AI startup Subquadratic has claimed to have solved a mathematical bottleneck that has been hindering large language models (LLMs), unveiling its own model named 'SubQ'. Independent evaluations suggest the model is significantly faster, cheaper, and more energy-efficient than existing models, capable of processing up to 12 times more text at once.
AI startup Subquadratic has claimed to have solved the mathematical bottleneck of 'dense attention' that has been hindering large language models (LLMs), unveiling its own model named 'SubQ'. The model is reported to process up to 12 times more text at once than existing models, offering significant advantages in speed and cost-efficiency.
Most current LLMs rely on the transformer architecture and the dense attention mechanism, which causes a bottleneck where computational complexity increases sharply as the length of text to be processed grows. Subquadratic claims to have fundamentally solved this problem to greatly improve LLM efficiency. While initial announcements were met with skepticism due to a lack of evidence, the company has recently bolstered its claims by releasing results from an independent evaluation by the third-party firm Appen.
If SubQ's performance is verified, it could dramatically reduce the cost of training and inference for LLMs and enable large-scale data processing tasks—such as analyzing hundreds of documents or entire codebases—that were previously difficult. This has the potential to trigger a paradigm shift towards a 'new age of efficiency' in AI model design and could alter the future development trajectory of transformer-based architectures.