OpenAI's 'Embarrassing' Math (techcrunch.com)
- Reference: 0179837062
- News link: https://science.slashdot.org/story/25/10/20/1827201/openais-embarrassing-math
- Source link: https://techcrunch.com/2025/10/19/openais-embarrassing-math/
> "Hoisted by their own GPTards." That's how Meta's Chief AI Scientist Yann LeCun [1]described the blowback after OpenAI researchers did a victory lap over GPT-5's supposed math breakthroughs. Google DeepMind CEO Demis Hassabis added, "this is embarrassing." The Decoder [2]reports that in a since-deleted tweet, OpenAI VP Kevin Weil declared that "GPT-5 found solutions to 10 (!) previously unsolved Erdos problems and made progress on 11 others." ("Erdos problems" are famous conjectures posed by mathematician Paul Erdos.)
>
> However, mathematician Thomas Bloom, who maintains the Erdos Problems website, said Weil's post was " [3]a dramatic misrepresentation " -- while these problems were indeed listed as "open" on Bloom's website, he said that only means, "I personally am unaware of a paper which solves it." In other words, it's not accurate to claim GPT-5 was able to solve previously unsolved problems. Instead, Bloom wrote, "GPT-5 found references, which solved these problems, that I personally was unaware of."
[1] https://techcrunch.com/2025/10/19/openais-embarrassing-math/
[2] https://the-decoder.com/leading-openai-researcher-announced-a-gpt-5-math-breakthrough-that-never-happened/
[3] https://x.com/thomasfbloom/status/1979254235075059732
LLMs are a red herring (Score:2)
Physics will always win.
A better search engine, or... (Score:2)
That he didn't bother to search, or...
The volume of garbage that AI spews out has made traditional search engines less useful, or...
The people who actually solved the problems did not bother to contact the guy who tracks them.
Quite possibly all these "or" should be replaces with "and".
Great Name! (Score:2)
GPTards, I love it! I'll be using that name to refer to all ChatGPT users going forward.
Er ... (Score:2)
... does sound kinda useful?
Brilliant Scientist Presents His Creation (Score:2)
From Young Frankenstein:
[1]https://www.youtube.com/watch?... [youtube.com]
[1] https://www.youtube.com/watch?v=_IUv5U37Ilk
What's unfortunate here (Score:2)
What's really unfortunate here is that due to OpenAI's drastic exaggeration of what happened here it distracts from the real capabilities here. Being able to efficiently find sources in the literature is an incredibly useful tool. And even aside from that there are now multiple examples where professional mathematicians have used GPT-5 in the thinking mode to make progress on math problems. Nothing as major as any Erdos problem, but still clear use. Terry Tao for example used GPT-5 in thinking mode to help
it's not doing math (Score:2)
It's hallucinating solutions
Re: (Score:2)
No, it's not hallucinating, it's showing math from the reference data
"GPT-5 found refs, which solved these problems" (Score:5, Interesting)
As a language models, GPT-5 finding references that solved Erdos problems that the professional mathematician and maintainer of the Erdos problem website was not aware of, shows an actual useful skill for the LLM.
Re: (Score:2)
Agree. It's like saying "I could've known that". Yeah, but you didn't. Even if they are not at the level of Erdos proofs they might have seemed to claim, GPT claims are not "embarrassing", except perhaps to competitors.
Re: (Score:3, Insightful)
It's embarrassing that they claimed these problems were "previously unsolved".
Re: (Score:3)
The part that is embarrassing is that they didn't check their sources.
Re: (Score:2)
OpenAI's tweet was worded to suggest that GPT-5 found something new, when it "just" did a literature (or Google for Gen Z) search. If pulled that kind of stunt in a PHD thesis, it would be revoked post haste.
Re: (Score:3)
"... GPT claims are not "embarrassing", except perhaps to competitors."
Yes they are, the claims were utterly false.
Re: (Score:2)
Reminds me of a time I was at the Houston Art Museum. We're looking at a Jackson Pollack painting (abstract paint splatters). My friend remarked "that's awful, that's so easy, anyone can do that, I could do that!!" Then the lady just behind us says to him. "Yes, you COULD have done, but you DIDN'T, and that's whole point here" We laughed so hard...
splat (Score:2)
I've been there.
I was so frustrated that I taped my banana to the wall!
not nearly in the same realm (Score:1)
One is asserting that GPT-5 can do theoretical math at a level above practicing mathematicians. Other is saying that if you spend a hundred billion dollars digesting all of human knowledge you can build a better search engine.
A better search engine is not going to do anything of the things AI zealots are promising. We already know that ChatGPT is pretty good as a search engine. But that won't replace human labor; it will just help highly skilled workers who are able to precisely formulate the questions they
Re:"GPT-5 found refs, which solved these problems" (Score:4, Interesting)
It's useful, but not groundbreaking. The equivalent to a minimal spanning tree in directed graphs is called an optimal branching. The algorithm for that was originally known Edmonds' algorithm in most of the world after the discoverer. It was later discovered that Chu and Liu already published essentially the same idea two years earlier in a Chinese journal. Even now, it is still often referenced only by Edmonds' name. This happens often enough.
Re:"GPT-5 found refs, which solved these problems" (Score:4, Interesting)
I invented the Bresenham line algorithm, a while later I learned I was not the first. I'm sure that's been replayed a million times.
Gallo discovered HTLV-3, the virus that causes AIDS. He discovered it by stealing the virus from French researchers who isolated it first and made it available in good faith. So, you know, the details are important. Here, promoters of AI made claims that are utterly untrue, it would be interesting to know if they made these errors out of deceit or ignorance.
Re: (Score:2)
I absolutely agree. It's a major difference to find a publication somewhere that was unknown and discovering something independently. It's useful to be able to do the former, but vastly different from being able to do the latter. Not verifying what the computer did is the real embarrassment for OpenAI.
Re: (Score:2)
my thoughts exactly, and also proof that GPT-5 may actually be more intelligent that the a-holes that created it and promote it.