This seems to be the mathematical equivalent of using AI to submit 200 unverified pull requests to an open source project and then telling the maintainers it’s their job to figure it all out.
The one mathematician saying he’s not going to spend hours of his time to verify if a slop report is true, let alone do it for dozens of reports, reminds me of all those projects making rules that unverified AI pull requests will be trashed.
I think OpenAI has retracted some of them already. For OpenAI it’s a really low-risk thing: if it’s wrong, then just retract the paper. If it’s right though, the fame all goes to OpenAI. Meanwhile, people who spent their whole life researching on this topic, needs to confirm this manually for OpenAI.
It’s a DDoS attack on human capacity, effort and attention, always has been.
I mean, this is a step worse actually. We’ve already seen mathematicians claiming that these systems have actively scooped them. Lots of academics are using these systems regularly.
At this point, every time I see a paper being “published” by an AI company, I’m wondering who they stole the result from.
Also makes me wonder if it found and exploited (or got caught out by) some bugs in the Lean programming language…
(Not saying Lean is buggy, but finding bugs seems more likely to me, as a programmer who knows not much about mathematical proofs since I haven’t looked at anything like that since uni, about a decade ago)
deleted by creator
At least half of these are formally verified. Although they did retract 3 papers. Out of about 700
No they aren’t, about 40% of them claim to have been self-verified. But when actual mathematicians looks at them, it isn’t even clear if the “proof” included is proving the thing the paper claims to prove. It all has to be looked at with great scrutiny to find out if any of it even has merit.
OpenAI just dropped a bunch of busy work on the entire field of mathematics that may or may not turn into anything at all…
Even worse, the software-based proofs are not the same as the plain text ones that they’re giving out.
There’s literally no reason to trust them, because it’s completely divorced from the actual text.
Saying “self-verified” is massively downplaying it. They are formalized and checked in Lean4 1, which is a programming language used by mathematicians today, to mechanically check their proofs to rule out human error. In other words a theory being stated in Lean generally means it’s more likely to be correct than one stated in mere human language.
Now, Lean, like any piece of software, has had bugs. A while back someone exploited a Lean bug to “prove” the Collatz conjecture 2. So it is possible that AI agents found a bug and used it. But the vibes I got from mathematicians in the field is that that’s not very likely.
It’s interesting that the one person who actually seems to know what they’re talking about is the one getting downvoted. AI is bad at many things for many reasons, but that doesn’t mean we should just assume that anything derived from AI is automatically slop.
Just throw that garbage in the bin, where it belongs.
This is the paradox of AI… generate lots of plausibly good sounding material, but you need a human to spend loads of time validating it. Software is already going through this issue.
Then you’re getting into the infinite monkey theorem, and one important part is that for the random monkeys to create works of art like Shakespeare you also need a reviewer to determine when the output is interesting, in that case you are really shifting the burden of creativity onto the reviewer
I’m confused. Did OpenAI release an album? What did they drop? The article is unclear.
A series of academic papers spanning across multiple mathmatic fields.
yes. let’s trust the machine that can’t count to 50.
You don’t have to. That’s how math works.
It’s like firing 5,000 gallons of raw sewage against a wall and then picking out the pieces of undigested sweetcorn to put in your soup.
Are you claiming the corn would still be sweet? Citation needed.
One of the main fucking points of a mathematical proof is to be clear and convincing. If it is not convincing and reducible to simple steps, it’s not a proof at all.
All a proof needs to do is logically derive the conclusion from the premises. Sure, there are qualities we like to see, but complexity is no disqualifier. These AI proofs are of course horrendous and need a lot of work, but even a bad proof can allow new work to continue on top of it.
There’s difference between complexity and unreadability
I’m confused whats in the papers? Anything important?
- L=RL=BPL
- multiplying two numbers can be faster than nlogn
- matrix multiplication in n^2.25
3sum is subquadratic- Hilbert tenth problem over rationals
- Unique game conjecture - now theorem
To name but a few. Each of these alone can be ground breaking.
Edit: sorry, the 3sum result was not from this batch. it happened around the same time so it got jumbled in my brain, but this one was proved by Claude.
Thank you. So why so little info out?
my guess is that probably these problems aren’t that well known to the general public? unlike Navier-Stokes, etc.
if the likes of Riemann Hypothesis, P vs NP, etc. got proven i bet everyone will hear about it immediately.
