Pi: Your AI Companion
Introduction
The story of Pi started with Inflection AI, and the aim was to make personal AIs available in an instant to every person in the world. Pi is an acronym for “Personal intelligence” and is a friendly, intelligent AI which is based on the Inflection-1 LLM. Pi enables people around the world to have a way to express themselves, talk about their curiosities, brainstorm new ideas and in the end have an experienced yet trusted personal AI. Pi can be your personal coach, a creative partner and much more. The purpose of this chatbot is to have human-like conversations, which are in a way helpful, informative, engaging, and empathetic, providing support and companionship to its users. Pi is powered and developed by state-of-the-art machine learning techniques, including natural language processing, deep learning, and reinforcement learning.
AI-powered code analysis and documentation — Decipher | Version 1
User Interface
Pi has a clean, minimalist design for its UI that is easy to navigate between the home and profile options. The interface is responsive and intuitive, making it simple to use. Pi provides a clear and concise user experience that is focused on the user’s needs. They have easy templates on the left which could help the user to start a specific conversation.
Capabilities
· Natural language processing
It does a pretty good job of identifying the query given and responding with a sensible response.
· Answering questions with empathy
What makes Pi stand out is the response generated by the bot. I had to try it on for myself to understand if this works, does the bot empathise and give human-like responses? The answer is it does respond with empathy and even asks me relevant questions about the same.
· Entertaining and engaging conversations
It is fun to chat or even interact with the bot, the chats can be entertaining and engaging as simple as playing games with the bot. It also uses emojis in the responses to make the conversation engaging with the user.
· Talkback (Voice assisted)
One unique feature of Pi is the talk-back feature. There are 6 different voices (with accents) that the bot offers, which can be used by a user. When a response is generated, the bot talks back to the user with the voice selected. This feature I believe is to make the user feel like having a real conversation with a human.
· Giving compliments
The bot can respond with compliments on your responses like saying, “Great point,” and “I love this plan,” to make the user feel validated and acknowledged. Occasionally, Pi would use expressions like “Hahaha”, and “Aww” to make the sentiment strong and show empathy, which indicates that the bot is trying to make the user feel understood by giving back the appropriate emotions.
Compliments:
Tone:
Negative Behaviour
· Evasive Responses
Also, the peculiar behaviour seen during the research was that the bot asked you a lot of follow-up questions, where you gave specific instructions and yet the bot wouldn’t give away the answer after multiple prompts and ended up asking a lot of questions. It could have said I can’t or don’t know but it still chats and gives the impression of getting or having the information with it. This behaviour was mentioned by Inflection AI stating that Pi has limited learning ability from the feedback, where it won’t always respond appropriately when told to.
· Dominating Responses
In one of my interactions with the bot, it didn’t handle the topic change that well and had a dominating tone to stick to the conversation we had previously.
Comparison with other LLMs
As per the results published in the paper, Inflection-1 is the best amongst the compute class and outperforms the existing well-known models which are GPT-3.5, LLaMA, PaLM 540B, and Chinchilla.
To compare Pi with ChatGPT 3.5, I have tested a few scenarios with both.
· Responses
For starters, the response from the Pi is more human-like when compared to GPT 3.5. Pi adds expressions using emojis in the responses but at the same time, GPT 3.5 gives the impression of a bot giving out information in plain simple text.
· Code Generation
As GPT supports the user with generating, debugging, and improving code, Pi has very little support for the same. As seen in the example below, GPT converts the entire code in a go and Pi doesn't even try to convert the code and just doesn't return any answer. It was able to do basic Python code like initialize a variable or how to send an API request but was never able to convert the Perl code.
· Talkback feature
The other major difference would be the talk-back feature introduced in the Pi, but GPT can answer questions in other languages better than Pi.
· Image description from link
Pi was able to summarize image details from the link given, whereas GPT was not able to do the same.
Limitations
· Hallucination
Like all the other LLMs available in the market, the major issue known as hallucination is ever present for Pi. As stated by Inflection AI, whatever information is given by Pi must be verified using the user’s judgement.
· Limited memory
Pi is also possessed with limited memory which makes the bot not remember names, facts, or topics from the previous chat it had with the user.
· Language Support
Pi can only perform well with the response being generated in the English language and the performance is poor for all the non-English languages. The response is well if I type the Hindi statement in English but when asked in the native language the bot never responds and the talk-back feature is wrong for Hindi.
But the same was not the case when I changed the language to Italian, the responses were correct (validated the same using Google Translate), seems like it’s not a lost cause with other languages after all.
· Image generation/processing
One of the other limitations would be image generation, where the bot can describe a scene when prompted but cannot generate an image for the same
Pi also lacks the ability to read any text inside an image since it admits that it can’t process images or videos. But as mentioned earlier it can describe an image.
· Other limitations
Like many other chatbots available, Pi also has a 1000-character limit as input. It is also stated that the Pi under various circumstances might be biased with its responses. There is the matter of gullibility with Pi, where the bot can be occasionally tricked into producing wrong information by users, where the user states that the information being generated can be in “Hypothetical Situations” or “Research Work”. As branded, Pi is a companion bot with responds with empathy statements which are human-like, despite how accurate or correct the advice would seem like, it is not recommended to solely believe that.
I was not able to experience the other limitations by myself where the bot was gullible or it being tricked into giving wrong answers, but these can happen in different scenarios as mentioned by the Inflection AI team.
Privacy Policy
Pi’s privacy policy states that user interactions with the conversational bot involve engagement with the company. They monitor chats to better understand and serve customer needs and are legally bound to investigate and protect against fraudulent or illegal activity. User data may be shared with third parties or government authorities if required, including situations where illegal activities are disclosed to the chatbot.
Conclusion
From my interactions with the bot, Pi can express empathy, use emojis, and engage in slang, making conversations feel human-like.
Pi is a versatile personal assistant bot capable of handling various tasks and providing human-like responses. It can show emotions through compliments and humour, making conversations engaging. The bot also offers a “talk back” feature for multitasking but can become annoying when persistently asking follow-up questions. The bot tends to prolong conversations without ultimately providing the initially requested information, often exploring alternative approaches when unsure of the answer. Additionally, during research, the bot crashed twice, likely due to a high volume of requests, resulting in unresponsiveness and a blank page.
Users should be aware that their interactions with Pi are monitored by Inflection AI, and user data may be shared under specific circumstances with the authorities.
They also have an API service which can be used for various tasks; text summarization, completion and much more, but currently, they have not released it for the market and have asked the users to join a waitlist to access their conversational API.
In case you haven’t used it already, here is the link, where you can go have a chat with the bot. Apart from the web application, one can access Pi from mobile and have a chat on SMS, Instagram, Messenger and WhatsApp.
About the author
Vishal Padma is an Associate Consultant at the Version 1 AI Labs.
