If it just found existing solutions then they obviously weren't "previously unsolved" so the tweet is wrong.
He clearly misunderstood the situation and jumped to the conclusion that GPT-5 had actually solved the problems because that's what he wanted to believe.
That said, the misunderstanding is understandable because the tweet he was responding to said they had been listed as "open", but solving unsolved erdos problems by itself would be such a big deal that he probably should have double checked it.
Note that “solved” does not equal “found”
But yes, as an edge case handler humans still have an edge.
It's not obvious to me that they're better at admitting their mistakes. Part of being good at admitting mistakes is recognizing when you haven't made one. That humans tend to lean too far in that direction shouldn't suggest that the right amount of that behavior is... less than zero.
they feed an internet data into that shit, they basically "told" LLM to behave because surprise surprise, human sometimes can be more nasty
(Yes, not everyone, but we do have some mechanisms to judge or encourage)
This claim is ambiguous. The use of the word "Humans" here obscures rather than clarifies the issue. Individual humans typically do not "hallucinate" constantly, especially not on the job. Any individual human who is as bad at their job as an LLM should indeed be replaced, by a more competent individual human, not by an equally incompetent LLM. This was true long before LLMs were invented.
In the movie "Bill and Ted's Excellent Adventure," the titular characters attempt to write a history report by asking questions of random strangers in a convenience store parking lot. This of course is ridiculous and more a reflection of the extreme laziness of Bill and Ted than anything else. Today, the lazy Bill and Ted would ask ChatGPT instead. It's equally ridiculous to defend the wild inaccuracy and hallucinations of LLMs by comparing them to average humans. It's not the job of humans to answer random questions on any subject.
Human subject matter experts are not perfect, but they’re much better than average and don’t hallucinate on their subjects. They also have accountability and paper trails, can be individually discounted for gross misconduct, unlike LLMs.
This is more and more clearly false. Humans get things wrong certainly, but the manner in which they get things wrong is just not comparable to how the LLMs get things wrong, beyond the most superficial comparison.
Worst case (more probable): Lying
Works for Elon.
Off topic, but I saw The Onion on sale in the magazine rack of Barnes and Noble last month.
For those who miss when it was a free rag in sidewalk newsstands, and don't want to pony up for a full subscription, this is an option.
But it's only a matter of time before AI gets better at prompt engineering.
/s?
The inevitable collapse could be even more devastating than the 2008 financial crisis.
All while so vast resources are being wasted on non-verifiable gen AI slob, while real approaches (neuro-symbolic like DeepMind's AlphaFold) are mostly ignored financially because they don't generate the quick stock market increases that hype does.
2008 was a systemic breakdown rippling through the foundations of the financial system.
It would lead to a market crash (80% of gains this year were big tech/AI) and likely a full recession in the US, but nothing nearly as dramatic as a global systemic crisis.
In contrast to the dot com bubble, the huge AI spending is also concentrated on relatively few companies, many with deep pockets from other revenue sources (Google, Meta, Microsoft, Oracle), and the others are mostly private companies that won't have massive impact on the stock market.
A sudden stop in AI craze would be hard for hardware companies and a few big AI only startups , but the financial fallout would be much more contained than either dot com or 2008.
An AI bust would take the stock price down a good deal, but the stock gains have been relatively moderate. Year on year: Microsoft +14%, Meta +24%, Google +40, Oracle +60%, ... And a notable chunk of those gains have indirectly come from the dollar devaluing.
Nvidia would be hit much harder of course.
There is a good amount of smaller AI startups, but a lot of the AI development is concentrated on the big dogs, it's not nearly as systemic as in dot com, where a lot of businesses went under completely.
And even with an AI freeze, there is plenty of value and usage there already that will not go away, but will keep expanding (AI chat, AI coding, etc) which will mitigate things.
Well, an enormous amount of debt is being raised and issued for AI and US economic growth is nearly entirely AI. Crypto bros showed the other day that they were leveraged to the hilt on coins and it wouldn't surprise me if people are the same way on AI. It is pretty heavily tied to the financial system at this point.
If the stock market crashes, there’s lots of talk about how wealth and debt are interlinked. Could the crash be general enough to start calls on debt backed by stocks.
My recollection in 2008 was that we didn’t learn about how bad it was until after. The tech companies have been so desperate for a win, I wonder if some of them are over their skis in some way, and if there are banks that are risking it all on AI. (We know for some tech bros think the bet on AI is a longtermist like bet; closer to religion than reason and that it’s worth risking everything because the payback could be in the hundreds of trillions)
Combine this with the fact that AI is like what - 30% of the US economy? Magnificent 7 are 60%?
What happens if sustainable PE ratios in tech collapse. Does it take out Tesla?
