x
Minimising hallucinations through semantic token selection — LessWrong