Google Research has unveiled TurboQuant, a compression algorithm that reduces large language model memory consumption by at least 6x with zero accuracy loss, potentially democratizing AI deployment on consumer hardware.
A security researcher at Cambridge discovered nearly 3,000 internal Anthropic assets publicly accessible, including documentation of an unreleased model called Mythos that the company calls its most capable AI system ever trained.
Google TurboQuant algorithm achieves 6x KV cache compression with zero accuracy loss 鈥?and runs up to 8x faster on H100 GPUs.
Google researchers have developed TurboQuant, an innovative compression algorithm that can reduce AI memory usage by at least six times while maintaining zero accuracy loss, potentially transforming the economics of AI deployment.