Maybe the contagion is just the impact on the US economy which, classically anyways has been intermingled with everything.
I would bet almost everything that there is some lie at the center of this thing that we aren’t really aware of yet.
The US admin has been (almost desperately) trying to prop up markets and an already struggling economy. If it wasn't AI, it could have been another industry.
I think AI is more of a sideshow in this context. The bigger story is the dollar losing its dominant position , money draining out into Gold/Silver/other stock markets, India buying oil from Russia in Yen, a global economy that has for years been propped up by government spending (US/China/Europe/...), large and lasting geopolitical power balance shifts, ...
These things don't happen overnight, and in fact over many years for USD holdings, but the effects will materialize.
Some of the above (dollar devaluation) is actually what the current admin wanted, which I would see as an admission of global shifts. We might see much larger changes to the whole financial system in the coming decades, which will have a lot of effects.
Nowhere close. US GDP is like $30 trillion. Open AI revenue is ~$4 billion. All the other AI companies revenue might amount to $10 billion at most, and that is being generous. $10 billion/ $30 trillaion is not even 1%.
You are forgetting all those "boring" sectors that form the basis of economies like agriculture and energy. They have always been bigger than the tech sector at any point, but they aren't "sexy" because there isn't the potential "exponential growth" that tech companies
The Open AI revenue was ~$4 billion for the first half of the year; Anthropic recently reported a rate (which isn't total revenue, I know) equivalent to about $10 billion/year; NVIDIA's sales are supposed to be up 78% this quarter due to AI sales, reaching $39.33 billion, so plausibly ($39.33/1.78)*0.78 ~= $17 billion from AI in that quarter (rate, again yes I know, of $68 billion/year). So I can believe AI is order-of $100 billion/year economically… to US businesses with customers almost everywhere important except possibly China.
But just to re-iterate, this doesn't change your point. Even 100 B / 30 T is only one third of a percent.
Due to exorbitant privilege, with the dollar as the only currency that matters, every country that trades with America is swapping goods and services for 'bits of green paper'. Unless buying oil from Russia, these bits of green paper are needed to buy oil. National currencies and the Euro might as well be casino chips, mere proxies for dollars.
Just last week the IMF issued a warning regarding AI stocks and the risk they pose to the global economy if promises are not delivered.
With every hiccup, whether that be the dot com boom, 2008 or the pandemic, the way out is to print more money, with this money going in at the top, for the banks, not the masses. This amounts to devaluation.
When the Ukraine crisis started, the Russian President stopped politely going along with Western capitalism and called the West out for printing too much money during the pandemic. Cut off from SWIFT and with many sanctions, Russia started trading in other currencies with BRICS partners. We are now at a stage of the game where the BRICS countries, of which there are many, already have a backup plan for when the next US financial catastrophe happens. They just won't use the dollar anymore. Note that currently, China doesn't want any dollars making it back to its own economy, since that would cause inflation. So they invest their dollars in Belt and Road initiatives, keeping those green bits of paper safely away from China. They don't even need exports to the USA or Europe since they have a vast home market to develop.
Note that Russia's reserve of dollars and euros was confiscated. They have nothing to lose so they aren't going to come back into the Western financial system.
Hence, you are right. A market crash won't be a global systematic crisis, it just means that Shanghai becomes the financial capital of the world, with no money printing unless it is backed up by mineral, energy or other resources that have tangible value. This won't be great for the collective West, but pretty good for the rest of the world.
I just think that effects of the AI bubble bursting would be at most a symptom or trigger of much larger geopolitical and financial shifts that would happen anyway.
The first is how much of the capital expenditures are being fueled by debt that won't be repaid, and how much that unpaid debt harms lending institutions. This is fundamentally how a few bad debts in 2008 broke the entire financial system: bad loans felled Lehman Brothers, which caused one money market fund to break the buck, which spurred a massive exodus from the money markets rather literally overnight.
The second issue is the psychological impact of 40% of market value just evaporating. A lot of people have indirect exposure to the stock market and these stocks in particular (via 401(k)s or pensions), and seeing that much of their wealth evaporate will definitely have some repercussions on consumer confidence.
Solving problems that humanity couldn't solve is super-AGI or something like that. It's not there indeed.
Which, actually is not a real thing. Nor has it ever really been meaningful.
Trolls on IRC "beat the turing test" with bots that barely even had any functionality.
They're good at the Turing test. But that only marks them as indistinguishable from humans in casual conversation. They are fantastic at that. And a few other things, to be clear. Quick comprehension of an entire codebase for fast queries is horribly useful. But they are a long way from human-level general intelligence.
Of course they can sound very human like, but you know you shouldn't be that naive these days.
Also you should of course not judge based on a few words.
Another case of culture flowing from the top I guess.
