Startup Claims Breakthrough in AI Processing Bottleneck
▼ Summary
– Subquadratic claims its model, SubQ, solves the quadratic scaling problem in AI by using dynamic sparse attention instead of dense attention.
– Independent tests by Appen showed SubQ ran 56 times faster than FlashAttention and scored 89.7% on a coding benchmark, with costs as low as $8 versus $2,600 for a competitor.
– The startup started from an existing open-weight model and swapped in its attention method, which tempers claims of a full reinvention of how LLMs work.
– Skepticism remains because benchmarks differ from real-world use, and SubQ is not widely available, with only a small group having access.
– If validated, SubQ’s efficiency could enable cheaper, faster processing of large texts like codebases or contracts, addressing AI’s energy and cost challenges.
A Miami-based startup is making a bold claim: it has solved a core mathematical problem that has plagued artificial intelligence models for nearly a decade, making them slow and energy-intensive. The assertion was so audacious that it invited comparisons to Theranos. Now, however, the company has independent test results that largely validate its claims.
The company is called Subquadratic. It emerged from stealth mode in May with $29 million in seed funding and a new language model named SubQ. According to the startup, SubQ is faster, cheaper, and significantly more energy-efficient than today’s leading models. It can also process up to 12 times as much text simultaneously.
The Decade-Old Bottleneck
Understanding why this matters requires a look at how most large language models function. At their core is a “transformer,” introduced by Google researchers in 2017. The transformer relies on a process called dense attention.
Dense attention is thorough but computationally expensive. It compares every word in a text with every other word. As a result, when you double the length of the text, the workload roughly quadruples. This “quadratic” scaling is the primary reason large language models consume so much computing power and energy.
Subquadratic’s Solution
Subquadratic’s approach is to replace dense attention with “sparse attention.” Instead of comparing every word with every other, sparse attention only keeps the pairs that are important. This concept is not new; many teams have tried it. However, until now, none had matched the quality of dense attention.
The company claims its version finally achieves this. Crucially, it dynamically selects which words to focus on, based on the content rather than a fixed pattern. “That’s kind of where the secret sauce is,” explains co-founder and chief technology officer Alex Whedon.
The Receipts
Initially, the claims rested on a handful of self-published scores. Unsurprisingly, the reaction was skeptical. One AI engineer summed it up on X: SubQ is “either the biggest breakthrough since the Transformer … or it’s AI Theranos.”
So the company brought in a third party. It tasked Appen, a firm that evaluates other companies’ models, to run the tests. The results were striking. On a raw speed test, SubQ ran 56 times faster than FlashAttention, a leading existing method. On a tough coding benchmark, it scored 89.7 percent, close to the best models available.
The cost gap appears equally wide. According to the startup, running one long-context test on Anthropic’s top model costs about $2,600. On SubQ, the same test cost only eight dollars.
Still Too Good to Be True?
Even with these results, caution is warranted. Benchmarks are not the same as real-world usage. SubQ is also not widely available yet. Tens of thousands have joined the waitlist, but only a handful have access.
There is also a wrinkle in the origin story. Rather than training SubQ from scratch, Subquadratic started from an existing open-weight model and replaced its attention method. This is common practice, but it sits awkwardly alongside the claim of fully reinventing how large language models work.
“They may have built something real and useful,” says Will Depue, an independent researcher who previously worked at OpenAI. “But the public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck.”
Why It Matters
If the results hold up, the payoff is enormous. Cheaper, faster long-context models could read entire codebases, contract sets, or document troves in a single pass. They would also reduce the cost and energy required to run AI.
That prize is one the entire industry is chasing. AI already strains against the rising economics of AI agents, and other startups, such as Thomas Reardon’s Flourish, are attacking efficiency from different angles. Subquadratic, however, is betting the whole field will follow its lead. “We don’t think anybody will be building on transformers in a few years,” says chief executive Justin Dangel.
(Source: The Next Web)
