A startup claims it broke through a bottleneck that’s holding back LLMs

Miami-based AI startup Subquadratic announced last month that it has solved a mathematical bottleneck hindering large language models (LLMs) for nearly a decade. The company claims its new LLM, SubQ, is faster, cheaper, and more energy-efficient than existing models. SubQ reportedly can process up to 12 times more text simultaneously, enabling complex tasks like analyzing extensive document sets or entire codebases. Subquadratic asserts that SubQ matches the performance of leading models from Google DeepMind, OpenAI, and Anthropic on key tasks such as coding. Initial claims were met with skepticism due to a lack of substantial evidence beyond self-published test scores and limited public access to SubQ. AI engineer Dan McAteer characterized the situation as potentially "the biggest breakthrough since the Transformer... or it’s AI Theranos." To address this skepticism, Subquadratic has since published additional information, including results from independent tests conducted by third-party firm Appen. Subquadratic cofounder and CTO Alex Whedon acknowledged the initial skepticism, stating that releasing third-party benchmarks alongside the first announcement would have mitigated it.