The case of mathematics shows that the hype threatens to explode in their faces

A group of OpenAI researchers claimed to have “found solutions to 10 previously unsolved Erdös problems, and progress has been made on 11 others.” The statement seemed to indicate that GPT-5 had made an important qualitative leap in the field of mathematics, but the reality was very different. In fact, it all turned out to be an exaggeration that may harm OpenAI’s reputation going forward.

what has happened. The OpenAI engineers’ claim was promising, but exaggerated. The original message from Mark Selke, one of them, was added to those of other researchers such as Boris Power—who he apologized after realizing that they had screwed up—or Sebastian Bubeck—who also ended up modifying the tweet and acknowledged the error—. The original tweet seemed to make it clear that GPT-5 had managed to solve several of the famous Erdös mathematical problems. I hadn’t really solved them.

GPT-5 served to find solutions. The mathematician Thomas Bloom, who is precisely in charge of managing the website where all these open problems are managed, quickly clarified the situation. As explained on X/TwitterOpenAI’s claims were “a dramatically misinterpretation.” When he talks about “open” problems on the website, what he means is that he doesn’t know the solution, not that the problem has not been resolved. The only thing GPT-5 did was find recent research and studies that Bloom had not found. Here we must say that AI has managed to make striking mathematical advances recently: Meta AI, for example, managed to generalize the Lyapunov function.

Demis Hassabis and Yann LeCun criticize OpenAI. Demis Hassabis, CEO of DeepMind, indicated in X that this event had been “shameful”, while Yann LeCun, one of the top AI managers at Meta, highlighted how OpenAI had believed its own hype sales message with the message “Hoisted by their own GPTards”, which plays on GPT and “tards” (a suffix derived from “retards”), in reference to the gullible expectations that OpenAI usually sells.

Expectations are everything. Although OpenAI researchers and engineers admitted their mistake, what we see here is a dangerous pattern: one in which even the company’s own employees—or the enthusiasts who follow it—can end up falling victim to those expectations. It is very likely that internally the pressure to achieve great advances with their models is enormous, but that can lead to oversights and exaggerations like this that can cost the company’s reputation dearly.

GPT-5 didn’t do badly at all. Although the role of GPT-5 in this process was exaggerated, what must be recognized is that this model demonstrated its ability to become a very valuable assistant for researchers. Thus, this AI model can search the Internet and scientific study libraries in a very powerful way, and can “find solutions” already published where academics had not yet seen them when trying to solve related problems.

Research assistant. For mathematician Terence Tao, this is precisely a very striking element of these AI models: they may not solve the most complex mathematical problems, but can speed up tedious tasks such as those of the search for academic literature that helps solve them. For this expert, AI can help “industrialize” mathematics and act as a catalyst or “lubricant” for mathematicians’ hypotheses and theories.

But this is important. OpenAI is a machine for creating expectations, and its CEO, Sam Altman, does not hesitate to make vague and impossible to verify promises to attract more interest in his generative artificial intelligence models. A year ago promised that the AGI would arrive “in a few thousand days”something that sounds like one of those “Musk’s promises”.

risky bet. In recent weeks we have seen how OpenAI has reached unique circular financing agreements with NVIDIA, amd either Broadcom to create data centers, but the reality is that all these projects focus on one promise: that AI will be a fundamental part of our lives sooner rather than later. That can happen, of course, but if it doesn’t, the domino effect can be an absolute catastrophe given the tens of billions of dollars invested in such projects.

Image | Vitaly Gariev

In Xataka | If the question is whether there is an AI bubble, Sam Altman has just given the answer. One with which he wins