This post talks about Emergence, roughly re-iterating the point of an earlier post on neural networks.

AI Progress

AI is becoming a mainstream phenomenon. Everybody has an AI-generated avatar. Increasingly, we come to ChatGPT for answers instead of Google or StackOverflow. AI is assisting with writing code, and technically can assist in any task that requires mental effort: story writing, accounting, 3D modeling, testing, etc. It becomes a real extension of our natural brain.

Is AI magic? My understanding is - current achievements of AI are well within the original framework of neural networks (NN), which was established in the middle of the previous century (some 70-ish years ago). We knew for a long time that NNs can learn knowledge, and can be built on top of each other, thus operating at higher and higher level. The capacity is only limited by the available compute and storage resources, both of which grew quickly in accordance with Moore’s law. Thus, mechanically, it’s not magic.

There are many examples of other things where the mechanics can be simple but it doesn’t explain the result. My favorite is the Mandelbrot set. The authors created this simple equation without even using a computer: z' = z^2 + c, which was visualized by Mandelbrot only 2 years after the discovery:

zooming sequence

Emergence

To me, the Mandelbrot set is infinitely rich. I can stare at different parts of it, explore in depth, and always end up puzzled about how it’s all generated by a simple equation. The visual richness of the set is emergent: the new set of phenomena could not be directly predicted from the equation, or at least we don’t know how. Nature loves emergence: all of the most amazing things in nature are emerging through multiple layers. Our consciousness and intelligence are also emergent. A good description of this process can be found in “The Infinite Staircase” book.

Coming back to the current progress in AI: yes, we know the mechanics. But no, we don’t know what emergent properties are going to be available through these mechanics if pushed far enough. This is real magic, and it’s called Emergence. The question we are asking ourselves should not be “can NN reach the intelligence of a human?”. The real question should be “how do we predict emergence?” and “how do we shape the basic mechanics in the way of producing certain emergent properties?”. Once we truly understand emergence, building AI would be straightforward. In addition, hopefully, we’ll better understand (and control) the emergence of systems we already have, such as economics and society.