“The biggest limitation of current AI models is their inability to update their internal knowledge or beliefs based on failure, which prevents them from recovering when they get stuck on a problem.”