1) What good is your open problem set if really its a trivial "google search" away from being solved. Why are they not catching any blame here?
2) These answers still weren't perfectly laid out for the most part. GPT-5 was still doing some cognitive lifting to piece it together.
If a human would have done this by hand it would have made news and instead the narrative would have been inverted to ask serious questions about the validity of some these style problem sets and/or ask the question how many other solutions are out there that just need pieced together from pre-existing research.
But, you know, AI Bad.
They are a community run database, not the sole arbiter and source of this information. We learned the most basic research back in highschool, I'd hope researchers from top institutions now working for one of the biggest frontier labs can do the same prior to making a claim, but microblogging has and continues to be a blight on any accurate information so nothing new there.
> GPT-5 was still doing some cognitive lifting to piece it together.
Cognitive lifting? It's a model, not a person, but besides that fact, this was already published literature. Handy that a LLM can be a slightly better search, but calling claims of "solving maths problems" out as irresponsible and inaccurate is the only right choice in this case.
> If a human would have done this by hand it would have made news [...]
"Researcher does basic literature review" isn't news in this or any other scenario. If we did a press release every journal club, there wouldn't be enough time to print a single page advert.
> [...] how many other solutions are out there that just need pieced together from pre-existing research [...]
I am not certain you actually looked into the model output or why this was such an embarrassment.
> But, you know, AI Bad.
AI hype very bad. AI anthropomorphism even worse.
Please explain how this is in any way related to the matter at hand. What is the relation between the incompleteness of an math problem database, and AI hypesters lying about the capabilities of GPT5? I fail to see the relevance.
> If a human would have done this by hand it would have made news
If someone updated information on an obscure math problem aggregator database this would be news?? Again, I fail to see your point here.
The real problem here is that there's clearly a strong incentive for the big labs to deceive the public (and/or themselves) about the actual scientific and technical capabilities of LLMs. As Karpathy pointed out on the recent Dwarkesh podcast, LLMs are quite terrible at novel problems, but this has become sort of an "Emperor's new clothes" situation where nobody with a financial stake will actually admit that, even though it's common knowledge if you actually work with these things.
And this directly leads to the misallocation of billions of dollars and potentially trillions in economic damage as companies align their 5-year strategies towards capabilities that are (right now) still science fiction.
The truth is at stake.
If a purported expert in the field can is willing to credulously publish this kind of result, it's not unreasonable to assume that either they're acting in bad faith, or (at best) are high on their own supply regarding what these things can actually do.
Edit: we are in peak damage control phase of the hype cycle.
Once i told a coworker that a piece if his code looked rather funky (without doing a more deep CR), and he told me its "proven correct by AI". I was stunned, and asked him if he knows how LLMs generate their responses? He was genuinely in the belief that it was in fact "artificial intelligence" and was some sort of "all knowing entity".
So, it's not a matter of them not being able to do a good job of preventing the model from doing it, therefore giving up and instead encouraging it to do it (which anyways makes no sense), but rather them having chosen to train the model to do this. OpenAI is targetting porn as one of their profit centers.
I don't know about the former, but the latter absolutely has sexually explicit material that could make the model more likely to generate erotic stories, flirty chats, etc.
Tell that to the thousands of 18 year olds who'll be captured by this predatory service and get AI psychosis
I would argue that AI generated porn might be more ethical than traditional porn because the risk of the models being abused or trafficked is virtually zero.
That's not really true. Look at one if the more common uses for AI porn: taking a photo of someone and making them nude.
Deepfake porn exists and it does harm
I was just pointing out that when you're talking about the scale of harm caused by the existing sex industry compared to the scale of harm caused by AI generated pornographic imagery, one far outweighs the other.
What if you get a model that is 99% similar to your “target” - what we do with that?
Before only rich can afford to pay a pro to do photoshop. Now any poor person can get.
So why when rich can is fine and when everyone can is a problem?
Definitely not anti-AI here. I think I have been disappointed though, since then, to slowly learn that they're (still) little beyond that.
Still amazing though. And better than a Google search (IMHO).
That seems out of character for him - more like something I'd expect from Elon Musk. What's the context I'm missing?
Possibly entered the language as a saying due to Shakespeare being scurrilous.
I remember a public talk, where he was on the stage with some young researcher from MS. (I think it was one of the authors of the "sparks of brilliance in gpt4" paper, but not sure).
Anyway, throughout that talk he kept talking above the guy, and didn't seem to listen, even though he obviously didn't try the "raw", "unaligned" model that the folks at MS were talking about.
And he made 2 big claims:
1) LLMs can't do math. He went on to "argue" that LLMs trick you with poetry that sounds good, but is highly subjective, and when tested on hard verifiable problems like math, they fail.
2) LLMs can't plan.
