The use of computers in mathematics has been somewhat controversial from the very start.
There are of course all the computer-assisted proofs (see 4 color theorem), as well as the partially-assisted ones (see Viazovska et al on packing problems in dimensions 8, 24). But even finding a solution numerically, then rigorously verifying its properties can leave a lingering sense of incompleteness, of a gap in understanding. I like this one quote by (allegedly) Wigner that illustrates it well:
"It is nice to know that the computer understands the problem, but I would like to understand the problem, too."
Tests only work for a limited set of programming verification. In many cases you don’t actually know what the output for any given input should be, so there’s no way of verifying the AI-generated code. You just kind of have to trust it. The only exception I can think of is robotics and quantitative trading. Which have already been extensively utilizing AI.
So… more peer review backlog. That sounds fun. Oh, you want someone to review your paper, Mr phd in mathematics with 20 years of experience? Get in line behind chatGPT.
Human mathematicians could become “priests to oracles.”
Priests were interpreting the oracles (at least at a place like Delphi) according to the context of the people asking the questions aka participating in politics of that ancient times.
Subjectivity was a feature and I’m not sure that fits to mathematics though.
I wonder if mathematics as a science field moves more into engineering or if a different branch will emerge that is closer to that because to the point of the article, science is about understanding not just results.
It’s a well known problem in higher mathematics that even if you’ve solved a problem, often the proofs are incredibly long and complex and require an extensive amount of time spent by peers to review it.
It would be great if someone could explain to me how AI improves this situation. Even if AI thinks it’s solved a problem, unless the proof is incredibly efficient and well explained, it will be difficult to verify the correctness. One hallucination in 300 steps of logic is enough to destroy the entire proof.
The article poses if AI will be a tool, a collaborator or an oracle. Why not all 3?
If mathematics is human understanding of logical consequences, understanding is the priority. But if AI proves something we can't understand but can utilize, that is a different sort of useful.
We are getting awfully close to "the answer of the universe is 42" and having it not be a joke...
I don’t know about “close”,
but there are certainly results in math that are considered deep because they require the use of a “Hard Theorem” at some point. That kind of building on top of something Very Difficult is still possible without understanding the “Very Difficult” part. I’d say a lot of not-amazing math is built by believing the platform works but not being able to built it yourself.
I couldn’t build an internal combustion engine or even a plastic box, so maybe there’s nothing wrong with this approach.
It's amazing how much attention this issue has gotten. What is lost in the hype is no AI can tell you if a proof is correct. An AI can produce a convincing looking proof, but it can have a subtle but critical error or make an assumption that is unfounded. Thus, it ultimately comes down to humans. A mathematician has to craft the prompt, and mathematician to interpret/check the results. Also, these programs are very expensive and propitiatory. They are not like the commercial AI that regular people use. It takes considerable prompting and trial an error to solve even Olympiad/Putnam problems, and tons of work by humans pouring over the results to see if it's correct. For every Erdos problem that captures the headlines, there are many where it failed or untold hours of prompting and token burn to get that result, and manhours verify it.
AI can't yet come up with any new ideas to make the inductive leap to solve a math problem. New ideas are what get the accolades and using an old idea just means the original author missed something. We are still at the author missed something stage that AI is doing today.
It can definitely be a good research assistant though
I don't think you understand the workflow. Terrence Tao has done a lot of work using them in conjunction with LEAN.
You aren't having the AI check the proof, you interactively work on the same LEAN proof, handing off between the AI assistant and having LEAN check it and provide feedback for both of you when there's a mistake.
There are of course all the computer-assisted proofs (see 4 color theorem), as well as the partially-assisted ones (see Viazovska et al on packing problems in dimensions 8, 24). But even finding a solution numerically, then rigorously verifying its properties can leave a lingering sense of incompleteness, of a gap in understanding. I like this one quote by (allegedly) Wigner that illustrates it well:
"It is nice to know that the computer understands the problem, but I would like to understand the problem, too."
If you don’t understand the problem you can’t be sure that the computer does.
But I also definitely don't understand the problem if I can't get the computer to understand it, with tests.
In some sense I always considered programming to be more trustworthy than maths arguments without the certainty of a solver proof.
With all of these questions in the air, epistemology might be making a comeback.
Priests were interpreting the oracles (at least at a place like Delphi) according to the context of the people asking the questions aka participating in politics of that ancient times.
Subjectivity was a feature and I’m not sure that fits to mathematics though.
I wonder if mathematics as a science field moves more into engineering or if a different branch will emerge that is closer to that because to the point of the article, science is about understanding not just results.
This is a decidedly anti-enlightenment statement.
It would be great if someone could explain to me how AI improves this situation. Even if AI thinks it’s solved a problem, unless the proof is incredibly efficient and well explained, it will be difficult to verify the correctness. One hallucination in 300 steps of logic is enough to destroy the entire proof.
i.e. You have to be a good engineer to understand the well generated LLM code and a program
Would some one with tokens to burn mind checking that statement out and post back. Be sure to use long dashes too.
If mathematics is human understanding of logical consequences, understanding is the priority. But if AI proves something we can't understand but can utilize, that is a different sort of useful.
We are getting awfully close to "the answer of the universe is 42" and having it not be a joke...
I couldn’t build an internal combustion engine or even a plastic box, so maybe there’s nothing wrong with this approach.
It can definitely be a good research assistant though
You aren't having the AI check the proof, you interactively work on the same LEAN proof, handing off between the AI assistant and having LEAN check it and provide feedback for both of you when there's a mistake.
(edit: lol didn't realize the sibling comment below is essentially my comment)