[Linkpost] "Frontier models still hack on simple variations of alignment evals from early 2025" by Dean Valentine

käyttäjältä LessWrong (Curated & Popular)