Well, merely one year later, here we are. AIME is saturated (with tool use), gold at IMO, and current agentic uses clearly can plan (and follow up with the plan, re-write parts, finish tasks, etc etc).
So, yeah, I'd take everything any one singular person says with a huge grain of salt. No matter how brilliant said individual is.
Edit: oh, and I forgot another important argument that Yann made at that time:
3) because of the nature of LLMs, errors compound. So the longer you go in a session, the more errors accumulate so they devolve in nonsense.
Again, mere months later the o series of models came out, and basically proved this point moot. Turns out RL + long context mitigate this fairly well. And a year later, we have all SotA models being able to "solve" problems 100k+ tokens deep.
PS: So just we're clear: formal planning in AI </> making a coding plan in Cursor.
Sure, but isn't that moving the goalposts? Why shouldn't we use LLMs + tools if it works? If anything it shows that the early detractors weren't even considering this could work. Yann in particular was skeptical that long-context things can happen in LLMs at all. We now have "agents" that can work a problem for hours, with self context trimming, planning to md files, editing those plans and so on. All of this just works, today. We used to dream about it a year ago.
It can be considered as that, sure, but anytime I see Lecun talking about this, he does recognize that you can patch your way around LLMs, the point is that you are going to hit limits eventually anyways. Specific planning benchmarks like Blockworld and the like show that LLMs (with frameworks) hit limits when they're exposed to out-of-distribution problems, and that's a BIG problem.
> We now have "agents" that can work a problem for hours, with self context trimming, planning to md files, editing those plans and so on. All of this just works, today. We used to dream about it a year ago.
I use them everyday but I still woulnd't really let them work for hours in greenfield projects. And we're seeing big vibe coders like Karpathy say the same.
Personally i do not see it like that at all as one is referring to LLMs specifically while the other is referring to LLMs plus a bunch of other stuff around them.
It is like person A claiming that GIF files can be used to play Doom deathmatches, person B responding that, no, a GIF file cannot start a Doom deathmatch, it is fundamentally impossible to do so and person A retorting that since the GIF format has a provision for advancing a frame on user input, a GIF viewer can interpret that input as the user wanting to launch Doom in deathmatch mode - ergo, GIF files can be used to play Doom deathmatches.
The original point was about the capabilities LLMs themselves since the context was about the technology itself, not what you can do by making them part of a larger system that combines LLMs (perhaps more than one) with other tools.
Depending on the use case and context this distinction may or may not matter, e.g. if you are trying to sell the entire system, it probably is not any more important how the individual parts of the system work than what libraries you used to make the software.
However it can be important in other contexts, like evaluating the abilities of LLMs themselves.
For example i have written a script on my PC that my window manager calls to grab whatever text i have selected on whatever application i'm running and passes it to a program i've written in llama.cpp to load Mistral Small with a prompt that makes it check for spelling and grammar mistakes which in turn produces some script-readable input that another script displays in a window.
This, in a way, is an entire system. This system helps me find grammar and spelling mistakes in the text i have selected when i'm writing documents where i care about finding such mistakes. However it is not Mistral Small that has the functionality of finding grammar and spelling mistakes in my selected text, it only provides the part that does the text checking, the rest is done by other external non-LLM pieces. An LLM cannot intercept keystrokes in my computer, it cannot grab my selected text nor can create a window on my desktop, it doesn't even understand these concepts. In a way this can be thought as a limitation from the perspective of the end result i want, but i work around it with the other software i have attached to it.
So weird that you immediately move the goalposts after accusing somebody of moving the goalposts. Nobody on the planet told you not to use "LLMs + tools if they work." You've moved onto an entirely different discussion with a made-up person.
> All of this just works, today.
Also, it definitely doesn't "just work." It slops around, screws up, reinserts bugs, randomly removes features, ignores instructions, lies, and sometimes you get a lucky result or something close enough that you can fix up. Nothing that should be in production.
Not that they're not very cool and very helpful in a lot of ways. But I've found them more helpful in showing me how they would do something, and getting me so angry that they nerd-snipe me into doing it correctly. I have to admit, 1) however, that sometimes I'm not sure that I'd have gotten there if I hadn't seen it not getting there, and 2) sometimes "doing it correctly" involves dumping the context and telling it almost exactly how I want something implemented.
But isn't tool use kinda the crux here?
Correct me if I'm mistaken, but wasn't the argument back then on whether LLMs could solve maths problems without e.g. writing python to solve? Cause when "Sparks of AGI" came out in March, prompting gpt-3.5-turbo to code solutions to assist solving maths problems over just solving them directly was already established and seemed like the path forward. Heck, it is still the way to go, despite major advancements.
Given that, was he truly mistaken on his assertions regarding LLMs solving maths? Same for "planning".
