Microsoft has a new AI research story to share, and the human involvement might be the most interesting part. The company says researchers used AI to help resolve a mathematics problem dating back to 1995, while separate attempts without human guidance failed to reach the same goal.
The problem is the Kannan-Lovász-Simonovits conjecture, better known as KLS. Microsoft Research’s Xinzhi Zhang and independent researcher Zhao Song have published a paper presenting a proof, addressing a question with implications for probability, optimization, and computer science.
If high-dimensional geometry sounds intimidating, the underlying question is easier to picture. Imagine dividing a solid shape into two equal volumes. You could make a flat cut or use a more complicated curved surface. In a simplified description, KLS asks whether the best flat cut always stays within a fixed factor of the best possible cut, even as the number of dimensions increases.
The computing connection comes from algorithms that explore these mathematical spaces through random walks. Bottlenecks affect how efficiently they can sample a distribution. The authors’ result establishes a bound that does not grow with dimension, strengthening the mathematical foundations behind sampling, volume calculations, and related optimization work. That does not mean every algorithm suddenly runs faster, but it resolves a question underlying how researchers analyze their efficiency.
According to Microsoft’s announcement, the team explored more than 100 candidate approaches with AI assistance. Humans evaluated intermediate results, discarded unpromising directions, and guided subsequent work. AI also assisted with calculations, manuscript preparation, and internal checks, while the researchers contributed the central mathematical reasoning.
Microsoft says unguided attempts with several leading AI models used roughly the same total token budget but fell short. That comparison is interesting, although one project cannot establish how AI will perform across scientific research. It also leaves open how different models, instructions, or research methods might have affected the outcome.
There is supporting work beyond the company’s announcement, too. Pierre Bizeul, Boaz Klartag, and Joseph Lehec published a separate proof on October 4. Their abstract explicitly credits a criterion developed by Song and Zhang as the crucial step. The separate paper therefore builds on the original researchers’ contribution, rather than arriving at the result with no shared foundation.
Today’s news is Microsoft’s explanation of the work, rather than the first publication of the completed proof earlier this month. The company also reports subsequent formal verification using Lean 4.
Microsoft’s emphasis on human involvement raises another question, though. Could it be overselling the human contribution to soften the news for the anti-AI crowd? Presenting AI as something that needs expert supervision may sound less threatening than suggesting it can perform intellectual work previously reserved for highly trained specialists.
There is no evidence that Microsoft tailored this announcement for that purpose. The company identifies specific mathematical contributions from the researchers, and its account of unsuccessful unguided attempts supports the argument that their direction mattered. A failed unguided run, however, does not tell us exactly how much of the successful discovery came from the AI.
That is where I would like more transparency. Examples showing which ideas the systems proposed, what researchers rejected, and how the successful argument developed would help readers assess the division of work. A description of the process can explain who supervised it without fully revealing who, or what, supplied its most useful insights.
For anyone tired of AI announcements built around another benchmark score, this is a more interesting application. But we should scrutinize claims about human indispensability just as carefully as claims about AI independence. The mathematics deserves attention, and so does an honest accounting of how it got done.
Support independent tech journalism
NERDS.xyz is independently owned and operated. If you enjoy my coverage of Linux, AI, hardware, cybersecurity, and tech culture, consider supporting the site on Ko-fi.
Support NERDS.xyz
