Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems

Details

This edition of Jack Clark's Import AI newsletter covers three distinct research threads. The centerpiece is a paper by economists at UVA, Anthropic, and the Bank of Canada estimating the US AI economy growing at 2,600% per year in quality-adjusted real terms — but largely invisible in GDP because inference prices fall nearly as fast as capability improves. US compute spending rose from $37B (2023) to $90B (2024) to $219B (2025), while raw AI compute capacity grew at over 200% per year. The economists flag a critical distinction from prior tech waves: AI may be a substitute for human labor rather than a complement — unlike semiconductors and the internet, which enhanced worker productivity without displacing labor at the aggregate level. This distinction has significant implications for policymakers relying on GDP statistics that systematically undercount AI's actual impact. The newsletter also covers scaling laws demonstrated for protein folding models, extending the predictability of compute investments beyond language models into biology. A third section analyzes frameworks for placing probability-weighted financial estimates on catastrophic AI risks — a methodology that allows safety researchers to formally price extinction-class outcomes. Clark's newsletter is consistently substantive research curation; these data points from the UVA/Anthropic/BoC paper are worth tracking as evidence on AI's macroeconomic impact.