They really can’t. Token prediction based on context does not reason. You can scramble to submit PRs to ChatGPT to keep up with the “how many Rs in blueberry” kind of problems but it’s clear they can’t even keep up with shitposters on reddit.
And your 2nd and third point about planning and compounding errors remain challenges.. probably unsolvable with LLM approaches.
Debating about "reasoning" or not is not fruitful, IMO. It's an endless debate that can go anywhere and nowhere in particular. I try to look at results:
https://arxiv.org/pdf/2508.15260
Abstract:
> Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages modelinternal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional model training or hyperparameter tuning and can be seamlessly integrated into existing serving frameworks. We evaluate DeepConf across a variety of reasoning tasks and the latest open-source models, including Qwen 3 and GPT-OSS series. Notably, on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.
Thats kind of the whole need isn’t it? Humans can automate simple tasks very effectively and cheaply already. If I ask my pro versions of LLM what the Unicode value of a seahorse is, and it shows a picture of a horse and gives me the Unicode value for a third completely related animal then it’s pretty clear it can’t reason itself out of a wet paper bag.
Ignoring conversations about 'reasoning', at a fundamental level LLMs do not 'do math' in the way that a calculator or a human does math. Sure we can train bigger and bigger models that give you the impression of this but there are proofs out there that with increased task complexity (in this case multi-digit multiplication) eventually the probability of incorrect predictions converges to 1 (https://arxiv.org/abs/2305.18654)
> And your 2nd and third point about planning and compounding errors remain challenges.. probably unsolvable with LLM approaches.
The same issue applies here, really with any complex multi-step problem.
> Again, mere months later the o series of models came out, and basically proved this point moot. Turns out RL + long context mitigate this fairly well. And a year later, we have all SotA models being able to "solve" problems 100k+ tokens deep.
If you go hands on in any decent size codebase with an agent session length and context size become noticeable issues. Again, mathematically error propagation eventually leads to a 100% chance of error. Yann isn't wrong here, we've just kicked the can a little further down the road. What happens at 200k+ tokens? 500k+ tokens? 1M tokens? The underlying issue of a stochastic system isn't addressed.
>While Yann is clearly brilliant, and has a deeper understanding of the roots of the filed than many of us mortals, I think he's been on a debbie downer trend lately
As he should be. Nothing he said was wrong at a fundamental level. The transformer architecture we have now cannot scale with task complexity. Which is fine, by nature it was not designed for such tasks. The problem is that people see these models work on a subset of small scope complex projects and make claims that go against the underlying architecture. If a model is 'solving' complex or planning tasks but then fails to do similar tasks at a higher complexity it's a sign that there is no underlying deterministic process. What is more likely: the model is genuinely 'planning' or 'solving' complex tasks, or that the model has been trained with enough planning and task related examples that it can make a high probability guess?
> So, yeah, I'd take everything any one singular person says with a huge grain of salt. No matter how brilliant said individual is.
If anything, a guy like Yann with a role such as his at a Mag7 company being realistic (bearish if you are a LLM evangelist) about what the transformer architecture can do is a relief. I'm more inclined to listen to him than a guy like Altman who touts LLMs as the future of humanity meanwhile is path to profitability is AI Tik-Tok, sex chatbots, and a third party way to purchase things from Walmart during a recession.
Nobody does that. You can't "submit PRs" to an LLM. Although if you pick up new pretraining data you do get people discussing all newly discovered problems, which is a bit of a neat circularity.
> And your 2nd and third point about planning and compounding errors remain challenges.. probably unsolvable with LLM approaches.
Unsolvable in the first place. "Planning" is GOFAI metaphor-based development where they decided humans must do "planning" on no evidence and therefore if they coded something and called it "planning" it would give them intelligence.
Humans don't do or need to do "planning". Much like they don't have or need to have "world models", the other GOFAI obsession.
https://x.com/SebastienBubeck/status/1977181716457701775:
> gpt5-pro is superhuman at literature search:
> it just solved Erdos Problem #339 (listed as open in the official database https://erdosproblems.com/forum/thread/339) by realizing that it had actually been solved 20 years ago
https://x.com/MarkSellke/status/1979226538059931886:
> Update: Mehtaab and I pushed further on this. Using thousands of GPT5 queries, we found solutions to 10 Erdős problems that were listed as open: 223, 339, 494, 515, 621, 822, 883 (part 2/2), 903, 1043, 1079.
It's clearly talking about finding existing solutions to "open" problems.
The main mistake is by Kevin Weil, OpenAI CTO, who misunderstood the tweet:
https://x.com/kevinweil/status/1979270343941591525:
> you are totally right—I actually misunderstood @MarkSellke's original post, embarrassingly enough. Still very cool, but not the right words. Will delete this since I can't edit it any longer I think.
