I started to work in Artificial Intelligence (AI) in the mid-1960s – AI had been around for about ten years by that time. Within a few years, I began to understand that I was likely to be perpetually out of step with most other people working in AI. The other AI people I met wanted to create intelligent machines that did intelligent things. They didn't much care if the machine played chess the way a person would play chess. They just wanted to build a program that played chess well. Similarly, in what was then called computational linguistics, they wanted computers to parse sentences and translate the sentences, but they didn't much care about whether the programs operated the way people did when they understand or compose language, nor did they try to shed any light on how people did it.

I wanted to know how individuals’ minds worked. Also, I thought that there might not be any way to process natural language (this is a term, now widely in use, that I coined to get the AI work on language out of the linguistics camp) other than the way individuals did it. At least we knew individuals had some way in which they could understand and respond. Why not study individuals instead of enhancing algorithms?

I was an undergraduate at Carnegie Tech (now Carnegie Mellon) but there was no psychology major nor was there a computer science major. I majored in mathematics. It was difficult to find colleagues who cared about what I cared about. I wanted to study learning, but in the 1960s, the study of learning was dominated by behaviorist theory and methodology. Learning experiments focused on animals to study stimulus and response relationships. There were lots of rats and pigeons in the psychology department. At that time, no one believed there was any connection between the behaviorist approach to learning, which focused on observable behavior and was opposed to any theories or models about cognitive structures or processes, and the new field of artificial intelligence, which was focused on exactly the opposite: understanding and modeling cognitive structures and processes.

By the late 1960s, when I had become a professor, many of my students wanted to work on learning as well. I had many students who independently proposed to build a computer that started out knowing nothing and learned little by little.