Stanford team explores the RSIHF framework to evaluate the popularity and safety of AI automatic training

Stanford team proposes the RSIHF framework, which for the first time confirms that the 'popularity' and 'safety' of models during AI automatic training cannot be equated; stack of computing power cannot guarantee the win rate of model self-improvement, catering to human tendencies may lead to a decline in safety indicators; the RSIHF framework needs to evaluate popularity and safety separately.