Published -
July 2, 2025
Or: How I Learned to Stop Worrying and Love Metaphysics in Machine Learning
Picture this: You're at a party, and someone asks you to explain quantum physics using only kitchen utensils. Meanwhile, your friend is trying to describe the mating rituals of sea slugs using the same kitchen utensils. Now imagine if both of you had to use the exact same explanation format. Sounds ridiculous, right?
Well, welcome to the current state of artificial intelligence, where we're essentially doing the same thing—forcing every domain of knowledge into identical architectural boxes, regardless of whether it makes any philosophical sense.
Here's the thing that's been bugging me (and probably should bug you too): we've built incredibly sophisticated AI systems that can write poetry, diagnose diseases, and even beat humans at complex games. But underneath all that impressive performance, they're basically very fancy pattern-matching machines with about as much understanding of reality as a goldfish has of quantum mechanics.
Current AI development suffers from what I call "ontological blindness"—fancy words for "we forgot to think about what reality actually is." Whether we're predicting how atoms behave, designing new molecules, modeling ecosystems, or discovering materials, we just slap the same transformer architecture on everything and call it a day.
It's like using a hammer for everything because hammers are really good tools, even when you need to perform brain surgery or make a soufflé.
Here's where things get interesting (and where philosophy crashes into computer science like a caffeinated philosopher at a tech conference). What if I told you that physics, chemistry, biology, and materials science don't just study different things—they assume reality works in fundamentally different ways?
Physics thinks reality is all about processes. Everything flows, everything transforms, everything follows conservation laws. It's like reality is one giant, continuous dance where energy and momentum are passed around but never lost.
Chemistry is obsessed with building blocks. It's all about how simple things combine to make complex things, following specific rules. Think LEGO, but with atoms, and the instruction manual is written in the language of molecular orbitals.
Biology believes in the power of teams. Nothing interesting happens unless you have a bunch of specialized components working together. It's like reality is run by the world's most complex corporation, where cells are employees, organs are departments, and evolution is the ultimate HR manager.
Materials science is all about relationships. How things are connected determines what they can do. It's like reality is a massive social network, but instead of posting cat videos, atoms are sharing electrons.
So here's my wild proposition: what if we built AI systems that actually match how these domains think about reality? Instead of forcing everything through the same architectural meat-grinder, what if we created AI that thinks like a physicist when doing physics, thinks like a chemist when doing chemistry, and so on?
I call this "Ontological AI"—AI systems that are designed to match the fundamental assumptions about reality that each domain makes. It's like giving each AI system the right kind of brain for the job, instead of making them all use the same generic brain and hoping for the best.
Now, before you run away screaming about mathematical frameworks, let me break this down with some fun analogies:
Category Theory is like the ultimate filing system for different types of AI architectures. Imagine if you could organize every possible AI design into neat little boxes and then figure out how to transform one box into another. That's essentially what we're doing—creating a cosmic library of AI designs organized by their philosophical assumptions.
Information Theory helps us measure how "surprised" different AI systems are by reality. If your AI system keeps getting shocked by how chemistry works, maybe it's not thinking about chemistry correctly. We can actually measure this surprise and use it to build better systems.
Topology is like studying the shape of knowledge itself. Some knowledge is bumpy, some is smooth, some has holes in it. By understanding these shapes, we can build AI systems that navigate knowledge more effectively.
Here's where things get really exciting. Based on this framework, I'm predicting some pretty dramatic improvements:
For example, imagine an AI system designed for physics that actually respects conservation laws while it's learning. Not only should it perform better, but it might discover new quantum phases that we've never seen before. Or picture a chemistry AI that thinks hierarchically—it might design catalysts that human chemists would never think of.
This isn't just philosophical hand-waving. I've built actual software that can:
It's like having a philosophical consultant for your AI systems, except the consultant is also a mathematician, a computer scientist, and probably drinks way too much coffee.
If this actually works (and I'm pretty confident it will), we're looking at some science fiction-level implications:
Meta-Ontological AGI: Imagine AI systems that can figure out the right way to think about any new domain they encounter. They'd be like philosophical chameleons, adapting their worldview to match whatever reality they're trying to understand.
Scientific Discovery on Steroids: When AI systems actually understand domains the way human experts do, they might discover things that are completely invisible to current approaches. We're talking potential breakthroughs in climate science, drug discovery, materials design—the works.
AI That Actually Understands: Instead of just matching patterns, these systems might develop something approaching genuine understanding. They'd know not just what patterns exist, but why they exist and what they mean.
Here's the thing: we're at a crossroads in AI development. We can keep building bigger and bigger pattern-matching machines, or we can start building systems that actually understand the world they're trying to model.
The difference is like the difference between a parrot that can recite Shakespeare and a scholar who understands what Shakespeare means. Both might give you the right answer, but only one actually knows what they're talking about.
This ontological approach to AI isn't just about building better systems—it's about building systems that think about reality the way reality actually works. And if that doesn't get you excited about the future of artificial intelligence, I don't know what will.
The revolution in AI isn't just about making machines smarter. It's about making machines that actually understand what they're being smart about. And honestly? I think that's going to change everything.
Want to dive deeper into the technical details? Check out my full research paper or explore the open-source software framework. Warning: contains significant amounts of mathematics, philosophy, and possibly reality-bending insights.
