The Sigmoids Won't Save You

Scott Alexander argues that the common dismissal of AI capability growth via 'all exponentials become sigmoids' is logically weak. Using examples of failed sigmoid predictions in solar deployment and UN birthrate forecasts, he shows that trends often continue far longer than forecasters expect. He introduces Lindy's Law as the appropriate default when modeling AI progress as a black box: if a trend has continued for ~7 years, expect it to continue for roughly another 7. The burden of proof, he argues, lies with those claiming the AI capability curve will plateau before reaching transformative levels — they must either present an explicit model or explain why Lindy's Law doesn't apply.