Obviously embarassing, but completely overblown reaction. Just another way for people to dunk on OpenAI :)
He, more than anyone else, should be able to for one parse the original statements correctly and for another maybe realize that if one of their models had accomplished what he seemed to think GPT-5 had, that may require some more scrutiny and research before posting it. That would have, after all, been a clear and incredibly massive development for the space, something the CTO of OpenAI should recognize instantly.
The amount of people that told me this is clear and indisputable proof that AGI/ASI/whatever is either around the corner or already here is far more than zero and arguing against their misunderstanding was made all the more challenging because "the CTO of OpenAI knows more than you" is quite a solid appeal to authority.
I'd recommend maybe a waiting period of 48h before any authority in any field can send a tweet, that might resolve some of the inaccuracies and the incredibly annoying need to just jump on wild bandwagons...
My boss always used to say “our only policy is, don’t be the reason we need to create a new policy”. I suspect OpenAI is going to have some new public communication policies going forward.
The deleted tweet that the article is about said "GPT-5 just found solutions to 10 (!) previously unsolved Erdös problems, and made progress on 11 others. These have all been open for decades." If it had been posted stand-alone then I would certainly agree that it was misleading, but it was not.
It was a quote-tweet of this: https://x.com/MarkSellke/status/1979226538059931886?t=OigN6t..., where the author is saying he's "pushing further on this".
The "this" in question is what this second tweet is in turn quote-tweeting: https://x.com/SebastienBubeck/status/1977181716457701775?t=T... -- where the author says "gpt5-pro is superhuman at literature search: [...] it just solved Erdos Problem #339 (listed as open in the official database erdosproblems.com/forum/thread/3…) by realizing that it had actually been solved 20 years ago"
So, reading the thread in order, you get
* SebastienBubeck: "GPT-5 is really good at literature search, it 'solved' an apparently-open problem by finding an existing solution"
* MarkSellke: "Now it's done ten more"
* kevinweil: "Look at this cool stuff we've done!"
That seems like an easy mistake to entirely honestly make, and I think the pile-on is a little unfair.
Previously unsolved. The context doesn't make that true, does it?
Don't get me wrong, effectively surfacing unappreciated research is great and extremely valuable. So there's a real thing here but with the wrong headline attached to it.
If I said that I solved a problem, but actually I took a solution for an old book, people would call me a liar. If I was prominent person, it would be academic fraud incident. No one would be saying that "I did extremely valuable thing" or "there was a real thing here".
Henrietta Leavitt's work on the relation between a stars period of pulsation and brightness was tucked away in a Harvard Journal, which had revolutionary potential not appreciated until Hubbel recalled and applied her work years later to demonstrate galactic redshift in Andromeda, understanding that it was an entirely separate galaxy, that it was receding away from us and contributing to the bedrock of modern cosmology.
The pathogenic basis for ulcers was proposed in the 1940s, which later became instrumental to explaining data in the 1980s and led to a Nobel prize in 2005.
It is and has always been fundamental to the progress of human knowledge to not just propose new ideas but to pull pertinent ones from the literature and apply them in new contexts, and depending on the field, the research landscape can be inconceivably vast, so efficiencies in combing through it can create the scaffolding for major advancements in understanding.
So there's more going on here than "lying".
No, Weil said he himself misunderstood Sellke's post[1].
Note Weil's wording (10 previously unsolved Erdos problems) vs. Sellke's wording (10 Erdos problems that were listed as open).
Survivor bias.
I can assure you that GPT-5 fucks up even relatively easy searches. I need to have a very good idea how the results looks like and the ability to test it to be able to use any result from GPT-5.
If I throw the dice 1000 times and post about it each time that I got a double six. Am I the best dice thrower that there is?
It is pretty hard to fuck that up, since you aren't expected to find everything anyway. The idea of "testing" and "using any result from GPT" is just, like, reading the papers and seeing if they are tangentially related.
If I may speak to my own experience, literature search has been the most productive application I've personally used, more than coding, and I've found many interesting papers and research directions with it.
It sounds like the content of the solutions themselves are perfectly fine, so it's unfortunate that the headline will leave the impression that these are just more hallucinations. They're not hallucinations, they're not wrong, they're just wrongly assigned credit for existing work. Which, you know, where have we heard that one before? It's like the stylistic "borrowing" from artists, but in research form.
Google's result has more recently been generalised: https://arxiv.org/abs/2506.13242
Some people just read the "48 multiplications for a 4x4 matrix multiplications" part, and thought they found prior art at that performance or better. But they missed that the supposed prior art had tighter requirements on the contents of the matrix, which meant those algorithms were not usable for implementing a recursive divide and conquer algorithm for much larger matrix multiplications.
