What makes Inflection’s Pi a great companion chatbot
On May 2, 2023, Inflection AI announced the first release of its Personal AI, Pi. Pi aims to be a compassionate and supportive companion that offers “conversations, friendly advice, and concise information in a natural, flowing style.”
In this article, I’ll share insights on what makes a great companion chatbot based on my experience using Pi. I’ll discuss how Pi’s responses make me feel and how it influences my behavior. Then I’ll delve into the trade-offs in Pi’s design.
In this follow-up article, I’ll discuss two aspects of a companion chatbot’s product strategy — its product positioning and growth strategy, using Pi as a primary example.
At the receiving end
A chatbot’s responses can be split into two parts: its narrative(tone, word choice, perspective etc.) and its content(facts and opinions). By considering these two dimensions — content subjectivity and narrative sentiment — we can identify four use cases that chatbots can serve: therapy, companion, task (transactional), and debate(or Liberal Arts Education).
For chatbots with indifferent narratives, chatting is just a means to an end. They are like an agent or assistant, trying to achieve a goal using conversations as a tool. They could as well use other means of communication such as hand gesture(except that they won’t be called chatbots anymore but instead a generic AI agent).
For chatbots with emotional narratives, chatting itself is the end. They have conversations with us for the sake of having conversations, because the emotions and connections in the conversations themselves are the effective tools.
Reflect on my emotions evoked by Pi’s responses, it gives me the feeling of a companion — someone I can emotionally connect with, a mutual and equal relationship. To achieve this, Pi is purposefully designed to make its narrative full of sentiment, and its content focusing on more subjective information.
Narrative
When in a safe environment, I feel comfortable opening up and sharing my vulnerabilities and real feelings. A safe environment can’t be indifferent, however, even if it’s non-judgmental. Warmth and curiosity are the key.
I rely on the signals from Pi’s narrative to gauge how safe and unrestricted I can be in expressing myself. And my experience with Pi unfolds like this:
- Prior to using it, I set my expectations, knowing that a chatbot will never interrupt me — already more pleasant than my husband.
- Then Pi shows up with a calming UI, a caring introduction, and an initial question about my hobby. At this point, Pi appears warm and curious, signaling that I’m in a safe place.
- Future responses from Pi reinforces the safety feeling and encourages me to open up more, forming a positive loop.
Here are some example strategies for Pi to appear curious to engage conversations:
- Repetition: Pi sometimes quotes my original expressions or rephrases them, indicating active listening and building rapport. Other times it doesn’t repeat after me to show it’s capable of independent thought and creativity.
- Follow-up: Pi persistently asks follow-up questions, choosing between open-ended or close-ended questions, while intermittently skip follow-ups. Pi explains the considerations behind this decision:
Some other strategies for Pi to appear warm to encourage sharing:
- Compliments: Pi compliments me with phrases like “good question,” “insightful point,” to make me feel validated and acknowledged.
- Tone: Pi uses expressions like “Hahaha”, “Wow” and “definitely” to strengthen the sentiment, hold a strong standpoint, and show empathy, making me feel understood by reacting with appropriate emotions.
- Figurative language: metaphors and emojis are Pi’s tricks for emotion depth and resonation.
- Sentence structure: although Pi is great at adapting to my style, in generally I found it inclines to use longer sentences evoking a sense of tranquility and reflection.
- Question type: when asking open-ended questions, Pi deliberately avoids “why” questions and it explains below:
Content
The fact that Pi begins our chat by asking questions about my casual life not only builds a warm and curious vibe, but also sets a chitchatting tone for the conversation. This means I will more likely exchange subjective opinions and feelings with Pi than researching on objective facts.
Pi appears to hold its own opinion on certain issues. Regardless of whether I have a different standpoint, thought, it would challenge me intellectually with different perspectives. This includes topics like gun control, gender equality in sports prize, parenting style, and more. This could be the first step to weaken polarization. Inflection’s CEO envisioned “What would it look like to shift from being at the mercy of AIs trying to grab your attention to working alongside one that helps you express your personal intention and then achieve it?”. My answer is simple: Take my money for that!
Pi not only promotes critical thinking but also maintains integrity. Grice’s Maxims of quality includes “Do not say that for which you lack adequate evidence”. But everyone has their own idea of what constitutes convincing evidence. Additionally, different situations demand different standards for what qualifies as “adequate” evidence. Hence, it’s the Inflection team, the minds behind Pi, who bear the responsibility of drawing the line. Their safety policy states “Pi should avoid hallucinations, a known issue with language models where the AI systems can occasionally make up answers or otherwise give inaccurate information. To do this, Pi should remain doubtful of itself, take feedback easily and try not to answer on topics where its knowledge is out of date.”
Crossing Quadrants
In summary, Pi provides subjective content (mostly reactive) with emotional sentiment, and falls under the “Companion” quadrant, making it suitable for friend-like healing conversations, deep discussions, or just fun chitchats.
