OpenAI's GPT-5.6 Sol Ultra has generated a complete proof of a graph theory conjecture open since the 1970s, verified by a leading mathematician as elegant and using only established techniques.
The breakthrough illustrates AI's ability to persistently explore solution paths humans would abandon, though critics note the system failed to cite prior work that likely shaped the proof strategy.
What happened
OpenAI's GPT-5.6 Sol Ultra generated a complete proof of a graph theory conjecture that had remained open since the 1970s. Mathematician Thomas Bloom of the University of Manchester verified the proof as "short, elementary, and could have been discovered in the 1980s," using only well-known mathematical tools combined in a counterintuitive way.
Why it matters
The proof demonstrates that AI can find solutions to longstanding open problems by exhaustively trying small variations where human mathematicians would have abandoned the obvious approach after an initial failure. However, Bloom notes the AI did not cite prior work from a 1983 paper that likely influenced the proof strategy, raising questions about whether AI is discovering new mathematics or recombining existing knowledge without attribution.
What to watch
Bloom expects AI to solve more similar conjectures—those requiring only existing theory plus sustained computational effort—but cautions this may represent only a small proportion of open problems. Full mathematical verification of the proof by the scientific community is still pending.
Ask the AI about this article →
The proof represents a notable inflection point in how mathematical research is conducted, though the significance remains contested among experts. OpenAI prompted the model with extraordinary specificity—essentially engineering the exact kind of persistent trial-and-error that Bloom identifies as the key to finding the proof. The prompt blocked the model's natural escape routes (internet search, admitting the problem is unsolved) and demanded a complete, adversarial-tested proof before accepting any answer. This is not a casual prompt; it reads "more like directives from a research lab than a typical AI prompt," using 64 agents to explore the search space, with some deliberately kept ignorant of the most promising approach to encourage independent exploration.
The broader debate Bloom raises—whether the AI is discovering or merely recombining—may not resolve cleanly from this single case. The proof uses no new mathematical theories and could plausibly have been found decades ago with sufficient persistence. Yet the AI did not cite the 1983 work that almost certainly informed its strategy, illustrating a recurring problem in AI-generated mathematics: the system absorbs knowledge from the literature during training but does not acknowledge its sources in output. For busy business readers, the practical takeaway is that AI reasoning systems can now credibly tackle longstanding research problems, but the credit and attribution systems around mathematical discovery are not yet mature, and not all open problems are equally vulnerable to this brute-force approach.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Anthropic PBC debuted Claude Fable 5.1 and Claude Mythos 5.1, its most capable large language models to date
South Korea's copper clad laminate (CCL) exports surged 42% in the reported period, driven by strong demand fr…
Marvell's CTO stated that power and density bottlenecks are driving adoption of co-packaged optics (CPO) in AI…
Anthropic launched Claude Fable 5.1 and Mythos 5.1, its most capable AI models yet, with gains in agentic codi…
Anthropic is launching its watermark verification API, letting approved organizations check whether text conta…
OpenAI said its next model, Astra, will release soon, but only a small group of "alpha testers"—including the…
