mathbabe

Fantastic essay! Something similar has happened in the field of linguistics with opposite consequences. The success of LLMs to learn and generate human language has led neural network engineers to believe that they have obviated the science of linguistics, and they view with annoyance Chomsky’s insistence that LLMs tell us nothing about human language. In fact, neither LLMs nor their creators understand how the bots learn language. Like children, bots have no introspection on how they learn. The “reasoning” they report, if asked, is tracking a sequence of running operations, not their initial learning. Far from obviating the field of linguistics, the engineers have created yet another language learner that needs linguists to explain to them. The goal of technology is to achieve a task for a consumer, be it an individual, corporate or government, and often military, consumer. It doesn’t matter how it achieves its task. The inventor of the bicycle didn’t know that merely steering the front wheel wouldn’t steer the bicycle unless the rider also counter-leaned into the turn. He didn’t need to know. Understanding what that leaning contributes is a scientific goal. The goal of science is not to change the world in any way, but to understand it — explain it — as it is, including understanding and explaining the technologies that engineers introduce into our world. So we linguists not going away just yet.
–and btw, on the funding-consumer problem: Chomsky was funded by the military, but his work has absolutely no military value, never has. Bell Labs funded a lot of linguists who pursued their own scientific interests. Maybe that will change in the AI future, but maybe we should try to understand why there was so much freedom and breadth in the past so we can continue or reproduce it in case it lags.

LikeLike