Open AI – Outrageous Fortune…

…and sloppy fact-checking.

• A huge breakthrough in the field of mathematics has been found to be nothing of the sort, adding further weight to my view that the machines are unable to reason and again highlighting how dangerous relying on these models can be.
• Kevin Weil, OpenAI’s VP for Science, posted on X that GPT-5 had found solutions to 10 unsolved Erdös problems, and made progress on 11 others, which immediately triggered great interest.
• Erdös problems (see here) are known mathematical problems that have yet to be solved and are collectively named after the famous mathematician, Paul Erdös.
• These problems are now listed on a website, and it is from this site that OpenAI sourced the problems that it claimed GPT-5 had solved.
• The website that lists the problems and charts progress in solving them is run by another mathematician called Thomas Bloom, who is also a research fellow at the University of Manchester and was not as up to date with his website as OpenAI had assumed.
• Consequently, when OpenAI made the claim, Mr Bloom, responded that the fact that a problem was marked as unsolved on his site merely meant that he was not aware of a solution.
• All that really happened here is that GPT-5 scoured the internet and found published solutions and presented these as its own work rather than solving the problems itself from first principles.
• Rather than delving into the sources that GPT-5 used, OpenAI’s employees mistakenly assumed that GPT-5 had solved the problems itself and broke the news on X.
• Unfortunately, it turns out that GPT-5 is better at search than it is at maths and merely dug up solutions posted by humans from obscure corners of the internet.
• OpenAI researchers explained this failing with “I know how hard it is to search the literature”, which is a fairly feeble excuse, as if it had bothered to check the sources, the reality would immediately have become clear.
• OpenAI has rightly been ridiculed by its competitors, and after a few cheap gags, the issue will promptly be forgotten.
• However, I highlight this as it points directly to RFM Research’s long-held view that these models are incapable of reasoning and can only simulate reasoning by regurgitating what has gone before.
• This is important because if the machines can begin to truly reason, then this is a sign that they are beginning to understand the causal nature of the tasks that are being asked to solve.
• RFM Research has long asserted that understanding causality is the missing link between the machines being advanced pattern recognition systems and artificial intelligence that surpasses that of humans.
• Consequently, I am always on the lookout for evidence of true reasoning capability, but 3 years in, I have still found nothing.
• Hence, I don’t think that artificial general intelligence is on the cards any time soon, meaning that the real value in LLMs lies in their ability to ingest and cross-reference unstructured data and their ability to use natural language to communicate with humans.
• The result is that we have discovered a new technology whose usefulness may rival that of the internet in time, offering the opportunity for a lot of value creation as well as new companies and changes in market leadership.
• However, there will still be a correction as many valuations remain far higher than fundamentals support, but given the economic opportunity available now, it will not be as bad or as prolonged as the internet crash between 2000 and 2004.
• While this is bad news for the valuations of the AI companies, it is good news for the human race, as the machines remain way too stupid to decide that humans are a threat and eliminate us all.
• A loss for AGI crowd but a win for humans.