How quantum tunneling breaks chips—and how AI can live with it

As transistors shrink, electrons leak through gate oxides via quantum tunneling, introducing errors in AI chips. Rather than fix every bit, researchers modeled the tunneling process from first principles using WKB approximation and found that these errors have structure—skewed toward important bits and dependent on network weights. Their algorithm TAC corrects for this structure at deployment time, achieving 95% clean accuracy with 3.4–33.6× less error-correction overhead than naive approaches, without retraining or runtime cost.