AI and Functions

What would new developments in AI look like? There’s been many advances in this area, particularly around pattern matching, but the newer methods are mostly reverse engineering a function out of data, much like linear regression used in simple problems.

I have a feeling the new direction has to look something like this: a new math around pattern matching is developed. Maybe there is a first-class citizen in the math for a pattern function. I am not talking about approximately representing a function with deep learning, reverse engineering it from data. I am talking about in the abstract, there will be a pattern function, with its own algebraic rules, operations around the matching of the pattern function in various forms. We will be able to play with these rules, and perform symbolic manipulation in any direction we want, and these derivations in the abstract will take us to places that’ll surprise us, which means discoveries. This is how Calculus worked for physics.

It can be argued that when physics was developed there were no computers, so physicists back then had to use symbolic math to do stuff, now we have computers, so we don’t have to. People who argue this are underestimating what mathematics did for science. There isn’t enough time in the world to run the kinds of variations needed that will get us the developments required for huge discoveries, and even then, you would be tied to a particular set of data. So we need some kind of new abstraction.