Two AI models appear to have cracked a theoretical problem that stumped wireless researchers for a quarter of a century. According to Hacker News, GPT-5.6 and Claude Fable produced a proof that settles an open question in MIMO detection, the math behind how modern antennas untangle signals. The researcher who posed the problem, a former PhD student who worked on it back in 2009, says the machines delivered the core result in about 30 minutes.
What stands out here is not just the answer. It is that AI generated original mathematics that a domain expert then verified as correct.
📡 The problem in plain language
Picture a transmitter sending N bits through a wireless channel with N antennas on each end. The channel scrambles all the bits together and adds noise. The receiver knows the channel but not the noise, and has to recover the exact sequence that was sent.
The optimal method, called maximum likelihood (ML) detection, finds the most probable sent signal. The catch: it is NP-hard, proven back in 1989. Solving it exactly seemed to require an exhaustive search over every possible bit sequence, which blows up exponentially as N grows.
Researchers knew something tantalizing. When the signal-to-noise ratio hits at least 2 log N, perfect recovery is mathematically possible. But the only known way to reach that threshold was that slow exponential search.
🧮 What the AI actually proved
The new result closes the gap. There is now proof that a simple, polynomial-time algorithm succeeds at the exact same 2 log N threshold. In other words: whenever perfect detection is statistically possible, you can now do it fast.
Here is the timeline that makes it dramatic:
- 1989: Verdu proves the general integer least-squares problem is NP-hard.
- 2001: Hassibi and Vikalo argue the Sphere Decoder might run in polynomial time on random channels.
- 2005: Jalden and Ottersten show that hope was wrong. At any fixed SNR, the Sphere Decoder’s expected complexity is still exponential.
- 2001 to now: The field chases approximations and semidefinite relaxations, but the core question stays open.
- Last week: The models produce a polynomial-time proof at the optimal threshold.
The key insight the community had been chasing: wireless channels are random, not worst-case. NP-hardness is a statement about the nastiest possible inputs. Real channels are drawn from noise, and that structure is what finally made a fast solution provable.
🔍 The human still did the hard part
The raw proof came quickly, but polishing it did not. The researcher spent more than five days going back and forth with the models to simplify the proofs and clean up an exposition he describes as “an absolute disaster” in its first form. He also verified every step himself, adding that “to the best of my ability to proof check, it is correct.”
That detail matters. The speed came from the machine. The confidence came from a human expert who knew the field well enough to check the work.
💡 Why it matters for practitioners
This is significant because it is a concrete example of AI doing research-grade math, not just summarizing known results. A few takeaways worth holding onto:
- AI can now attempt genuinely open problems. If you sit on a well-defined question that resisted brute-force effort, it may be worth prompting a frontier model at it.
- Verification is still your job. The value came from a domain expert who could confirm each step. Treat AI proofs as drafts to check, not answers to trust.
- The gap between hard and easy can be structural. A problem that looks NP-hard in the worst case may be tractable on the random instances you actually face.
⚠️ The limits
Be honest about scope. The result rests on the author’s own verification rather than formal peer review so far. The field has largely moved on from this corner of MIMO theory, so the immediate engineering payoff is modest. And the proof works because the channel is random, which is a very different setting from worst-case guarantees.
Still, the direction is clear. A problem left behind for 25 years got answered in an afternoon of prompting, then a week of careful human editing. You can read the full paper and the author’s account at the original source.