Subquadratic Launches SubQ: A New AI Model That Claims to Break LLM Bottlenecks

The Miami-based AI startup Subquadratic emerged from stealth mode last month with a bold announcement. The company claims it has resolved a significant mathematical bottleneck that has hindered large language models (LLMs) for nearly a decade. Initially met with skepticism, Subquadratic has begun to provide evidence backing its claims, including results from an independent evaluation of its innovative technology.

Subquadratic unveiled its new model, named SubQ, which the company asserts is faster, cheaper, and requires significantly less energy compared to existing models. The startup claims SubQ can process up to 12 times more text simultaneously than most models, enabling it to tackle data-intensive tasks such as analyzing hundreds of documents or entire codebases.

SubQ's Performance Compared to Leading Models

According to Subquadratic, SubQ performs on par with leading models from Google DeepMind, OpenAI, and Anthropic in key areas like coding. Despite this, the company initially faced backlash due to a lack of substantial evidence beyond self-published test scores. As skepticism grew, industry experts began to question the validity of Subquadratic's claims.

After a month of scrutiny, Subquadratic released further details about SubQ, including results from independent tests conducted by the firm Appen. “We expected healthy skepticism,” said Alex Whedon, cofounder and CTO of Subquadratic. “In hindsight, releasing the third-party benchmarks alongside the initial announcement would have preempted much of the skepticism.”

Independent Validation of SubQ's Claims

Appen's evaluation provided a more comprehensive look at SubQ's capabilities, validating many of Subquadratic's assertions. Jeanine Sinanan-Singh, Appen’s director of generative AI research, expressed excitement over the results, stating, “That was really exciting to me; it validated their architecture.” She added that while SubQ may not completely replace existing top models, it could significantly enhance speed and reduce costs for specific tasks.

Subquadratic's innovation relies on a shift from traditional dense attention mechanisms to a new approach known as sparse attention. This method drastically reduces the number of computations required, allowing for more efficient processing. “If you’re reading a book, you’re not going to look at the first and second words, first and third—that’s insane,” Whedon explained.

The Future of LLMs: A Shift in Paradigm?

Subquadratic aims to revolutionize how LLMs are constructed, with cofounder and CEO Justin Dangel asserting, “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.” This statement underscores the potential impact of SubQ on the future landscape of AI and machine learning.

The traditional LLM architecture relies heavily on transformers, which utilize dense attention processes that result in high computational costs. As the length of text increases, the number of computations escalates exponentially due to the nature of dense attention. By contrast, SubQ's sparse attention approach selects only relevant token relationships, streamlining the process.

• SubQ can process up to 12 times more text than other models.
• It requires significantly less energy than existing models.
• Performance is comparable to top models from Google, OpenAI, and Anthropic.

As Subquadratic continues to refine SubQ and gather further independent validation, the AI community watches closely to see if this new model can indeed deliver on its promises and reshape the future of large language models.