AI math problem solving: the workflow is the IP

A mathematician at Anthropic used Claude in AI math problem solving to help prove a 150-year-old riddle. We think the headline misses the point: AI didn't author the breakthrough. A quarter of a million pages of human-directed work did — and that workflow, not the model, is the real asset.
TL;DR: A mathematician at Anthropic used Claude to help crack a 150-year-old problem — but the headline hides the real lesson: the breakthrough lives in the human-directed workflow, not the model.

Key takeaway: AI didn't "solve" the problem on its own. A trained mathematician ran a quarter of a million pages of structured work through it, and that process — not the chatbot — is the intellectual property.

Why it matters: For marketing teams, the same truth holds. The model is a commodity; the repeatable, human-authored workflow around it is the thing competitors can't copy.

What happened

A researcher on Anthropic's technical staff says he used the company's Claude model to prove a problem that had resisted mathematicians for over 150 years. As ABC News reported in its wire story, he's claiming a genuine "Holy Grail" result with AI's help.

The details matter. Per AFP's reporting carried by Yahoo, the researcher, Levent Alpöge, consulted Anthropic chief executive Dario Amodei before announcing, and the result landed a day after the company's record-breaking IPO. He said more than 250,000 pages came out of his work with the AI.

Source: AFP via Yahoo, 2026

Coverage was careful to call it a claim. NPR noted it was the second time in recent months that an AI has been credited with a serious mathematical result — and, like any big proof, it needs to survive peer review before it's accepted.

Most people will read this as "the AI did the maths"

The consensus take writes itself: frontier models have crossed from calculation into discovery, Claude is now a research collaborator, and this is a proof point in the race against OpenAI and Google DeepMind. The timing — a day after the biggest IPO in history — makes that reading irresistible.

It's not wrong, exactly. But it quietly credits the machine with the authorship, and that's where the story gets misleading. Context helps: on Epoch AI's research-level FrontierMath benchmark, no model achieved even a 2% success rate on the full set. Raw AI maths is still bad. Something else did the heavy lifting here.

Source: Epoch AI (FrontierMath), 2025

Our take on AI math problem solving: the workflow is the IP

Here's what we think everyone's skipping. A 250,000-page body of work isn't a chatbot being clever. It's a trained mathematician setting up the problem, choosing the attack, catching the dead ends, and feeding structure back in — thousands of times. The model was the engine. The human-authored process was the invention.

That distinction isn't pedantic. It's legal and commercial. In most jurisdictions, raw AI output isn't copyrightable — there's no human author of record for a machine's text. What is protectable, and what actually holds value, is the workflow: the sequence of human decisions, prompts, checks and verification steps that turn a generic model into a specific result. The proof survived because a person owned that layer.

In our experience building agents, teams get this exactly backwards. They obsess over which model they're on, then wire it up with a thin, undocumented process that anyone could replicate in an afternoon. The model is a commodity you rent by the token. The workflow is the asset you own — and the only part a competitor genuinely cannot copy, because it encodes judgement, not just capability.

This is why we build around repeatable, inspectable processes rather than magic prompts. An agent that drafts your content is only as valuable as the documented workflow behind it — the brief structure, the brand rules, the human sign-off gates. That's the thinking behind our content creator agent: the model writes, but the workflow is what makes the output yours and keeps it defensible.

The Anthropic story is the same principle at the extreme. Strip out the human-designed 250,000-page process and you don't have a proof — you have a very expensive autocomplete. The headline says AI solved it. The footnotes say a person built the method, and the method is the thing worth protecting.

What this means for marketing teams

  • Document your workflows before your tools. Write down the human steps — brief, draft, check, approve — so the process is an asset you own, not tribal knowledge that walks out the door.
  • Treat model choice as a 12-month commodity decision, not a strategy. Assume you'll swap models at least once a year; build so a swap takes days, not a rebuild.
  • Add explicit human sign-off gates to any AI content pipeline. One verification step before publish catches the 5% of output that would otherwise cost you a correction.
  • Keep an audit trail of prompts and edits. If you ever need to defend ownership of AI-assisted work, the human-authored process is your evidence.
  • If you're starting from scratch, map one workflow end to end before buying anything — talk to us about where your process actually lives.

Frequently asked questions

Did an AI really solve a 150-year-old maths problem on its own?

No. A mathematician at Anthropic used Claude as a tool across more than 250,000 pages of work. The human directed the process; the result still requires peer review before it's accepted.

Is AI-generated content protected by copyright?

In most jurisdictions, raw AI output isn't copyrightable because there's no human author. The protectable, valuable layer is the human-authored workflow — the decisions, prompts and checks around the model.

Why does the workflow matter more than the AI model?

Models are commodities you rent and swap. The workflow encodes human judgement, is hard to copy, and is the part you can own and defend. It's where durable value sits.

Written by the Anjin team - we build AI marketing systems and remain professionally unimpressed by hype.

Continue reading