When AI became decoupled from any understanding of the human mind

I’ve been thinking a bit about the history of AI and how it led to the current situation where it has become effectively decoupled from any attempt to understand the human mind. Now, I’m not thinking about AI across the board, but rather the regnant forms of machine learning that dominate in the commercial market and so dominate current discussions about the implications of AI. Those discussions have become effectively divorced from the study of the human mind in the cognitive sciences, but also the humanities, something I’ve discussed in an article at 3 Quarks Daily, Aye Aye, Cap’n! Investing in AI is like buying shares in a whaling voyage captained by a man who knows all about ships and little about whales.

AI began as an attempt to simulate the human mind. The people who did the work also thought about the mind. AI work on chess led to psychological investigation of how humans played chess. The most commercially successful early AI programs were so-called expert systems, from the 60s on into the 80s. To develop such a system you would ask human experts to think through problems out-loud so you could record their thoughts. The recordings would then be transcribed. This became developed into a systematic methodology called “protocol analysis.” My point is simple, this AI work was closely linked to work on human thinking.

The big breakthrough in machine learning came in 2012 with a machine vision system called AlexNet, developed by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton. It was based on something called a convolutional neural network. CNNs are based on Fourier analysis, which had been used in understanding the visual system going back to the late 1960s. So, at this point the technical basis of the artificial system remained in touch with the study of human perception.

That changed with the development of GPTs. The technical basis of those systems had nothing to do with the technical system of language and cognition. With GPT-3 things exploded. Its language capacity was far beyond anything else that had been done. The field quickly figured out that they could improve performance simply by scaling up, more data, more compute, more parameters. Doing this didn’t require deeper insight into language and thought. It required two things: 1) knowledge of how to scale systems up, a highly developed craft skill, and 2) the money needed to pay for the increased resources. The enterprise was now effectively decoupled from any attempt to understand the human mind.

Of course, no one’s happy that the inner workings of LLMs are mysterious. It makes so-called “alignment” a hellish problem. At the same time, the fact of that mystery makes it easy to imagine whatever you wish about the technology. Thus the black box nature of these systems is convenient for the generation of hype. You can imagine future capacities to be whatever you will. Reality is not going to get in your way, and least not in the present.