A new AI chatbot called Pi is designed to serve as your personal assistant — here's how it works

• A new chatbot called Pi, launched by Inflection AI, offers personal advice and support.
• LinkedIn cofounder Reid Hoffman and DeepMind cofounder Mustafa Suleyman cofounded Inflection AI.
• Pi was designed to be friendly, but makes clear to users that it can't feel emotions.

There's a new AI chatbot on the scene — and this one wants to get personal.

Pi, which stands for "personal intelligence," was released Tuesday by Inflection AI. The AI startup has three cofounders: Mustafa Suleyman, the cofounder of DeepMind; Karén Simonyan, the company's chief scientist; and LinkedIn cofounder Reid Hoffman.

Pi is designed to be conversational and offer emotional support. In its opening dialogue, Pi tells users it's meant to be "useful, friendly and fun," and prompts users to ask it for advice or answers, or to talk about whatever's on their mind.

Suleyman told Bloomberg Pi is meant to be a "neutral listener" that asks follow-up questions and offers advice.

When I told Pi "life has been overwhelming," the chatbot probed me to share more.

"Many people feel like they just want to be heard and they just want a tool that reflects back what they said to demonstrate they have actually been heard," Suleyman told Bloomberg.

Like OpenAI's ChatGPT and Google's Bard chatbots, Pi is based on LLMs, or large language models. This means the bots are trained to generate text answers by analyzing large quantities of content available online.

In another conversation, I told Pi that I was considering graduate school, and that I worried about missing the application deadline.

Pi gave me advice on other options to consider, and asked how I could keep myself busy. 

The bot is designed to offer human-like support and advice, but also made clear that it's not actually human.

When I asked Pi how it was doing, it responded, "Well, as an AI I'm not really affected by the kind of things that stress humans out, but I am programmed to be empathetic and understanding. I know that humans experience a wide range of emotions, both positive and negative, and I'm here to be a supportive friend. So I guess you could say I'm doing pretty well!"

Suleyman told Forbes that his vision for starting Inflection AI was to make a chatbot that could respond to real-life problems.

"When we set out on this project a little over a year ago, I had the core question, what makes for a great conversation?" he told Forbes.

At one point, I asked Pi to share museum recommendations for a friend visiting New York City. Pi gave itself some autonomy by mentioning "my favorites," and included some museums "a little more off the beaten path." 

In general, Pi kept our conversation flowing with follow-up questions. At one point I asked if I could restart our conversation. Pi said I could change topics, but didn't offer a full reset of our interaction.

An Inflection AI spokesperson later told me that users can't technically "restart" their conversations with Pi, unless they use different login credentials.

Users can delete their accounts through the iPhone or iPad apps, or by sending an email request to Inflection AI with the phone number they used to register, the spokesperson said. Users can also request via email that Inflection AI delete its copy of conversations.

Pi has been trained on data as recent as November 2022, Forbes reported. Details that can make a person identifiable are not used in the training, Suleyman told Bloomberg.

The chatbot is currently free on a web browser or mobile download for iOS, and prompts users to add a phone number after 10 interactions so Pi can text them to ask how they're doing. Users can text the chatbot "PAUSE" when they want to stop hearing from Pi. Pi is also available to DM on Instagram and Facebook Messenger, and to text on WhatsApp or SMS. 

Like other AI chatbots, Pi can make wrong or inaccurate statements. Inflection AI said it's working to "minimize hallucinations." Inflection AI also said it's "creating a new form of 'boundary training' that will redefine how AIs learn and are trained." Users can flag problematic conversations on the web browser and apps.