The distillation panic

Reproduced with permission of author.

Lambert pushes back on the conflation of legitimate AI distillation techniques with illicit API abuse, arguing that using stronger model outputs to train smaller ones is standard industry practice—not an attack. The concern he raises is semantic but consequential: mislabeling this as "distillation attacks" risks triggering regulatory overreach that damages the open-source ecosystem while proving ineffective against determined adversaries, potentially ceding long-term competitive advantage abroad.