Subquadratic Startup Claims Breakthrough in LLM Performance
Photo: MIT Technology Review
Quick answer
Subquadratic's SubQ model delivers a 12x increase in processed text volume, reduced energy consumption, and maintains parity with top-tier LLM performance.
U.S.-based startup Subquadratic has unveiled its new language model, SubQ, which it claims could revolutionize the large language model (LLM) industry. The company asserts that SubQ processes up to 12 times more text per batch than comparable models from Google DeepMind, OpenAI, and Anthropic, while maintaining comparable performance on key tasks such as code generation.
However, Subquadratic's initial claims faced significant skepticism. Critics highlighted the absence of independent verification and drew comparisons to Theranos' infamous downfall in medical technology. In response, the startup released additional test results conducted by independent firm Appen, which partially validated SubQ's efficiency claims.
According to Janine Sinanan-Singh, Appen's Director of Generative AI Research, the test results were surprising: «This could be a game-changer, as existing models struggle with speed and inefficiency.» Nevertheless, SubQ remains unavailable for widespread testing, leaving questions about its real-world capabilities unanswered.
Subquadratic's co-founder and CTO, Alex Widon, acknowledged that the company could have avoided some skepticism by providing independent data upfront. «Going forward, we will rigorously validate results before publishing,» he stated.
Common questions
- What is SubQ and how does it differ from other LLMs?
- SubQ is Subquadratic's new language model, reportedly processing 12x more text per batch while being faster and more energy-efficient than competitors. Its output quality remains comparable to models from Google DeepMind and OpenAI.
- Why were Subquadratic's claims met with skepticism?
- The startup initially lacked independent validation of its performance claims, relying solely on internal tests. This drew parallels to high-profile failures like Theranos until partial confirmation emerged from third-party evaluations.
- What tasks can SubQ handle?
- SubQ is designed for large-scale data processing, including analyzing hundreds of documents or entire codebases. It excels in high-speed, high-efficiency scenarios requiring rapid inference.
Dzen feed: /feed/dzen.xml · RSS: /feed.xml