Writing

2026

• May 17 Compressing the KV cache without retraining, and what mostly doesn't work
• May 15 Trying to push sparse attention further, and what didn't work
• May 13 90% of attention is slack in pretrained transformers
• May 11 Knowledge in weights, not in prompts
• Apr 23 A runtime for agent-authored apps
• Apr 22 Sandboxing agent-generated code with disposable unikernels
• Apr 8 Matching frontier LLMs with diverse small ensembles
• Apr 6 Code optimization with LLM-imagined ideas and evolutionary search