myrmepropagandistfuturebird@sauropods.win
Aug 20, 2026, 6:10 PM"AI solves Impossible Math Problem"
*looks inside*
"Mathematicians use machine learning to help find key counterexample in still unproven conjecture"
Replies
- myrmepropagandistfuturebird@sauropods.winAug 20, 2026, 6:11 PM
Hey at least glamorous theoretical stubborn old math problems are getting a little more attention these days since a lot of people are very excited if you'd be willing to say their software helped you crack it.
Yakyu Night OwlYakyuNightOwl@mastodon.worldAug 20, 2026, 6:14 PM- myrmepropagandistfuturebird@sauropods.winAug 20, 2026, 6:24 PM
Fortunately everyone was very normal about celebrating what these nice kids did and I have NO bad memories of this event.
- Yakyu Night OwlYakyuNightOwl@mastodon.worldAug 20, 2026, 6:26 PM
@futurebird Seeing them recognized and celebrated by the New Orleans City Council on a livestream is one of the good memories I'll keep.
prom™️promovicz@chaos.socialAug 20, 2026, 6:11 PM@futurebird Mathematics, proving that it used to be so biased for internal and external reasons, that a random squirrel can make serious progress by banging buttons on a computer algebra system?
David NashdpnashAug 20, 2026, 6:20 PMWhat’s even more frustrating to me is that it’s looking like there may be *some* utility in letting a state-of-the-art LLM churn out some putative mathematical proofs that can be run through theorem-proving tools like Lean, but the hype level here is *utterly ridiculous*:
Mathematician proves Theorem X with their own understanding of math: “Mathematicians prove Theorem X!”
Same, but add symbolic computing tools like Wolfram Alpha: “ Mathematicians prove Theorem X!”
Same, but add proof checking software, which was the stuff of fantasy till quite recently: “Mathematicians prove Theorem X!”
Same, but add theorem proving software like Lean, likewise the stuff of fantasy till quite recently: “Mathematicians prove Theorem X!”
Same, but add an LLM: “AI proves Theorem X! Crisis in mathematics! Do we even need mathematicians anymore?”
- myrmepropagandistfuturebird@sauropods.winAug 20, 2026, 6:22 PM
Baby, we've known this since they cracked the 4 color problem. Of course these tools can be useful. But someone is thinking "we don't need mathematicians anymore now that we have AI" and ... yeah.
Nemoiinavpov@mastodon.onlineAug 20, 2026, 8:28 PM@futurebird
The thing is. 4 colour is proven. fine. Whatever, it's somewhat useful. You can use it in other proofs.At some point, someone will give up with an amazing insight as to *why* which will help us all.
Later, someone will give a way to *explain* that insight. And that's when maths will have moved forward truly.
ShadSterlingShadSterling@mastodon.socialAug 20, 2026, 7:44 PM@dpnash @futurebird they’re so desperate to be able to claim “AI” can do something valuable, and most people don’t understand math well enough to be able to tell that “AI” can only generate proofs extremely inefficiently
Assimilateborgassimilateborg@kind.socialAug 20, 2026, 8:10 PM@futurebird it's the same thing as LLM is good at bullshitting code. It bullshits so long until the code breaks and you found a bug. In maths it bullshits so long until a counterexample is found. And code bugs or counterexamples are easy to prove that they work.
Doing real maths or code is hardly possible with LLM, because you cannot prove easily that the math or code is correct.
Risa Spirellirisa_spirelli@nerdculture.deAug 20, 2026, 9:03 PM@assimilateborg @futurebird the thing about maths in contrast to most other forms of knowledge is that in fact you can. that's why it seems finding mathematical proofs is a thing LLMs with the help of proof checkers like Lean are very much able to do.
- Risa Spirellirisa_spirelli@nerdculture.deAug 20, 2026, 8:43 PM
@futurebird there have been proofs by LLM + Lean without human assistance recently, the funny maths youtuber made a video on it https://www.youtube.com/watch?v=ZJ8KThKAfbs
- Risa Spirellirisa_spirelli@nerdculture.deAug 20, 2026, 8:48 PM
@futurebird in the video he also quotes Terence Tao who's here on mastodon and apparently writes a lot about maths and AI
Nazonazokiyoubinbou@urusai.socialAug 20, 2026, 9:26 PM@futurebird It's really weird for me to see them use a LLM for anything to do with math because it literally — by definition — can't do math. It can try to approximate "statistically speaking, this probably follows that question" and be trained to raise the statistics such that it's usually right, but it still can't actually perform the math because that's fundamentally just not how it works... Or, more fundamentally, they can't do logic for the same reason — their methodology of logic is "this statistically seems to follow that a lot in the training data, thus I output it" not "ah hah, if this is true, then that must be true."
Thus they can only be used to pretend at math... They might get it right but if they do it's by luck, not by actually applying logic or mathematics.
