MIT researchers developed Attention Matching, a KV cache compaction technique that compresses LLM memory by 50x in seconds — ...
Google Squoosh replaced TinyPNG in my workflow with faster offline compression, flexible quality controls, real-time comparison, and support for more image formats.