These HIs don’t predict the best next token to use based on their training, prompts and memory, so weird