The Horizon of Cognition and the Breakthrough Called AI
Recommended Article : 【Philosophy】 Philosophy of Artificial Intelligence Table of Contents
It was not long ago that news about Google DeepMind’s AlphaFold2 being considered a candidate for the Nobel Prize surprised us. On May 9th, AlphaFold3 was announced, allowing us to see interactions between proteins, nucleic acids, or protein-nucleic acid interactions. Unlike AlphaFold2, AlphaFold3’s source code is not public, and Google’s intent to monopolize the field of artificial intelligence proteomics with a strict emphasis on non-commercial only is evident. Of course, since the pseudo code for AlphaFold3 has been released, there is exciting news of attempts around the world to reverse-engineer AlphaFold3.
The emergence of AlphaFold3 shows the possibility that AI can create new knowledge. Another example of AI’s ability to generate knowledge is FunSearch, developed by Google DeepMind as well. The AI that creates new knowledge is almost exclusively being done by Google DeepMind, which always greatly stimulates my imagination, and I have concluded that our cognitive range can be expanded.
Recently, I thought of the concept of the horizon of cognition. The event horizon, which inspired this term, is the point where the expansion speed of the universe and the speed at which light reaches us balance, defining the boundary of the observable universe. Similarly, it seems that humans also have a boundary of the observable universe, such as not understanding the world of molecules (**microscopic boundary**) or failing to predict macroeconomics or the future a day ahead (**macroscopic boundary**). This is because in this world, where diverse knowledge forms layers through emergence, we can only understand ranges similar in size to us, and much smaller or larger worlds seem incomprehensible.
For example, why has the grand unified theory, which attempts to merge quantum mechanics and general relativity, still failed after 100 years? Could it be outside our range of cognition? Professor June Huh, who received the Fields Medal, often called the Nobel Prize of mathematics, solved several mathematical problems by combining combinatorics and geometry. Mathematics is divided into four areas: geometry, algebra, combinatorics, and number theory, and the number of ways to combine two or more areas is (24 - 4) = 12, but still, there are mathematical problems that cannot be solved by human effort alone, right?
However, with an AI model like AlphaFold3, I am beginning to think that our range of cognition can be expanded into realms we cannot understand. The understanding beyond the horizon of cognition can only be comprehended by AI, so it won’t eliminate existing jobs, but it could also lead to the creation of new ones. I refer to this new intellectual entity as an observational AI. It functions similarly to telescopes or microscopes. In fact, the computer program that solved the four-color problem could also be seen as one type of observational AI.
The reason I call it a new intellectual entity is because the diffusion model, which forms the basis of AlphaFold3, represents a new paradigm that has never existed on Earth before. Our brains seem to work like transformers, and our eyes like vision transformers. But the diffusion model, which intentionally blurs the original image and then restores it, represents a type of intellectual entity that has never historically existed on Earth. In the field of computer vision, vision transformers work so well that the diffusion model hasn’t been very active, but in proteomics, it seems that only the diffusion model works well. And behind it, there seems to be a cognitive system that necessarily needs to introduce noise in the world of molecules governed by quantum mechanical uncertainty.
So, what new observational AI could exist in the future? Considering 1) the existence of a large amount of training data, 2) exceeding human cognitive range, it might not be difficult to predict new fields. I still believe AlphaFold3 has not solved all the difficult parts of proteomics. For example, a diffusion model that predicts time-dependent molecular dynamics simulations or even byproducts, organic chemical reactions could soon emerge. And logical reasoning AI that exceeds logical reasoning to create new axioms, AI that reproduces the history of the universe, AI that predicts the macroeconomy a minute later (though the uncertainty of history means that predicting the macroeconomy an hour later is not possible) could be examples of this.
Input: 2024.06.08 08:19