Enterprise AI pilots fail by applying generative models to discriminative tasks

Cold-Interview6501 · reddit · 2026-08-20

The post argues that most enterprise AI pilots fail to reach production because they apply generative models to problems requiring discriminative ones. Comparing a successful fraud detection system with a stalled LLM pilot, it highlights fundamental differences: discriminative ML continuously updates parameters on operational data ($P(y|x)$), while generative LLMs rely on frozen external weights ($P(x|prompt)$). The author addresses counterarguments regarding fine-tuning, open weights, and data sovereignty.

Original post →

More from Companies & People

Companies & People channel →