A startup says it cracked the bottleneck holding back AI

Miami startup Subquadratic claims to have solved a nearly decade-old math problem that makes AI models slow and expensive. Its model, SubQ, uses "sparse attention" instead of the standard method, comparing only relevant word pairs rather than all of them. Independent tests showed SubQ running 56 times faster than leading methods, with one test costing $8 versus $2,600 on a competing model — though experts urge caution until broader access is available.