Blog
If AI supposedly only guesses the next word, how can it creatively solve a maths problem?
How can models that only learn from existing data come up with a genuinely new solution? And can they really do that?
In September 2026, OpenAI announced a successful proposed solution to a complex mathematical problem. Around 10,000 AI agents worked on it for 88 hours. Headlines declared that AI had beaten mathematicians. So where does the truth lie?
How is that possible? A language model, according to the usual (and heavily simplified) explanation, only guesses the next likely word. It cannot think independently and be creative... Or can it?
The answer, I think, is hiding in the word only.
AI does not "just" guess the next word blindly
A language model really does build its answer in small pieces called tokens. They might be words, parts of words or punctuation. This is how the model described in the GPT-4 technical report was trained (by the way, this is THE unassuming research paper that kicked off the entire AI craze for the general public three years ago).
But the model does not pick the next piece like a lottery number. It considers the question, the text already written and relationships learned from a huge number of examples (I will get into how that actually works in a future article).
Think of a chess player. On every turn, they also "only" choose the next move. To choose well, they must know the rules, see the board and anticipate what comes next. The description is true. It just leaves out the difficult part. Language models do something similar. That ability to track which words and ideas relate to each other sits at the heart of the Transformer architecture.
How a series of words becomes a mathematical proof
A mathematical proof is a very strict text. Every line must follow from the problem and previous steps. A model can suggest one step, then another, and gradually discover a new route.
This is where we need to slow the hype down. The scientists at OpenAI did not send one prompt to ChatGPT and go for coffee. They deployed around 10,000 autonomous AI agents that worked nonstop for over 88 hours. Another 17 hours went into formal verification using specialist software.
The announcement also sparked sharp controversy. Mathematicians Tristan Buckmaster and Levent Alpöge, who had been working on a related problem, accused OpenAI of rushing in the same direction after learning about their unpublished work. OpenAI denies the accusation and, after its own investigation, stated that their private prompts in the Codex tool could not have influenced the result, although they did not 100% rule out that data the scientists had entered into Codex (a developer tool by OpenAI) might have influenced the agents' "thinking".
Either way, the proposed proof still awaits independent scrutiny. Under the rules, it must be formally published, examined for at least two years and broadly accepted by the mathematics community before the million-dollar prize is even considered.
So what about creativity? Today's AI does not yet think the same way a human does. In practice, though, it offers something else: it can quickly and relatively reliably suggest possible next steps based on previous experience and instructions. Add reliable checking, and those suggestions can produce a genuinely new and creative result: new value.
What this means for your business
Use AI as a fast, senior colleague, not as the referee or decision maker. Let it draft an email, automate a tedious business process (I can help with that, by the way :)). Have it produce reports and clear charts. AI already excels at that. But do not leave product innovation or company strategy to AI. That path leads straight into the grey fog of generic answers and vague recommendations.
Do you run a business and want to make AI genuinely useful? Get in touch. We will design a solution. I will explain everything in plain language.