Stacity: a Lock-In Risk Benchmark for Large Language Models — LessWrong