Two rival AI models were asked the same 25-year-old engineering question. Both came back with a proof — and they were different proofs

Dimitris Papailiopoulos, a principal researcher at Microsoft Research, put an open problem from wireless engineering to GPT-5.6 and to Anthropic's Fable 5 separately. The question dates from the early 2000s and concerns MIMO detection: a transmitter sends N bits across a channel with N antennas at each end, the channel scrambles them together and adds noise, and the receiver must recover every bit exactly. Doing this optimally is NP-hard. It has been known since the 2000s that exact recovery becomes statistically possible once the signal-to-noise ratio reaches roughly 2 log N — but the only method that reached that threshold was exhaustive search. Sphere decoding was proposed as a fast alternative in 2001 and shown in 2005 not to close the gap. Both models returned complete proofs that a fast algorithm does reach the same threshold, and they picked different algorithms to prove it with: GPT-5.6 used approximate message passing, while Fable 5 used signed LMMSE followed by greedy bit-flipping — a method engineers already use in practice but had never proved works. Papailiopoulos says the first proof took about thirty minutes and that he then spent five to seven days simplifying the argument and the write-up with the models, and has checked it end to end himself. There is no paper yet and no peer review, so the result should be read as a researcher's verified claim rather than a settled one.

✓ Verified · 3 sources

Related stories