Google's TurboQuant algorithm compresses LLM key-value caches to 3 bits with no accuracy loss. Memory stocks fell within ...