A communications problem open for 25 years just got cracked by two "AI heavy guns."

Microsoft Research principal researcher Dimitris Papailiopoulos announced: using OpenAI's GPT-5.6 and Anthropic's Claude Fable 5, querying each separately and cross-checking, he proved in about five days of intense prompting a key proposition — exact polynomial-time maximum-likelihood MIMO detection exists in the statistically feasible region. The 25-year gap between "statistically possible" and "computationally possible" just closed.

Papailiopoulos worked on this exact problem during his own PhD. Seventeen years later, two models finished the proof for him.

A Problem That Stalled for a Quarter Century

The problem is classic: in an N×N Gaussian MIMO channel, the transmitter sends a bit vector, the receiver gets a mixed, noise-corrupted signal, and knowing channel matrix H, must decode the original bits error-free. Maximum-likelihood detection is optimal but reduces to exact search over a discrete least-squares problem — the naive approach requires enumerating everything, with exponential complexity.

Researchers knew for decades that at SNR 2logN, information theory permits perfect recovery. But between "statistically feasible" and "computationally feasible" sat a gap nobody crossed for over twenty years. Sphere decoding's expected complexity fell short; semidefinite relaxation gave no explicit threshold; heuristic flips lacked rigorous proof; statistical physics methods had no formal proof; MCMC analysis missed mixing-time bounds. The "statistical ceiling" and the "algorithmic floor" never met.

The breakthrough hinged on division of labor: GPT-5.6's initial proposal leaned on AMP-style algorithms with a complex analytic framework; Claude Fable's path was more pragmatic — a symbolic LMMSE start to get a high-quality initial point, then greedy bit-flipping refinement. Two steps, intuitively structured, drawn from ideas actually used in communications engineering. The final version adopted Fable's scheme: the initial proof took about half an hour, but polishing it down to a granularity checkable line-by-line with only basic linear algebra and probability took days.

AI Research's Next Form: Not an Oracle, a Puzzle Assembler

The most valuable part of this case isn't "AI solved another problem." It's the demonstration of how AI actually participates in mathematics: not the model independently completing a proof, but humans handling questioning, decomposition, and verification while the model assembles large bodies of known methods into a complete logical chain in the right order.

The proof introduced no new mathematics, no new concepts — mostly systematic assembly of existing tools and long-chain rigorous construction. That's precisely what current models excel at: assembling "routine steps" at the right granularity, doing the work humans can do but rarely sustain. What lowers the cost of hard problems isn't a flash of insight from some model — it's the work pattern of "trial-and-error plus assembly" itself.

This contrasts sharply with Issue 029's "AI math proof debunked in 24 hours." Last time, a model confidently proved something, then mathematicians pointed out it proved the wrong conjecture. This time, the researcher made verification part of the process — the preprint is public, and the community is invited to check it line by line. AI's value shifts from "giving answers" to "reducing the friction on the path to correct answers."

One judgment worth noting: roll these models back to earlier pretraining distributions and they might not complete this argument so smoothly. Model preferences are shaped by "collective intuition" — methods appearing often in the corpus feel "natural." Rather than treating them as oracles of truth, see them as a distillation of human experiential intuition, further sharpened by reinforcement learning. That's both the capability and the boundary.

For the industry, the bigger meaning: the LLM race has escalated to "solving real-world problems." Whoever's model produces verifiable results in research, engineering, and production owns the next stage's pricing power. A 25-year open problem is a milestone for mathematics — and a side portrait of frontier-model commercial value.

See you tomorrow.