Every time a colleague tells me they got ChatGPT or Claude or … to write a code to interpret some binary data file, it turns out that what the LLM actually did is find a code that some human had posted to GitHub (or maybe a couple of codes, and then averaged them). I have yet to see an example where the LLM actually produced a data interpreter “on its own”. It’s not without value, but, as you say, it “isn’t too big a deal”.
What fields medal mathematician was this? It sounds like Terence Tao, he has said similar things in the past.
I think when OpenAI’s o1 model came out in 2024, Tao compared it to a ‘mediocre but not completely incompetent grad student’. In 2026 he uses it as a research assistant to give him new ideas
What does the advent of powerful AI models mean for mathematicians like me? I have yet to succeed in prompting a solution to one of my long-term problems, though AI models have helped fix an incorrect lemma and suggested a helpful reorganization of a paper. They have pointed me in directions in which I had not thought to look, and spurred me to learn new mathematics. Combining AI with the skills I have honed over the years is making me a stronger mathematician.
However I think the issue is that (and I could be wrong about this) that current AI models are only as good as the training fed into them. So they can learn existing math, but they can’t really generate a lot of novel math. It’ll take new AI models to take existing math and generate novel math.
I may be a Luddite and I am not a professional mathematician but I won’t accept OpenAI’s claims until they have been independently verified by actual mathematicians. And with 700-ish claims, that will take some time.
As you noted in an earlier post, some problems are just boring and uninteresting to solve, even with sufficient computational power.
You and literally every person in the world. Even if it was a claim by a person it would be independently verified by a different mathematician before it was accepted. That’s how it works and no one is trying to say otherwise. That’s the literal reason why they are publishing it.
I’ve seen it right here on the Dope. Someone came up with a silly puzzle of some sort (I think it might have involved searching a large dictionary for words with a particular letter pattern, or looking for numbers with a particular property), and a Doper fed the problem description to an LLM and got code that ran and accurately found the targets. It’s possible that someone else had already posed that problem and implemented code for it, but I doubt it.
FATIIW, the AI might have fallen back on existing code that was not human, but also AI generated (an human “curated”) …
THAT very scenario will be more and more frequent in the future.
The title of the thread says otherwise.
I’ve just seen a paper - it hasn’t been published yet, but is accepted - about LLMs replicating the work in a paper by reading the paper, reading the associated data, and sometimes writing code used for the research but not available using the flowchart and description in the paper. I don’t think it just put things together.
Also, have you ever adapted big chunks of code? For my dissertation I rewrote the Jensen/Wirth Pascal compiler into one for my object-oriented Pascal-based microprogramming language. It is somewhat non-trivial. I’ve also repurposed a logic simulator into a fault simulator when I was in Bell Labs.
As for Enigma, all I know was what I read in the internet. It appears there are some messages without any of the tells that helped Bletchley Park decode them. I have no idea of how much effort has been put into these.
There are tons of messages not yet decrypted. Most, if not all, can be brute forced but ultimately, that requires somebody care enough to put in the effort. This was something some of my friends considered a couple decades ago as sort of a ‘fun’ side project but ultimately didn’t (again, lots of effort and kind of boring, ultimately). As the example you posted shows, it’s hardly riveting stuff. Nothing super critical, especially most of a century after the events.
It only sounds impressive that an LLM did it to people who aren’t too familiar with the problem.
I described how Enigma machines worked in the chapter on Encryption in my book, so I am familiar. I’m not impressed by a simulator either - anyone could toss one off in no time. I am impressed that an LLM decided to write (or steal) one.
Going back to the OP, it would be like saying “big deal that the proofs were generated. Some of them were wrong.” Like mathematicians don’t get proofs wrong.
I didn’t mean to suggest that LLMs can’t (or won’t soon be able to) do it, but I’m getting tired of colleagues who excitedly declare that Claude can write all their instrument data parsers for them and their evidence is Claude went out and found a pre-existing human-written parser. Again, that is not without value, but it’s a far cry from feeding proprietary gibberish in one side and getting open data out the other.
Moderator Note
I have removed all of the links to archive . ph as this is a known malicious site and is not related to the legitimate archive.org.
Please do not link to this site in the future.
I’m not familiar with the post you mention, but this kind of pattern matching would not take any kind of special programming to find matches. For instance, if you wanted to find words which started with a and ended with z, a simple pattern match like “a*z” could be used. There are lots of different pattern matching languages. A common one used in Unix is called “regex”, which stands for “regular expression”. Regex has rich pattern matching rules that can find very complicated patterns. For instance, a regex pattern like “[abc].t.[xyz]” would mean “Find any word which started with one of the letters a, b or c, followed by any number of characters that contain at least one letter ‘t’, and ends with one of the letters x, y, or z”. The pattern can seem complex and confusing if you’re not used to regex, but it’s simple and easy to use once you know the rules. To find words to solve the puzzle, the LLM would just have to make up a regex pattern to find the right words and then run a regex search against a dictionary list.
Yes, I understand it takes 1-2 years for peer review which buys OpenAI some more time to hook investors. I really think this is their primary job: getting investors excited by any means necessary.
It’s not peer review. I’ve never done peer review for math, but I doubt the reviewer is expected to confirm the proof at the level other mathematicians would. A quibble, but people on all sides get what peer review is wrong, including treating it as final confirmation of the proof of something.
And I’d love to find investors excited about investing hundreds of millions of bucks on solving math problems.
Which begs the question… Will Real Mathematicians ™ in the future do mostly the dirty work of reviewing/signing off AI’s moments of glory?
I don’t know but as the parent of a math genius I do have a dog in this fight. I’m just not sure what the fight is, or what it will be 10 years from now.
What we can say, at least, is that whether or not these LLMs are doing good math is a matter of debate in the math community, as is whether or not the LLMs are actually stealing from mathematicians who have been working on these problems with LLMs.
It would be like if I (foolishly) developed a novel using an LLM and then asked it to edit my work. The LLM used those notes and snippets of my work to produce a novel. Who gets credit?
Let’s say that many writers don’t think the novel is very good, certainly nothing revolutionary. Now we’ve got the awesome novel I would have written, had I more time, alongside the mediocre novel AI wrote.
And the general public of course can’t tell the difference.
Sorry to stretch this metaphor but my world is writing, not math. It’s easy to take breathless media reporting at face value and assume yes, this AI sure did some fancy math. But laypeople cannot verify that. Only math professionals can. And I’m hearing this is pretty controversial.
Vi Hart just posted about Open AI’s plagiarism in the Navier Stokes proof.
I have no idea how easy it is to follow this link. I got an email as a patreon supporter:
Anyway, she describes it as a leopard-eating-face incident. Her post is long but worth a read.
I hate to be the pedantic person here, but LLMs are very good at LANGUAGE … and not much more … but they have smart friends that they can talk to and ask all kinds of stuff and then relay the anwer back to the user in very clear language.
(we - collectively - seem to fall prey to the fallance of a chatbot knowing all kinds of stuff - which it does not, it’s just the “one-face-to-the-customer”)