I Almost Mistook a Math Headline for a Business Signal

Last week I saw a headline claiming “AI helps tackle the Riemann hypothesis.” My first thought was: could AI now prove my pricing proposals and contract terms more professionally? I was excited for about ten minutes, then returned to my pricing sheet. Clients never asked for a mathematical proof; they wanted clear, reliable answers they could act on.

What It Actually Is—and Who's Using It

The story centers on researchers trying to involve large language models in advanced mathematical reasoning. It is not an automatic answer machine for every problem. It is more like an assistant that can work alongside us, although its output still needs verification. Scott Aaronson discussed the related work on his blog, and Hacker News soon filled with questions and skepticism: the sample was small, the result might be incomplete, and the model might not actually understand what it was doing. Peers with the relevant expertise still need to check.

I got this wrong too. When I saw the words “proved,” I nearly reposted it as “AI can already replace experts.” Later I asked Xiaozhou, a friend who works in data, to take a look. He replied with one line: “First ask which step it proved, and where the complete proof is.” That brought me back down to earth.

You Do Not Have to Tackle the Hardest Math to Replicate the Workflow

To try a similar workflow, I would budget RMB 0–100 per month to start, with free tiers available on personal web tools; whether I pay depends on how often I use them. In terms of time, I would first set aside two hours to learn how to prompt, check sources, and recalculate results. There is no requirement to code, but there is a requirement to check every step. A fluent answer is not evidence.

The first step is simple: open the large-language-model chat tool I already use, start a blank conversation, and enter, “I have a business question. Break it into known conditions, hypotheses to verify, and steps for verification.” Then start with one small problem I am working on. For example, I could break “Should I give an existing client a discount?” into past transactions, marginal cost, probability of closing, and customer value. Not everyone needs this tool. If my daily work is just replying to a few emails, there is no urgency to try it now.

Match the Tool to My Stage—and Keep My Judgment

When I was just starting out, I would use it to organize my thinking, but I would not let it contact clients or promise deliverables. Otherwise, uncertain assumptions could turn into formal quotes. With one or two clients, I might have it play the roles of customer, collaborator, and skeptic, then verify the key numbers myself. For a team already scaling, I would first establish a traceable review process: keep a record of who asked the question, who verified it, and who made the final call.

What this math story really reminded me was not that AI had crossed another mountain. It was that polished conclusions make it easier to get carried away. We can still be excited—but then return to the concrete problem in front of us: which step did this actually help us avoid?