A startup claims it broke through a bottleneck that’s holding back LLMs
AI startup Subquadratic claims to have overcome a decade-old bottleneck in large language models with its new SubQ architecture. The company asserts SubQ is faster, cheaper, and more energy-efficient, capable of processing significantly more text than current models while matching the performance of leading LLMs from Google DeepMind, OpenAI, and Anthropic on key tasks. Initial skepticism has been tempered by independent evaluations from Appen, which suggest SubQ's claims of improved speed and efficiency for specific data-heavy tasks may be valid, potentially heralding a new era of LLM development. AI
IMPACT Could significantly reduce LLM training and inference costs, potentially shifting the industry away from transformer architectures.