What happened
New Scientist reporter Matthew Sparkes says ChatGPT generated a solution to a six-player dice problem after about 13 minutes of reasoning. Researcher Eric Harshbarger checked the numbers and found them correct. The construction uses five 720-sided dice and one 20-sided die, making it mathematically valid but impractical.
New Scientist reports that Sparkes asked a free version of ChatGPT to find six dice for six players such that every die had an equal chance of winning and no throw could result in a tie. The request followed mathematicians’ decade-long effort to construct a comparable five-die set. Sparkes says the chatbot produced an answer after roughly 13 minutes, following two clarifying prompts. Harshbarger, an Auburn University researcher involved in the five-die work, checked the numerical construction and told New Scientist that it was correct.
The reported construction is not the best known or a practical gaming product. Five dice have 720 sides and the sixth has 20 sides. New Scientist says Harshbarger and collaborators had already adapted their five-dice result into a six-dice construction in which every die has 360 sides. The researchers’ website also contains the five-dice solution and a method for extending solutions to more dice. New Scientist says a search found no trace of the exact six-dice answer, but the participants could not determine whether the chatbot located those public components and completed the work or produced the solution independently.
Source details: newscientist.com ↗
Why it matters
The report offers a concrete, independently checked example of a general-purpose AI system combining mathematical information into a valid solution. It does not establish that the model created a new mathematical concept, or that it can reliably solve harder problems without human oversight. The practical implication is that chatbots may help researchers generate candidate constructions, while people still need to verify them and judge whether they are useful.
This is a useful capability demonstration because the chatbot appears to have connected an existing mathematical construction with a known induction method and supplied a valid extension in plain English. Harshbarger’s numerical check is meaningful evidence that the answer was not simply incoherent output, but it is not the same as a formal proof or a controlled evaluation across many problems.
The report also shows the boundary of the result. The answer is awkward and physically unusable, and the source provides no independent test of the chatbot’s reasoning process. New Scientist quotes OpenAI’s Sébastien Bubeck predicting continued progress, but that is an expert view rather than evidence from this experiment. The article says current systems can find solutions or counterexamples to existing problems but have not demonstrated the creation of new mathematical concepts or fields.
What to watch next
The key unresolved questions are whether the result can be independently reproduced, whether a complete proof and transcript become available, and how much of the answer came from material already published online. The article describes the chatbot as free but does not identify the exact model, version, usage limits, or whether the capability is broadly available.
A stronger assessment would require the full prompt-and-response transcript, a formal verification of the construction, and tests by mathematicians who did not know the expected answer. It would also help to establish whether the model used the researchers’ website, how much prompting was needed, and whether the result can be reproduced with the same free access.
The broader question is whether AI can generalize from known mathematical ingredients to genuinely new problems, rather than solve variants whose relevant methods are already online. Until that is demonstrated, the report supports cautious use of chatbots as assistants for generating and checking candidate ideas, not as replacements for proof, judgment, or mathematical authorship.