Red_Shirt_no2Red_Shirt_no2Aug 20, 2026, 9:39 PM@futurebird @nazokiyoubinbou
Agentic AI uses tools (proof testers) that _do_ apply logic to test proofs generated by LLMs’ pattern-regurgitation capability. It seems to me this isn’t so far off from what people do; we try imposing patterns we’ve experienced on a problem, test the result, and iterate until we find a pattern that passes our tests. We rarely _start_ with logic; usually we start by choosing (somewhat randomly) how to look at a problem.- Nazonazokiyoubinbou@urusai.socialAug 20, 2026, 9:48 PM
@Red_Shirt_no2 @futurebird Actual logic isn't applying a pattern though. I'm not good at writing up word definitions (but I mean, you could just look at the dictionary) but "this statistically seems probably in a response" is not logic by any definition whatsoever. It's a facsimile of results.
And don't get me wrong. They can facsimile the process too. I get that. But again, they're not actually applying "if this, then that" reasoning. They're applying "to the statement if this, then that, x has a statistically higher probability of being what is in the training data."
For example, if you gave them something they weren't trained for sufficiently they wouldn't know how to solve it. Essentially "guessing." In some conditions you can even produce a 100% failure rate!
- Red_Shirt_no2Red_Shirt_no2Aug 21, 2026, 1:41 PM
@nazokiyoubinbou @futurebird
Yeah. What I’m saying is that “agentic AI” isn’t just an LLM making guesses; it has a deterministic framework _around_ the LLM that _does_ apply logic (deterministic programs) to test what the LLM spits out. The LLM hallucinates a response, the framework fact-checks it & feeds back errors.The LLM is not reasoning. The system seems to do something like what we do when we think we’re reasoning.
- Nazonazokiyoubinbou@urusai.socialAug 21, 2026, 7:08 PM
@Red_Shirt_no2 @futurebird I mean, even then it's not really. I suppose external tooling may have a check mechanism or two that applies computer logic, but they still lack understanding. The logic is a simple gate "is x true? Then y" type of thing. (Yes it gets more complex, but in other words it's not understanding, it's measurements.)
As for the "reasoning" mechanism, the LLM still is just applying statistical probabilities, just it has a training to produce a sort of double check to increase the likelihood of producing a more correct statistical probability. This helps reduce the effects of the randomization of the token production that produces what people call hallucinations without having to lower the temperature too much, bit it still lacks understanding or actual logic.
- Nazonazokiyoubinbou@urusai.socialAug 21, 2026, 7:12 PM
@Red_Shirt_no2 @futurebird As a side note, you say it looks like our thought process, but I guess you haven't actually seen the "thinking"? Local models that do it don't hide it from the user entirely. (It goes inside an HTML collapse.) The "thought" process goes something like this:
The user has asked us so-and-so. We should produce a response that fits within said criteria. First we should do x, then we should do y.
It's not really "thinking" of course. It's trained to produce an output that says "the user asked us so-and-so, we should so-and-so." That's just the training forcing that specific response first. But this raises the statistical chance of being able to produce a "wait, that isn't correct, I should say y instead of x" inside the hidden section.
- myrmepropagandistfuturebird@sauropods.winAug 21, 2026, 7:10 PM
@Red_Shirt_no2 @nazokiyoubinbou
I thought "agentic" just meant that you gave the LLM usernames and passwords to things like bank accounts and heating systems and instructed it to make a macro with a series of commands to be executed using those user accounts.
That is using a LLM to "do things" rather than just make text.
The additional systems that make this possible don't really change the fundamental model all that much, they just vastly expand the potential impact of any mistakes.
- Nazonazokiyoubinbou@urusai.socialAug 21, 2026, 7:13 PM
@futurebird @Red_Shirt_no2 Yeah, basically it's just sending specific outputs to tools designed to do the actual things. The tools still lack understanding in the same way notepad doesn't understand what you're doing when you type, just that it must produce text onto the screen as you type. They have specific, very narrow tasks that the API calls do.
They just train it to output the correct things to those APIs (most of the time...)
Aug 20, 2026, 10:11 PMThere was no link to any article so I’m not sure, but did they use an LLM? What I mean is there’s a big difference between “ask Chatgpt” and training a machine learning model for a specific problem. But these days both might be headlined under “use AI”.
The other day I saw an article about how “AI plants trees to help reforestation”. The project was about drones and yes, they were automated. In general terms it probably wasn’t wrong to use the acronym, but I hate how due to recent use of the word, the headline then reads like “chatgpt plants trees!” and the buzzword AI takes the credit when the system in use wasn’t actually the plagiarism-into-slop machine.
@futurebird@sauropods.win @nazokiyoubinbou@urusai.social- Nazonazokiyoubinbou@urusai.socialAug 20, 2026, 10:17 PM
@onion @futurebird Yes, this is a good point. The term "AI" has been falsely slapped onto things that don't even make sense for a very very long time now. Often enough it just means "really complex algorithms specifically designed for a specific purpose." (Which is kind of as not actually AI as anything can possibly be not actually AI since AI, by definition, must be more general.) It is frequently used as just a buzzword or marketing hype...
So "they used 'AI'" could very well just mean they have specially designed algorithms for the task, in which case there is a whole different meaning here...
As a side note, this is why I frigging hate everyone just letting these people just call everything "AI." It should be called what it is and the scam should not be perpetrated.
GabsGabs@mastodonapp.ukAug 20, 2026, 9:52 PM@futurebird ah, that makes way more sense
Aug 21, 2026, 5:19 AM@futurebird for real this time, or did the LLM exploit a bug in the prover software again?