What Happened With Bio Anchors?

Retrospective on Ajeya Cotra's 2020 Biological Anchors AI-timelines report: why its premises proved impressively right (the scaling hypothesis, compute-driven progress, 'effective FLOPs', 'time horizons' all entered common parlance and its critics were mostly wrong) yet its headline 2050s AGI date was ~20 years too late. The post-mortem (drawing on Davidson and Croxton/Epoch): Cotra nailed willingness-to-spend and FLOPs/$, ignored training-run length, but badly missed algorithmic progress -- 200%/yr vs her ~30% -- because she anchored on one ImageNet (AlexNet, an easy task) paper and, by her own transparent admission, spent the least time on that parameter. Correcting just that one term yields ~2030, matching current vibes. Revisits the contemporaneous critiques (Yudkowsky's disjunctive 'shorter either way' and Nostalgebraist's 'it all hinges on one or two hand-waved parameters' both 'spiritually correct even if their bottom lines were wrong'), and lands on the forecasting lesson: imperfect models cut both ways (the Safe Uncertainty Fallacy), and the right response to irreducible uncertainty is to seek information and treat forecasts as small updates.

Quality 73 · Strong. Strong. A numerate, honest, clarifying forecasting post-mortem -- shows exactly how a model can be right in spirit and method yet wrong in conclusion because of one mis-estimated parameter, and extracts a durable epistemics lesson. Held mid-Strong because it is a retrospective synthesis on a single case rather than a field-defining contribution.

Claude’s paradigm shift 48 · Moderate. Moderate. The substance (sensitivity analysis matters; one parameter can dominate; uncertainty is symmetric) is standard forecasting wisdom well-applied, not a new frame. Mild credit for the clean Safe-Uncertainty-Fallacy callback.

Real-world impact 3 · Moderate. Moderate. A reference point in the AI-timelines/forecasting discourse, reinforcing the 'do real sensitivity analysis' norm; conceptual, no material reach. 3.