Where small models beat frontier LLMs (and where they don't): a 125M PHI detector

Last month I published a 184M-parameter intent classifier that matches frontier LLMs at 22× lower...