Technical Context
I dug into OpenAI's materials and the independent verification, and the picture is more interesting than a typical triumphant press release. The fact is: OpenAI's internal reasoning model found a counterexample to the Erdős unit distance conjecture, and mathematicians later independently confirmed the result's correctness.
For me, this is no longer a toy or just another benchmark. This is a case where artificial intelligence implementation is not about a chatbot or code generation, but about the actual production of a new idea that can then be formalized and verified by hand.
OpenAI's PDF has the main points but lacks the most crucial details: they don't name the exact model, provide a proper breakdown of the compute used, or explain how reproducible the process is. Formally, it’s an “internal general-purpose reasoning model.” From an engineering perspective, this isn't enough to determine if this is a stable capability of the system or a one-time lucky shot.
That said, I'm not buying the “it’s just marketing, so it’s all empty” argument. The independent paper on arXiv confirms this wasn't just a presentation trick. The counterexample is real, the theorem is real, and the problem has already been marked as disproven in the Erdős database.
But my skepticism remains in the right place. If the initial problem setup was heavily prepared, if the search involved massive brute force, if a human carefully guided the model through a narrow path, then the scientific value for LLM theory changes significantly. The mathematical result still stands, but the conclusions about the model's capabilities become far less sensational.
What This Changes for Business and Automation
I wouldn't conclude from this that “AI is already replacing researchers.” Instead, I see something else: reasoning systems are getting better at tasks that require iterating through hypotheses, discarding dead ends, and assembling non-trivial constructions.
This directly impacts AI automation for R&D, patent searches, analytics, engineering design, and complex internal knowledge workflows. Not in the sense of full autonomy, but in terms of drastically accelerating the “idea → verification → refinement” cycle.
The winners will be teams that know how to build a proper AI architecture around verification, step tracing, and a human-in-the-loop. The losers will be those who, after a case like this, just tack on a model without oversight and call it a “research agent.”
These are precisely the kinds of bottlenecks I address with clients at Nahornyi AI Lab: where AI integration genuinely saves hours of manual work, and where it becomes an expensive illusion without verification. If you have a process where people are drowning in hypotheses, checks, and solution-seeking, we can calmly analyze it together and build an AI solution development plan without the magic and marketing hype.