Here is a HN poster claiming to be one of the authors rebutting the claim of prior art: https://news.ycombinator.com/item?id=43997136
Imagine if you were talking about your own work online, you make an honest mistake, then the whole industry roasts you for it.
I’m so tired of hearing everyone take stabs at people at OpenAI just because they don’t personally like sama or something.
However when representing an reputable organization, people are expected to be cautious or otherwise required to have their comments reviewed and most organizations would enforce this to protect their brand or reputation.
As Carl Sagan said it best, extraordinary claims require extraordinary evidence. This was a pretty extraordinary claim and multiple senior staff at the org endorsed the comment without even a cursory check first. I would think serious observers are more concerned by the process and controls in OpenAI or lack thereof here, rather than a specific single mistake.
Like them or hate them, OpenAI is the leader in the industry, and everyone lookup up to them and their employees in a public forum for credible information so will hold them to higher standard than a lesser known lab.
The burden of checking for quality comes with having a reputable brand . The burden is being monetized / compensated with the valuation or size of fund raise a org is able to command.
No, it does not. It only produces a highly convincing counterfeit. I am honestly happy for people who are satisfied with its output: life is way easier for them than for me. Obviously, the machine discriminates me personally. When I spend hours in the library looking for some engineering-related math made in the 70s-80s, as a last resort measure, I can try to play this gambling with chat, hoping for any tiny clue to answer my question. And then for the following hours, I am trying to understand what is wrong with the chat output. Most often, I experience the "it simply can't be" feeling, and I know I am not the only one having it.
Of the other 50% that are real, it's often ~evenly split into sources I'm familiar with and sources I'm not.
So it's hugely useful in surfacing papers that I may very well never have found otherwise using e.g. Google Scholar. It's particularly useful in finding relevant work in parallel subfields -- e.g. if you work in physics but it turns out their are math results, or you work in political science and it turns out there are relevant findings from anthropology. And also just obscure stuff -- a random thesis that never got published or cited but the PDF is online and turns out to be relevant.
It doesn't matter if 75% of the results are not useful to me or hallucinated. Those only waste me minutes. The other 25% more than make up for it -- they're things I simply might never find otherwise.
You can't just give one single global hallucination rate since the rates depend on the different use cases and despite the abundant amount of information available to people on how to pick the appropriate tool for a given task, it seems very few people care to take the time to actually first recognize that these LLMs are tools, and that you do need to learn how to use these tools in order to be productive with them.
The hallucination rates are about the same as far as I can tell. It depends mostly on how niche the area is, not which model. They do seem to train on somewhat different sets of academic sources, so it's good to use them all.
I'm not talking about deep research or advanced thinking modes -- those are great for some tasks but don't really add anything when you're just looking for all the sources on a subject, as opposed to a research report.
but the reasons search got hard was that it became profitable to become the "winner" of a search query. It's a hostile market that works to actively undermine you.
AI absolutely will have the same problem if it "takes over" except the websites that win and get your views will not look like blogspam, they will look like (and be) the result of adversarial machine learning.
The fact that LLM's understand your question semantically, not just with keyword matching, is huge.
Should you blindly trust the summary? No. Should you verify key claims by clicking through to the source? Yes. Is it still incredibly useful as a search tool and productivity booster? Absolutely.
I did check the translations were correct as part of this — while my German isn't great, it was sufficient for this — and it was fine up until reaching a long table about the timeline of events relevant to the subject, at which point it couldn't help but make stuff up.
Still useful, but when you find the limits of their competence, there's no point attempting to cajole them to go further. They'll save you whatever % of the task in effort, now you have to do all the rest; it's a waste of effort to think either carrot or stick will get them to succeed if they can't do it in the first few tries.
> No, it does not.
> It is excellent when just finding something is enough.
How could you say that with high confidence when you admitted it might be useful for others?
1. Get list of sources and their summaries from an LLM.
2. Read through, find a paper who's title and summary seem interesting to you.
3. Follow the LLM's link, usually to an arXiv posting.
4. Read the title and abstract on arXiv. You can now judge the accuracy of the LLM's summary.
It's really easy to tell if the LLM is accurate when it is linking to something which has its own title and summary, which is almost always the case in literature search.
Example: https://platform.sturdystatistics.com/deepdive?search_type=e...
Edit: Got home and checked the error logs. There was a very long search query with no results. Bug on my end to not return an error in that case.
If you were hoping to use the citation network, it needs the url as input rather than the title.
Here is the list of what the system proposed to take a look at: 1. Vascular at‐risk genotypes and disease severity in Lebanese sickle cell disease patients 2. Narrow band filter for solar spectropolarimetry based on Volume Holographic Gratings 3. Communication from Space: Radio and Optical by S. Weinreb 4. An Adaptive and High Coding Rate Soft Error Correction Method in Network-on-Chips 5. Plasma hemostasis in patients with coronavirus infection caus...
