Taylor Pospisil

Stop using LLMs as classifiers

LLMs seem good at classification but lack calibration, interpretability, and control. Treat them as feature engineers instead.

Minimally Sufficient · 3m