Miami-based AI startup Subquadratic has emerged from stealth mode with a claim so bold it would make a Transformer blush: they've supposedly cracked a mathematical bottleneck that has been holding large language models hostage for nearly a decade. The company's new model, SubQ, promises to be faster, cheaper, and more energy-efficient than anything else on the market, while processing up to 12 times as much text at once. That's like reading War and Peace in one gulp instead of flipping pages one by one.

But here's the catch: when Subquadratic made its grand announcement last month, the receipts were conspicuously absent. Skepticism was swift and merciless. Dan McAteer, an AI engineer, summed up the mood on X: "SubQ is either the biggest breakthrough since the Transformer ... or it's AI Theranos." Ouch.

A month later, the company has started bringing those receipts, publishing results from independent tests run by third-party firm Appen. The numbers look promising: SubQ scored 89.7% on LiveCodeBench, a coding test, and was 56 times faster than models using FlashAttention, a previous sparse-attention technique. "That was really exciting to me, it validated their architecture," says Jeanine Sinanan-Singh, Appen's director of generative AI research.

So what's the secret sauce? Subquadratic ditched dense attention - the core operation of transformers, which multiplies every token with every other token in a quadratic explosion of computation - in favor of sparse attention, which only multiplies selected pairs. The idea is that not every word relationship matters; you don't need to connect the first word of The Great Gatsby to the last word just to summarize it. "If you're reading a book, you're not going to look at the first and second words, first and third - that's insane," says cofounder Alex Whedon.

Subquadratic won't say exactly how SubQ chooses which words to focus on - because, you know, trade secrets - but claims it's dynamic and calculated on the fly. The cost savings are eye-popping: running Anthropic's LLM Opus through a test called RULER 128 costs $2,600, while SubQ allegedly does it for eight bucks. That's not a typo.

Before you toss your OpenAI subscription, some caveats: SubQ reused weights from the Chinese open-source model Qwen rather than training from scratch, which undercuts the claim of a full reinvention. And the model is still mostly behind a waitlist, with only a handful of users having actual hands-on experience. "The public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck," says independent researcher Will Depue.

Subquadratic, for its part, is philosophical about the skepticism. "We hope we're kicking off a new age of efficiency," says CEO Justin Dangel. "We don't think anybody will be building on transformers in a few years." Until then, the world waits - and wonders if this is the next big thing or just another well-funded math problem.