I would expect more papers like 3rd as the topic is communication systems. Unfortunately, inside the ref.3 there is noting said about distortions.
Maybe it is indeed a linguistical curse of this topic...
But then they flip to the next page and they read a story on a subject they're not an expert on and they just accept all of it without question.
I think people might have a similar relationship with ChatGPT.
And I guess a lot of LLM-hype critics have the trait to be much less capable of "being able to flip to the next page and read a story on a subject they're not an expert on and they just accept all of it without question".
Because this is an unusual personality trait, these LLM-hype critics get reprimanded all the time by the "mob" that they don't see the great opportunities that LLMs could bring, even though the LLMs may not be perfect.
(1) people getting caught using it to do their own jobs for them (i.e. they don't even realise it's wrong about the stuff they do understand);
(2) people who see the problems and therefore don't trust them anywhere at all (i.e. no amnesia, quite sensible reaction);
(3) people who see the problems and therefore limit their use to domains where the answers can be verified (I do this).
--
As an aside, I'm a little worried that I keep spotting turns of phrase that I associate with LLMs, for example where you write "you're absolutely right": I have no idea if that's all just us monkeys copying what we see around us (something we absolutely do), or if you're using that phrase deliberately because of the associations.
The only thing I'm confident of is that you're not doing is karma-farming with an LLM, but that's based on your other comments not sounding at all like LLMs so why would you (oh how surprising it was when I was first accused of being an LLM), but eh, dead internet theory feels more and more real…
I don't know if something like this already exists and I'm just not aware of it to be fair.
Currently, I am applying RDF/OWL to describe some factual information and contradictions in the scientific literature. On an amateur level. Thus I do it mostly manually. The GPT-discourse somehow brings up not only the human-related perception problems, such as cognitive biases, but also truly philosophical questions of epistemology that should be resolved beforehand. LLM developers cannot solve this because it is not under their control. They can only choose what to learn from. For instance, when we consider a scientific text, it is not an absolute truth but rather a carefully verified and reviewed opinion that is based on the previous authorized opinions and subject to change in the future. So the same author may have various opinions over time. More recent opinions are not necessarily more "truthful" ones. Now imagine a corresponding RDF triple (subject-predicate-object tuple) that describes that. Pretty heavy thing, and no NLTK can decide for us what the truth is and what is not.
Basically we are trying to combine the benefits of chat with normal academic search results using semantic search and keyword search. That way you get the benefit of LLMs but you’re actually engaging with sources like a normal search.
Hope it was what you were looking for!
Can I ask, how did you build your search database?
I have a recent example where it helped me locate a highly relevant paper for my research. It was from an obscure journal and wouldn't show up in the first few pages of Google Scholar search. The paper was real and recently published.
However, using LLMs for doing lit review has been fraught with peril. LLMs often misinterpret the research findings or extrapolate them to make incorrect inferences.
It correctly described the locations in text, then it offered to provide a diagram.
I said “sure”, and it generated an image saying the chest location is on the neck, and a bunch of other clearly incorrect locations for the other measurement sites.
It’s gotten better. But it’s still bad.
As to not trusting the generated text, you’re totally right. That’s why I use it as a search tool but mostly ignore the content of what the LLM has to say and go to the source.
https://www.youtube.com/watch?v=RvGE-xhroy0
[drinks pee twice]
[1] https://x.com/SebastienBubeck/status/1970875019803910478
edit: full text
It's becoming increasingly clear that gpt5 can solve MINOR open math problems, those that would require a day/few days of a good PhD student. Ofc it's not a 100% guarantee, eg below gpt5 solves 3/5 optimization conjectures. Imo full impact of this has yet to be internalized...
This just seems to be the Open AI culture, which for better or worse has helped foster the AI hype environment we are currently in.
What mathematician uses this as the definition for “open”? I don’t go around saying that most problems in this textbook are open questions, just because I don’t know how to do them.
amelius•3mo ago
* OpenAI researchers claimed or suggested that GPT-5 had solved unsolved math problems, but in reality, the model only found known results that were unfamiliar to the operator of erdosproblems.com.
* Mathematician Thomas Bloom and Deepmind CEO Demis Hassabis criticized the announcement as misleading, leading the researchers to retract or amend their original claims.
* According to mathematician Terence Tao, AI models like GPT-5 are currently most helpful for speeding up basic research tasks such as literature review, rather than independently solving complex mathematical problems.
HarHarVeryFunny•3mo ago
So GPT-5 didn't derive anything itself - it was just an effective search engine for prior research, which is useful, but not any sort of breakthough whatsoever.