12/03/2026
How Claude Opus 4.6 Surprised Donald Knuth – And Why It Matters
People talk a lot about “smart” AI, but this time we’ve crossed a real line. Claude Opus 4.6, Anthropic’s latest model, helped Donald Knuth – a legend of computer science and author of *The Art of Computer Programming* – solve an open math problem he’d been stuck on for weeks.
In a short note called *Claude’s Cycles*, Knuth explains that he was working on a graph theory problem about Hamiltonian cycles in an \(m \times m \times m\) cube. He had the base case and some special cases, but no general solution. Over roughly 31 guided explorations, Claude Opus 4.6 ended up proposing an elegant construction that works for all odd \(m\). Knuth then checked everything and wrote a full rigorous proof himself, and later a formal proof in a proof assistant confirmed the solution.
What’s interesting isn’t just “the AI found the answer”, but *how* it worked: trying ideas, switching strategies (brute force, Gray‑code patterns, decomposing into “fibers”), and eventually discovering a simple structure based on the sum of coordinates modulo \(m\). It looks more like the work of a motivated PhD student than a basic chatbot. Knuth himself says he has to “re‑examine” his view of generative AI and calls this a “delightful” step forward in automatic reasoning and creative problem‑solving.
My take: this story is a strong signal about the future of AI for developers, researchers, and builders. It shows that a general‑purpose model like Claude Opus 4.6 is no longer just a tool for summarizing text or generating boilerplate code, but a real research partner that can suggest *new* ideas experts actually find useful. Humans still sit in the center for verification, proofs, and connecting results to the bigger theory, but the AI “co‑pilot” clearly accelerates exploration.
For us, that means:
- using these models to explore options when we’re stuck on an algorithm or system design,
- documenting each exploration step (as Filip Stappers did with Claude) to stay in control,
- and seeing AI not as a threat, but as a new R&D tool that’s becoming accessible to everyone