https://twitter.com/AndrewYNg/status/1736577228828496179

This is one of those perspectives that feels wise but I quite disagree with. Consider Covid in Jan 2020. At that time, was it right to focus on (i) actual current realized harm, or (ii) hypotheticals based on projecting exponential functions? Clearly (ii).

but we have thousands of years of evidence about how the exponential growth of diseases works

we have no prior evidence to make good decisions about how the exponential growth of AI will unfold

intersubjective mechanism design @ poidh.xyz - come chill /poidh + follow @poidhbot

The thing about forecasts with wide priors is wide posteriors.

+1, it feels like we're extrapolating astronomic numbers from fitting an exponential curve on ~3 data points, assuming that it's only up forever into infinity. No thermodynamic constraints, no reactionary measures, no unexpected dead ends that we realistically hit over and over and over.