Will GPT-5 be able to self-improve? — LessWrong