Christopher Manning

Manning's body of work — GloVe, the foundational Stanford NLP curriculum, a generation of his PhD students who fanned out across industry labs — gives him a vantage point from which most current LLM coverage looks like a category error. His sharpest recent writing argues that when a model appears to "understand" something, we are watching statistical regularity in human language being mistaken for the conceptual structure those patterns ride on. The position is rare for someone with his standing and his institutional incentives, which is part of why it's worth taking seriously.

Worth following when

you want to read the gap between "the model produced the right words" and "the model knew what those words meant", written by someone old enough in the field to know what got lost in the rebranding from NLP to AI.

Topics

what distributional semantics can and can't claim about meaning; structural critiques of current LLM evaluation practice; the long arc of language technology before LLMs.

Key works

GloVe word vectors (2014, co-lead author with Pennington and Socher); "Foundations of Statistical Natural Language Processing" (1999, with Schütze); long body of public critiques of current LLM evaluation practice.