Sam Altman has said that people saying ‘please’ and ‘thank you’ to AI models is costing OpenAI millions of dollars in extra compute costs, but there might be a practical reason to be polite to AI models.
Google DeepMind scientist Murray Shanahan has offered a compelling perspective on this seemingly trivial act of politeness to LLMs. His explanation, surprisingly grounded in scientific reasoning, suggests that saying “please” and “thank you” to AI could actually lead to improved performance. Shanahan’s insight offers a fresh way of thinking about our interactions with these powerful tools.
“Do you say please and thank you?” Shanahan was asked on a podcast. “You can say please and thank you. There’s a good reason—good scientific reason—why that might get better performance out of it,” he said.
He continued, explaining the rationale: “Say it’s role-playing at a very smart intern, right? Then it’s going to just role-play maybe being a bit more stroppy (bad-tempered and argumentative) if they’re not being treated politely. It’s just mimicking what humans would do in that scenario.”
The mimicry, Shanahan suggests, might have practical consequences. “The mimicry might extend to kind of being, you know, not being as responsive if their ‘boss’ is stroppy.” He added with a touch of humor, “You know, ‘See, boss, I absolutely love that!'”
Shanahan’s analogy to a “stroppy intern” provides a relatable framework for understanding this complex interaction. Large language models, like skilled interns, learn by observing and mimicking human behavior. They are trained on massive datasets of text and code, which include examples of both polite and impolite interactions. Consequently, when users treat the AI politely, they are essentially providing it with positive reinforcement for desired behavior. Conversely, being rude might inadvertently reinforce less helpful responses.
This perspective has significant implications for the future of human-AI interaction. As AI models become increasingly integrated into our lives, understanding the nuances of their behavior will be crucial. While Altman’s concern about computational costs is valid, Shanahan’s argument suggests that politeness might not just be about social etiquette but also a practical strategy for optimizing AI performance. Perhaps a little courtesy could go a long way in shaping the future of our relationship with these powerful tools.