Dear LLM, Take a Deep Breath and Carry On

How to talk to your LLM effectively by leveraging interaction rituals

Welcome to the Culturescapes by Kursat Ozenc. This newsletter examines culture through the design lens. It discusses big ideas and small practices for crafting better cultures in our personal, work, and social spheres. 

Today, we're exploring a fascinating concept: devising interaction rituals when engaging with LLMs. This key idea can transform our interactions with these powerful tools.

The zeitgeist around LLMs is a mixture of many emotions: awe, excitement, and panic. The fast pace of innovation in the space puts most of us in a constant FOMO cycle.

With all this frenzy, we find ourselves jumping between different LLMs (ChatGPT, Gemini, and Claude), overwhelmed with a stream of prompting wisdom from all sides, irritated by quasi-realistic images and hallucinatory responses.

But it doesn't have to be this way. As early adopters of this technology, we play a crucial role in shaping the future of human-AI interactions. By shifting our mental model from transactional to relationship-driven thinking, we can pave the way for a healthier and more fulfilling future.

How do we do that? One way is to reframe our interactions with AI. For instance, we can see our chats with AI as conversations instead of prompt-response cycles. A conversation is a living, humane thing. It has a beginning, middle, and end. It has context, roles, and boundaries. When designed meaningfully, AI can understand and respect these boundaries, making our interactions more fulfilling.

Speaking of conversations, social psychologists can teach us much about their dynamics. In his seminal work, Interaction Rituals, E. Goffman articulated that people devise interaction rituals in daily conversations to create cohesion and predictability. Interaction rituals are standardized, routine interchanges, such as greetings, salutations, compliments, and turn-takings. They facilitate our conversations by marking a conversation's beginnings, middles, and endings. These rituals are not only social niceties but also serve as a framework for understanding and structuring our interactions.

What do these interaction rituals look like with our LLMs?

Interaction rituals can manifest in our chats with Large Language Models (LLMs) in several ways, reflecting the social dynamics typically seen in human interactions. Here are some examples:

Politeness

People use greetings and expressions of politeness, such as "hello” “please“ "thank you." One study investigated whether polite language such as “please” and concluded that politeness indeed increases the LLM’s quality of responses. They also discovered that a moderate politeness brings out the best results, meaning that overly polite and overly rude behavior result in poorer performance.

Turn-Taking:

Conversations with LLMs often involve a back-and-forth exchange, similar to human dialogues. This turn-taking is an interaction ritual that helps structure the discussion, ensuring that both the user and the LLM are carrying on the conversation.

Feedback and Acknowledgment

People may provide feedback or acknowledgment to LLMs, such as saying "I see" or "Got it," which are rituals that confirm understanding and keep the conversation flowing smoothly. This kind of reassurance could create a feedback loop to improve the LLMs’ performance.

Emotional Expression

Although LLMs do not experience emotions, people may project emotions onto interactions, using emojis or expressive language to convey feelings. These emotional expressions can create a quasi-social interaction where the LLM responds appropriately to maintain the ritual of emotional exchange.

Role-Playing and Social Norms:

While talking to LLMs, assigning them roles is an effective way (e.g., as an assistant or advisor) to orient them and provide them with the proper context. Giving them roles reflects the ritualistic aspect of role-based interactions in human society. People and LLMs play specific roles in these interactions, each with expectations and behaviors. By adhering to these roles and norms, we can create a more structured and predictable interaction with the LLM, similar to how we navigate social situations in the real world.

Now that we've explored the concept of interaction rituals, it's time to put it into practice. I invite you to treat your next LLM chat as a conversation, not just a prompting exercise. To help you make this shift, here are four specific ritual ideas you can try in your next chat.

Take a deep breath and follow a step-by-step approach

Before you dive into your conversation, suggest your LLM ground herself with a deep breath and take a step-by-step approach. I love to imagine LLMs as a living, breathing organism with billions of neurons. Before firing up all those neurons, think of all those neurons breathing in and out to tingle themselves for action. The reality is less poetic but still fascinating. DeepMind researchers discovered that LLM models perform better when prompted to take a breath and think step-by-step. The logical explanation for its behavior is that LLMs scrape content from the internet and find better answers from Q&A forums where people use calming expressions to guide people to resolving challenges.

Assign roles like you are in a role-playing game

When talking with your LLM, please give them a role so they can orient themselves to your context before responding to your question. For example, you would like to brainstorm about your long-awaited vacation to Barcelona. Start with something like this: "You are a highly sought-after local guide who knows Barcelona inside out. Help me..." This prompting works in at least two layers. First, you narrow down the context and provide boundaries. Second, you give a filtering lens for LLM to harvest the best knowledge available in its data universe.

Be Socratic in the Middle

Treat the middle part of the conversation with LLM like a Socratic dialogue, where you are there to investigate the truth rather than accept it at its face value. This part is similar to an interview, where you probe into more nuanced and deeper insights. You can ask, "How do you do that?" "What if..." "Can you give me a specific example?" "Why?" I am still learning how to do this effectively with LLMs, but I also see a pattern of zooming in and out working better for me to gauge the responses' quality. You first start asking questions to widen the conversation. Then, you narrow it down with more specific questions. And the cycle goes on.

Finish it with a Reflection

Having gone through the zooming in and zooming out cycle with your LLM could quickly get overwhelming. After seeing a long conversation thread, you could ask your LLM to reflect and summarize your discussion. This reflection exercise will allow you and LLM to synthesize your learning. After this, you can either continue on the same topic or pivot to a new direction that emerges from the reflection.

This is a wrap for this issue. Until next time, take good care of yourself and your loved ones!