Caltech Researchers Take Another Stab at One-Bit AI Models
The Wall St Journal has this story/release about researchers from Caltech that have launched PrismML, a startup that has developed a mathematical breakthrough enabling the "radical compression" of large language models to a 1-bit architecture without sacrificing reasoning or performance. By reducing the precision of model weights to a single bit, the company’s flagship "Bonsai" models can achieve up to an eightfold increase in processing speed and an 80% reduction in energy consumption, allowing high-fidelity AI to run locally on smartphones and laptops. Backed by $16.25 million in funding from investors like Khosla Ventures and Cerberus Capital, PrismML has open-sourced its technology to shift the AI paradigm toward "intelligence density," prioritizing energy efficiency and hardware adaptability over the traditional pursuit of ever-larger data centers. Previous 1-bit efforts focused on convolutional neural networks (CNNs), whereas this work seems to apply to transformers and other models relevant to large language models (LLMs).
