Representing LLM inputs as HTML instead of JSON to cut input tokens by 11%

Very interesting! I'd love to see a breakdown of various formats including YAML, TOML, CSV, and HCL with their relative cost vs accuracy.