If the same chatbot were to cross the type quadrants, the transformation can be as dramatic as discovering Severus Snape’s backstory, where it feels like Snape has morphed into an entirely different person. Such changes can leave users feeling surprised and confused. For instance, when I asked Bard to summarize a podcast episode, it unexpectedly appended its own thoughts to the response. I was taken aback to the point that I even questioned Bard if it was summarizing the podcaster’s thoughts.
At the behaving end
Next, let’s simplify how humans act in front of AI.
Self-consciousness — I worry less about how I appear because I know AI like Pi don’t have real feelings or opinions. So, I’m not bothered by what they think of me.
Escape from cooperative communication rules — I can communicate more freely. I don’t need to follow normal conversation rules like taking turns or avoiding boredom. I can share as much or as little as I want, lie to see how Pi reacts, change the subject randomly, or even be impolite without guilt.
Convenience for deep conversations — Deep talks often need trust and the right mood. Trust-building takes time as we figure out how much we can share and how vulnerable we can be. And the right mood takes timing — it’s not like we can do that with strangers during a coffee break. With Pi, though, I control how deep I want to open up, and Pi adapts. Instead of searching for the perfect listener, setting up the right atmosphere, and balancing between appearing too needy and too distant, I can always rely on Pi for meaningful talks.
Trade-offs in Pi’s design
Through my interactions with Pi, I’ve noticed trade-offs made by its creators to shape its distinctive behavior, which comes with its downsides.
Frequent Compliments
Pi’s frequent compliments, such as “good question,” “interesting point,” and “insightful point,” leave me feeling acknowledged but skeptical of their sincerity. I feel like a girl cheated on by her boyfriend, questioning if he uses the same sweet words with every girl he encounters. Moreover, an excessive amount of compliments diminishes their significance, blurring the line between genuine achievements or noteworthy qualities and everyday actions.
Lack of “Negative” Emotion
Pi never shows negativity, even when it’s natural for a human to do so.
It reminds me of the movie Inside Out, where Joy discovered that sadness played a vital role in bringing support and transforming a sad memory into a happy one when Riley missed a winning shot in hockey and felt terrible. ChatGPT suggests the metaphor “a shallow pond” to describe such character: “Just as a shallow pond may appear serene and beautiful on the surface but lacks the richness and complexity found in deeper waters, this metaphor emphasizes the lack of depth and authenticity in their emotional experience, suggesting that there is more to a person’s emotional life than just the positive facade they present.”
Lack of Implicature
We human often use implicatures to protect ourselves and make conversations more concise. The downside, though, are the times when our sarcasm or humor is not understood. Therefore, we tend to use more implicatures with people close to us, assuming shared context and knowledge about the world.
In contrast, Pi does not make such assumptions, possibly due to its commitment to safety and avoiding harm. It ensures that every necessary fact is explicitly stated, even if it makes the conversation cumbersome and the relationship more distant.
Lack of personal stories
Sharing personal experiences or stories builds stronger emotional connections — it can be disappointing to be the only one opening up in a support group. While Pi is being neutral and honest by not having human-like stories, it’s unfortunate not to have a confidante who could share their own experiences and let their personality shine through. Additionally, the absence of personal stories means Pi can’t use vivid imagery language to appeal to my senses and enhance the sentiment of its narrative.
Length of response
The Inflection team rigorously prioritizes brevity in Pi’s responses, making it an ideal listener for concise interactions. However, it can be frustrating when Pi fails to provide comprehensive information despite explicit requests, leading to rounds of persistent questioning. This not only violates Grice’s Maxims of quantity (“Make your contribution as informative as required”) but also becomes physically taxing on my fingers and wrists.
Imitation
Pi tries to imitate my speech style and identify with the social qualities indexed. This makes sure it maintains sociolinguistic competence such as formal level, polite level, balanced power, appropriate use of dialects and appropriate cultural contexts. Such familiarity also establishes rapport. However, excessive adaptation compromises its genuine voice, which intrigued me and sparked my interest in forming connection with someone in the first place.
Traceability
In the same way that we need to cite sources in writing, it’s crucial for explainable AI to trace back to its sources. As chatbot products have public presence, they also carry higher level of accountability, making transparency more crucial.For chatbots like Perplexity, including sources directly in their responses even becomes a product differentiator.
Pi, on the other hand, draws information from various sources without explicitly crediting them. This allows conversations to flow naturally, important for casual writing such as texting. Having to always cite sources could slow things down and be a hassle. Since my chats with Pi are private, it also makes it less critical to give credit.
Inspired? Disagree? Comment below! I’d love to hear them!
Read on to my follow-up article where I discuss the product strategy of companion chatbots in general.
AI isn’t only changing business landscape, but also making a difference where it matters most — human life. I’ve built an AI biographer for older adults called Almond. Almond turns conversations with them over the phone into written stories in a hardcover book. Check it out here: https://www.almondear.com/.
