About

I’ve been learning and thinking about artificial intelligence since I took my first class in university titled “Symbolic Computation and Artificial Intelligence”. It was in 2003, and we were in the AI Winter just after the dot-com bubble. Symbolic AI was losing popularity and Machine Learning was the new cool.

I spent quite a lot of time since then reading about cognitive science and neuroscience, looking for answers in understanding how the human brain works, because a lot of the problems in computing seemed to stem from our lack of understanding of how the human mind processes information and makes decisions.

Dismayed by the gap in our level of understanding, I began to look for answers in philosophy, psychology and anthropology. Computational Cognitive Neuroscience approaches cognition from below, attempts to recreate it from the building blocks. Philosophy starts from the phenomenon, and tries to work its way down. I quite liked the change in approach, where cognitive science is a lot more like engineering, trying to come up with discrete answers, whereas the humanities put greater emphasis on insights and questions, rather than rushing to an answer. Something I had to learn and unlearn.

The strong belief that LLMs will solve this problem (or have already solved it, according to some) annoys me to no end, but I wanted to find the right words to explain why I feel that way. I feel they have an effect on people, and our interaction with AI gives rise to interesting phenomena on its contours. How we perceive and react to it gives insights into how our mind works.

Over the years I’ve settled on a few positions and I will be writing in detail about them. Below I’m stating some of them to give a map of my thinking.

I don’t believe statistical methods will lead to a true understanding of how the mind works.

We’ve built systems to differentiate bird calls, but it gave us no insight into how we hear bird calls, why we pick them out through the noise, and more mysteriously, how we categorise them. When we try to explain these in concrete steps we always fail.

We need a new and more precise language to replace anthropomorphic terms

We use words like “knowledge”, “understanding” and “intelligence” to describe things we intuitively know but can’t explain in discrete terms or mathematically prove. Marvin Minsky called these words “suitcase words” because they contain so many different concepts encapsulated in a single word. We need to unpack all these concepts, just like physics had to do for weight and mass.

Machines have no subjectivity

We haven’t got a good definition of what meaning is and how it’s produced but I suspect it is produced through subjectivity. LLMs can generate text but they can’t generate meaning; it needs a human being to interpret the world and give meaning to their output.

AI is media

In Baudrillardian terms, the map precedes the territory: model comes before the real. Generative AI is hyperreal because it produces signs from descriptions of the world rather than from the world itself. It becomes a new way for a subject to experience the world, a new stand-in for reality.

When building products with artificial intelligence, these questions come up all the time and I find myself repeating these arguments. They also contain lots of conflicting positions and counterarguments, which is normal because nobody has figured out the answers to these open questions yet. I wanted to hash them out and clarify my thinking, because I want to communicate them better, be better at explaining why a certain thing doesn’t work, or at spotting a pattern before committing to building the thing.

Some of the pieces will be long essays on set positions, and I will be updating them as my thinking changes. Others will be shorter blog posts that I intend to write as I read a new thing.