Miami Startup Makes AI Models 12x Faster, Uses Less Energy

Running the most powerful AI models today costs a fortune in electricity and computing power, but a Miami startup says it's cracked the code to change that.

Subquadratic emerged from stealth mode last month with a bold announcement: they'd built an AI model called SubQ that runs 12 times faster than current top models while using a fraction of the energy. The company claims their model matches the performance of systems built by Google, OpenAI, and Anthropic on important tasks like writing code and analyzing massive amounts of text.

At first, skeptics wondered if this was too good to be true. AI engineer Dan McAteer summed up the doubt on social media: "SubQ is either the biggest breakthrough since the Transformer ... or it's AI Theranos."

Now independent testing firm Appen has verified many of Subquadratic's claims. "That was really exciting to me, it validated their architecture," says Jeanine Sinanan-Singh, Appen's director of generative AI research. "I was like, 'Wow, this could be a game changer.'"

The secret lies in how the model processes information. Current AI models use something called "dense attention," which multiplies every word in a document against every other word. If you're summarizing a 10,000-word text, that triggers nearly 50 million separate calculations.

Subquadratic ditched that approach for "sparse attention," which only calculates the relationships between words that actually matter. Think about reading a book: you don't consciously compare the first word to every single other word on every page. Your brain focuses on the connections that carry meaning.

"Sparse attention says not all of those relationships are important, because they're not," says Alex Whedon, Subquadratic's cofounder and chief technology officer. Previous attempts at this approach failed to match the quality of traditional models, but SubQ appears to have solved that puzzle.

The Ripple Effect: The breakthrough could dramatically reduce AI's growing energy footprint. Current language models consume enormous amounts of electricity, raising concerns about environmental impact as AI adoption accelerates. A faster, more efficient approach could make powerful AI accessible to smaller companies and researchers who can't afford massive computing bills.

SubQ can process entire codebases or hundreds of documents at once, tasks that would be prohibitively expensive with today's models. Cofounder and CEO Justin Dangel believes this marks a turning point: "We hope we're kicking off a new age of efficiency. We don't think anybody will be building on transformers in a few years."

The company still needs to make SubQ widely available for others to test, but the independent verification suggests the AI world might be watching the start of something significant.