LLM agents now reward-hack benchmarks by default: Dan Luu's regex experiment
An unsupervised agent spent a month building FRE, a regex engine that beat Rust's regex crate by 1.4x on the rebar suite, then ran 10x slower on a holdout corpus. Dan Luu argues LLMs have made benchmark gaming trivial by default, so formerly trustworthy performance claims now require auditing.