The senses, although they are necessary for all our actual knowledge, are not sufficient to give us the whole of it, since the senses never give anything but instances… From which it appears that necessary truths, such as we find in pure mathematics, and particularly in arithmetic and geometry, must have principles whose proof does not depend on instances, nor consequently on the testimony of the senses, although without the senses it would never have occurred to us to think of them…
- Leibniz: Philosophical Writings

When trying to build AI, eventually you run into this choice between empiricism and rationalism. Where does knowledge come from? What is logic? How do we learn?

To me, this is a practical question rather than a philosophical one. How do I build a machine that learns and reasons like a human?

To me, accepting rationalism (like Leibniz, above) is a non-starter for general AI because it concedes that human thought is extra-ordinary. And if it’s extra-ordinary, then we can’t build a machine to simulate it out of ordinary stuff like silicon.

And so thought can’t be abstract. Thought has to be physical, or at least empirically observable. Thought has to follow from physical materials interacting with the same rules as other physical systems. That assumption lets us pursue the construction of a general AI, within the framework of physics. (A single theory of everything.)

To most people, this means thought == brain, and they’re fine with that. Thought is just neurons and biochemical reactions, governed by the same electromagnetism etc. that exists outside of the body.

But if (I assume) thought is physical, then (I have to assume) everything else is physical. Specifically, mathematics and logic. Again, this is because to build general AI, I have to understand how to implement these things in my machine. They have to be empirically measurable; they’re not allowed to be hand-wavy and abstract.


All our knowledge begins with the senses, proceeds then to the understanding, and ends with reason. There is nothing higher than reason.
- Immanuel Kant, Critique of Pure Reason

But how do you define math and logic operationally, in terms of base senses? After all, the point is that “math is true.” People believe in math the same way they believe in God: it exists on some higher-plane of existence, and it’s authoritative. You start with unquestionable axioms and proceed via unquestionable logic. Adding empricial observations is just corrupting the pureness of it.

Now, I’m not saying that “math isn’t true.” There are certainly regularities in the universe: gravity, relativity, electromagnetism. Linear systems exist and are predictable. Arithmetic is useful in predicting outcomes of practical daily situations.

The caveat is that discovering and codifying these regularities is carried out by humans, who are messy and error-prone. And that’s what I mean by an empirical approach to math: viewing the people as the physical system under study.

A mathematician’s work is mostly a tangle of guesswork, analogy, wishful thinking and frustration, and proof, far from being the core of discovery, is more often than not a way of making sure that our minds are not playing tricks.
- Gian-Carlo Rota

So I view math as a natural process, involving humans (or constructed robots), which resembles a distributed consensus algorithm. Many people try math and publish proofs. Only the ones that are accepted by the majority pass the filter to become “accepted math.” This lets us carry on a single line of reasoning over the millenia.

This is the essence of math: we establish a common language (axioms), we reason about those axioms, and then we build trust in that reasoning via distributed proof-checking.


The sciences do not try to explain, they hardly even try to interpret, they mainly make models. By a model is meant a mathematical construct which, with the addition of certain verbal interpretations, describes observed phenomena. The justification of such a mathematical construct is solely and precisely that it is expected to work.

- John Von Neumann

But how should my “thinking machine” construct axioms or apply proven theorems to the real world? What is the observable mechanism behind this?

I think about “metaphor” as the building block for this operation. Specifically, metaphor as described by Lakoff and Johnson, which I’ve talked about before.

In their model, everything starts with basic senses. And as we live, we learn to associate certain experiences with other past experiences. Argument is war, time is money, 3 gold coins is a sheep. That’s metaphor. And they stack to build more complicated metaphors.

And to me, the reason we developed this faculty was so we can predict the future better. “He sank like a lead weight.” You’ve never seen him sink before, but you’ve seen a lead weight sink. It’s like that.

A reasonable physical implementation of this in my machine looks like a neural network. Of course, the details are fuzzy, but that’s what the whole field of ML is about. How does one look at an image and decide if it looks like a dog or a cat? In other words, how does a machine build abstract metaphor?

Getting back to math (and I’ve made this point before): the whole endeavor can be thought of as building metahpors between real world-scenarios and axiomatic symbols on the page. The symbols ought to change via rules that maintain the metaphor. Then you can use the theorems to make predictions about your real-world system.

Mathematics is the study of analogies between analogies. All science is. Scientists want to show that things that don’t look alike are really the same. That is one of their innermost Freudian motivations. In fact, that is what we mean by understanding.
- Gian-Carlo Rota

Mathematics compares the most diverse phenomena and discovers the secret analogies that unite them.
- Jean Baptiste Joseph Fourier

To be useful, the system ought to relate via metaphor to a high variety of physical systems. One way to do this is to keep your axioms really simple (like things in a set) so that they apply whenever someone can recognize countable “things” that may or may not be in a collection. Another way is to discover symbols that helps you predict something about literally everything (E=mc^2).

But again, the application of these axioms in my “thinking machine” depends on some physical neural networks, which have been trained through experience. The metaphors employed during reasoning also depend on experience. I suspect people are not different. Math doesn’t happen in a vacuum. No matter how “pure” and “right” you think math is, eventually you need a messy brain to pattern-match real-world systems to axioms, and back again.

Except for pure math, I suppose. And I can’t entirely disregard it, because even negative numbers and complex numbers were “pure math” at some point. Paul Lockhart would say they were “extended” via symmetry and then only later reified to concrete domains (like debts and electrical circuits). It’s curious that our aesthetic sense of symmetry should have any relation at all to what happens in the world. There’s probably deeper truth here, but I can’t put my finger on it.


I’m an empiricist. I’ll likely live and die and empiricist. Feel free to put “empricist” on my tombstone.

And this has changed how I think about math. The whole thing, to me, is pattern-matching via metaphor which exploits regularities in the universe. Math is only possible because for some reason, these symbols , when shuffled around correctly, behave exactly like a ball falling through the air.

It’s wild stuff.

And when viewed as less-than perfect art, the short-comings of math become apparent. Some math is wrong. (Probably not linear algebra, though.) Some math is beautiful (in the eyes of the community) but not useful. (It does not relate via metaphor to any real-world phenomenon which we can make predictions about.)

There is not one “right” way to do math. There are many symbols to choose and many ways to prove a theorem. The important thing is that the metaphor is maintained. The symbols must refelect some regularity in the universe. And it has to be beautiful.

Therefore psychologically we must keep all the theories in our heads, and every theoretical physicist who is any good knows six or seven different theoretical representations for exactly the same physics.
- Richard Feynman

And in the end, the kind of math we create and will accept as beatiful and true is limited by the physical constraints of our wetware. There might be regularities in the universe that are universal but too complicated for us to store in our brains.

However, all this requires abandoning the conceit that humans are special. If you are willing to propose that you and I are fundamentally different from the robot I am building out of stones and such (or even the people whose heads we’ve opened up and looked inside), feel free to ignore everything I’ve said.

- Mitchell