I built a programming language that LLMs can't use
2026-10-31
![]()
The author designed KillSwitch, an esoteric programming language built to defeat LLMs: specifications that contain lies, identifiers that require counting letters, constructs that trip safety guardrails, and code hidden in narrative noise. The best model on the benchmark, Claude Opus, scores 66.8%. The author reads that as evidence that LLM coding rests more on recognising patterns seen in training than on understanding code.
Was this useful?