When Will AI Transform the Economy? (andreinfante.substack.com)

The piece argues that AI’s slow macroeconomic impact is less mysterious than it seems: current models have very high “ceilings” (they can solve hard tasks sometimes) but low “floors” (they also make frequent, child‑like errors). That mix makes them impressive in demos but unreliable in production, so businesses enter a long hybrid phase—using models to augment human work—rather than abruptly replacing labor. The author frames this with a simple translation-firm model (human translation $60/hr, 300 hours = $18,000; careful human review 100 hrs = $6,000; machine costs negligible, reputational loss ~$50 per wrong sentence). Under that model, machines only start producing measurable cost and employment effects around ~35–50% sentence accuracy (hybrid strategies save ~20% at ~50%), while a sharp discontinuity to mass automation requires near‑perfect accuracy (high‑90s), and each extra “nine” of reliability is much harder and costlier to attain. The takeaway for AI/ML practitioners is practical: progress in headline capabilities isn’t enough—reducing error rates, verification, and domain‑specific reliability drive economic impact. Hybrid deployments can increase volume and shift labor rather than eliminate it (as happened after compilers), but once many tasks cross the critical quality threshold, labor demand can collapse rapidly. The author expects substantial, measurable economic effects in many fields within roughly five years if current improvement trends continue, barring major macro shocks.