Implications of the inference scaling paradigm for AI safety — LessWrong