131–140 of 187 posts
I'm a former research mathematician who worked for a little while in AI research, and this article matched up very well with my own experience with this particular cultural divide. Since I've spent a lot more time in the math world than the AI world, it's very natural for me to see this divide from the mathematicians' perspective, and I definitely agree that a lot of the people I've talked to on the other side of thi…
If AI-generated proofs become incomprehensible to humans, do they still count as -math- in the traditional sense?
As Feynman once said [0]: "Physics is like sex. Sure, it may give some practical results, but that's not why we do it." I don't think it's any different for mathematics, programming, a lot of engineering, etc. I can see a day might come when we (research mathematicians, math professors, etc) might not exist as a profession anymore, but there will continue to be mathematicians. What we'll do to make a living when that…
This is written in the first person, but there is no listed author and the website does not suggest an author...
Earlier quoted context omitted.
Serious theorem-proving AIs always write the proof in a formal syntax where it is possible to check that the proof is correct without issue. The most popular such formal language is Lean, but there are many others. It's just like having a coding AI, it may write some function and you check if it compiles. If the AI writes a program/proof in Lean, it will only compile if the proof is correct. Checking the correctness…
> Checking the correctness of proofs is a much easier problem than coming up with the proof in the first place. Just so this isn't misunderstood, not so much cutting-edge math is presently possible to code in lean. The famous exceptions (such as the results by Clausen-Scholze and Gowers-Green-Manners-Tao) have special characteristics which make them much more ground-level and easier to code in lean. What's true is th…
The main bottleneck is having the libraries that define the theorems and objects you need to operate at those levels. Everything is founded on axiomatic foundations and you need to build all of maths on top of that. Projects like mathlib are getting us there but it is a massive undertaking.
It’s not just that it is a lot of maths to go through, it’s also that most maths has not really been proven to this degree of exactitude and there is much gap-filling to do when trying to translate existing proofs, or the reasoning style might be quite distant to how things are expressed in Lean. Some maths fields are also self-consistent islands that haven’t been yet connected to the common axiomatic foundations, and linking them is a serious research endeavor.
Although Lean does allow you to declare theorems as axioms. It is not common practice, but you can skip high up the abstraction ladder and set up a foundation up there if you are confident those theorems are correct. But still defining those mathematical objects can be quite hard on its own, even if you skip the proving.
Anyways, the complexity of the Lean language itself doesn’t help either. The mode of thinking you need to have to operate it is much closer to programming than maths, and for those that think that the Rust borrow-checker is a pain, this is an order of magnitude more complex.
Lean was a significant improvement in ergonomics compared to the previous generation (Coq, Isabelle, Agda…), but still I think there is a lot of work to be done to make it mathematician-friendly.
Most reinforcement-learning AI for maths right now is focused on olympiad problems, hard but quite low in the maths abstraction ladder. Often they don’t even create a proof, they just solve problems that end with an exact result and you just check that. Perhaps the reasoning was incorrect, but if you do it for enough problems you can be confident that it is not just guessing.
On the other side of the spectrum you have mathematicians like Tao just using ChatGPT for brainstorming. It might not be great at complex reasoning, but it has a much wider memory than you do and it can remind you of mathematical tools and techniques that could be useful.
I'm a former research mathematician who worked for a little while in AI research, and this article matched up very well with my own experience with this particular cultural divide. Since I've spent a lot more time in the math world than the AI world, it's very natural for me to see this divide from the mathematicians' perspective, and I definitely agree that a lot of the people I've talked to on the other side of thi…
taking a helicopter to the top of a mountain is not the same thing as climbing it
I'm a former research mathematician who worked for a little while in AI research, and this article matched up very well with my own experience with this particular cultural divide. Since I've spent a lot more time in the math world than the AI world, it's very natural for me to see this divide from the mathematicians' perspective, and I definitely agree that a lot of the people I've talked to on the other side of thi…
This algorithm will happily predict whatever it was fed with, just ask Chat GPT to write the review of non-existing camera, car or washing machine, you will receive nicely written list of advantages of such item, so what it does not exist.
Earlier quoted context omitted.
> Checking the correctness of proofs is a much easier problem than coming up with the proof in the first place. Just so this isn't misunderstood, not so much cutting-edge math is presently possible to code in lean. The famous exceptions (such as the results by Clausen-Scholze and Gowers-Green-Manners-Tao) have special characteristics which make them much more ground-level and easier to code in lean. What's true is th…
Yes I definitely concur, I have spent significant time with it. The main bottleneck is having the libraries that define the theorems and objects you need to operate at those levels. Everything is founded on axiomatic foundations and you need to build all of maths on top of that. Projects like mathlib are getting us there but it is a massive undertaking. It’s not just that it is a lot of maths to go through, it’s also…
Could you elaborate on this? I'm interested to learn what the complexities are (beyond the mathematical concepts themselves).
Is it really a culture divide or is it an economic incentives divide? Many AI researchers are mathematicians. Any theoretical AI research paper will typically be filled with eye-wateringly dense math. AI dissolves into math the closer you inspect it. It's math all the way down. What differs are the incentives. Math rewards openness because there's no real concept of a "competitive edge", you're incentivized to freely…
> Many AI researchers are mathematicians. Any theoretical AI research paper will typically be filled with eye-wateringly dense math. AI dissolves into math the closer you inspect it. It's math all the way down. There is a major caveat here. Most 'serious math' in AI papers is wrong and/or irrelevant! It's even the case for famous papers. Each lemma in Kingma and Ba's ADAM optimization paper is wrong, the geometry in…
Wrong in the strict formal sense or do you mean even wrong in “spirit”?
Physicists are well-known for using “physicist math” that isn’t formally correct but can easily be made as such in a rigorous sense with the help of a mathematician. Are you saying the papers of the AI community aren’t even correct “in spirit”?