Rupture in matrix arithmetic: half-century power ladder compressed into single stride

A fifty-year argument about a hidden constant in matrix multiplication may have just been overturned by a single number: over the complex numbers, the exponent ceiling sits at most at 9 over 4, which is 2.25. It looks modest on paper. Against the scale of this field it is enormous. The previous best mark hovered near 2.371177, reached after more than a decade of combat over the fifth decimal place. Now a drop of 0.12 is on the table. In this piece I unpack what the argument in the video actually claims, why it matters, and where caution belongs.

Start with the ground floor. The plain method fills each cell of the product by pairing one row of the left factor with one column of the right factor, multiplying the aligned entries and adding them up. The narrator walks through a tiny case and then scales it: square factors of side n hold n squared cells, each costing about n operations, so the total cubic burden grows like n cubed. At side length two thousand that means roughly one billion operations, and at ten thousand it means about one trillion. The question is whether that cubic bill is a law of nature or merely a failure of imagination.

How the cubic wall first cracked

The first crack dates to Volker Strassen in 1969. He combined two-by-two blocks with sevenfold recursion instead of eight products, and by applying that single saved product recursively he drove the power down to log base two of seven, near 2.807. That was the first demonstration that the cubic bar was not fate, and it launched a race lasting more than half a century. I checked the numeric spine of that staircase against the Wikipedia record.

Next comes the quantity everyone hunts. Omega is the greatest lower value among all powers tau such that square multiplication fits, for every positive vanishing slack , into on the order of n to the tau plus slack operations. The slack lets us approach the boundary arbitrarily closely without demanding exact attainment. The crucial point: omega is an idealized long-run measure, silent about constant multipliers, cache behavior, rounding robustness, and performance at practical sizes. Read without that split, 2.25 misleads. I anchored that definition in the MIT survey of recent progress.

Then descends the record staircase . Tensor-rank work by Pan and Bini lowered the steps through the seventies and eighties, Coppersmith and Winograd reached about 2.376 with the laser method in 1987, and that figure haunted a generation. A near standstill of two decades followed. Stothers, Vassilevska Williams, Le Gall, Alman, and the Duan Wu Zhou line chipped the mark to 2.371339 by early 2024, and an AlphaEvolve-assisted refinement touched 2.371177 in August 2026 through loss-aware blending . I took the technical detail of that optimization story from the arXiv AlphaEvolve note. I matched the three stages of that pipeline, reformulation plus learned optimizer plus AlphaEvolve polish, to the AIWeekly rundown.

A different road toward 2.25

The October 2026 claim from OpenAI changes the grammar of the ladder: over the complex domain , omega is at most 9 over 4, so for every positive slack two complex square factors multiply in on the order of n to the 2.25 plus slack scalar steps. The text surfaced as an October 2 manuscript inside a broad mathematical release on October 6, arriving with a machine-checked proof of the bound in Lean. The release notes say most results cost roughly three Pro-thinking hours each. I pinned the scope of that release to the OpenAI announcement with its GitHub and Lean notes.

What persuades me is less the digit than the architecture behind it. The narrator builds the route through a spectral separation engine for tensors, a bridge to polynomial multiplication, a Fourier-flavored separation, and discrete growth inequalities, a genuinely different construction rather than a tighter squeeze of the old record. Companion manuscripts widen the picture: a dual exponent above 0.465 over every characteristic-zero field, a rectangular bound below 2.092, and a sub-2.258 square bound outside one uncomputed finite set of characteristics. I matched the scope table of those bounds to the BinaryVerseAI breakdown. I verified the companion list and its field-by-field splits against the VibeMathed theorem ledger.

A common misreading needs correction: nothing here promises faster training tomorrow morning. An upper bound is an existence statement; it names no competitive size, quotes no constants, and budgets nothing for memory traffic or parallel overhead. The narrator flags this honestly: the power describes how work scales as size runs to infinity. Hardware parallelism is a separate axis, because work that can run side by side differs from total arithmetic work. So read 2.25 as a redrawn ceiling, not as a shipped kernel.

Verification is under way, eyes on two

The echo, though, marks one of the busiest weeks mathematics has seen in years. The company published hundreds of papers at once; a researcher who spent 25 years on one solved problem voiced astonishment in the press, alongside warnings that only a handful of thirty number-theory papers looked striking and only one carried a Lean certificate. I carried the echo of that 722-paper wave through the NewScientist interviews with mathematicians. I gave the single-prompt-agent caveat and the demand for reproducible receipts through the objections in ScientificAmerican.

The last question is the deepest: might omega equal two? Two stands as a natural quadratic floor , since a side-n factor already holds n squared entries and the output holds n squared entries, so anything below two would process less work than the data itself in the usual model. Even equality would not mean linear time in n; it would mean roughly quadratic arithmetic in side length, near-linear in entry count up to lower-order drift. The quarter-point gap between 2.25 and two remains a vast unknown, and nobody knows whether it closes. I framed the open status of that descent to two with the EmergentMind problem file.