behind why the Bio-Anchors timeline missed, press play.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Astral Codex Ten Podcast