On The Risks of Emergent Behavior in Foundation Models — LessWrong