An Unexpected GPT-3 Decision in a Simple Gamble — LessWrong