I used LLMs to build a compression tool that beats xz on x86_64 ELFs
LLM-guided structural preprocessing beats xz -9e on 103/103 Alpine binaries tested.
Search compressed data without decompressing—beats lzma on numeric arrays.
Data engineers, researchers, anyone managing large numeric datasets (time series, sensor data, images).
zstd · lzma · ClickHouse (columnar + compression)
LLM-guided structural preprocessing beats xz -9e on 103/103 Alpine binaries tested.
Single-file C++ ANS kernel beats wrestling with zstandard for quantized data.
AI-designed ELF preprocessing beats xz by 6% on 103 real binaries, perfectly reversible.
Runs 405B model compression on a single 32GB GPU when others need enterprise clusters.
SHA-256 verifiable manifests prove lossless compression mathematically, not just statistically.
99.9% compression claims need peer review—zero stars, one commit, no standard benchmarks.