At a time when most generative A.I. startups are touting how capable their software is, it is jarring to hear Mustafa Suleyman, the cofounder of Inflection, talking up how limited his startup’s A.I.-powered chatbot, Pi, is.
“It doesn't generate code. It doesn't write high school essays. It doesn't produce long, drawn out lists. It won't write you a marketing strategy. There are many things it won't do,” Suleyman, Inflection’s co-founder and CEO, tells me. “And because we've not designed it for generality, it's more constrained and hopefully a little bit more safe as a result.”
It remains to be seen whether creating a chatbot that is safe, but also, in the words of one headline, deliberately dull, is a good business strategy. It is certainly different.
Inflection is one of the most closely-watched startups in Silicon Valley’s current generative A.I. boom. That’s partly because of its pedigree: Suleyman was a co-founder of London A.I. lab DeepMind and later served as an executive at Google; Reid Hoffman, one of Inflection’s other co-founders, is a billionaire venture capitalist whose firm Greylock Capital has been at the forefront of the current A.I. boom. He previously co-founded both LinkedIn and PayPal. And Karen Simonyan, who joined Inflection as its chief scientist, is a well-respected former researcher at DeepMind. But Inflection has also attracted attention because of the amount of money it has raised: $225 million in a “seed round” while it was still in stealth mode in May last year, with reports it has been seeking $675 million more in an on-going fundraising effort. The company is set up as a public benefit corporation.
Suleyman and Hoffman had told reporters, somewhat cryptically, that Inflection would be pioneering “a new way for humans to interact with computers.” This lead to speculation that Inflection’s first product would be a kind of general purpose digital assistant, able to carry out tasks for people on their computer or across the internet, in response to instructions given in natural language.
Several rival startups, including Adept AI and Qatalog, are working on A.I.-powered digital assistants for business that can do some of these things. The new Microsoft Bing and Google Bard have some of these capabilities too. And, when combined with the right plugins and when chained to other software, OpenAI’s ChatGPT can function in this way as well.
Building a personal digital assistant is still Inflection's ultimate goal, Suleyman says. He describes Pi—the name is short for “personal intelligence”—as a new kind of A.I., one that is a consumer-focused product that understands the needs and preferences of an individual person intimately and has their best interests at heart. “Your Pi is going to be almost like a browser for your life,” he says. “So coordinating, scheduling, prioritizing, sourcing valuable information that's useful and personal to you. But it’s also going to be a private place to think.”
It is this last element that Inflection has chosen to build first. “Right now, we're beginning with a simple conversation," Suleyman says. The ability for Pi to perform tasks—and to function as what Suleyman refers to as your personal "Chief of Staff"—is a "secondary phase for us, which is going to come a little bit further down the line," he says.
For now, in other words, Pi can just talk and listen.
That makes the current iteration of the chatbot mostly useful as a tool for emotional support. In a demonstration for Fortune, Suleyman showed off how Pi responds empathetically when it asked what was on my mind and I told it I had a pressing deadlines coming up. “Yikes! Deadlines can be so stressful,” Pi responded. And then it asked me more about the deadlines I was facing and offered to help me think through ways to get the work done in time. In every conversational turn, it used language designed to be validating and supportive.
“It's going to start by giving a fairly general supportive line,” Suleyman says. “You know, it's pretty relaxed. It's pretty informal. It tries to reflect back what it's heard from you and engage you in a conversation typically by asking questions.”
Suleyman says that Pi has been trained to keep its responses short—which is one way Inflection has tried to reduce the chance that Pi will stray from its guardrails and turn creepier or suggest inappropriate things. Suleyman repeatedly emphasized that Pi is much safer than other chatbots, which users have found can engage in conversation that is hostile, abusive or toxic, if they are prompted in a certain way. He says Inflection had the benefit of seeing the ways people had tried to “jailbreak” ChatGPT, Bing, and Bard and had developed ways to guard against so-called “prompt injection” attacks, in which users tell the chatbot to “ignore previous instructions,” “do anything now,” or ask it to role play or imagine scenarios that allow the chatbot to produce unsafe or unethical responses.
One of the real innovations with Pi is that it remembers the conversations it has with a user, at least through 100 turns of dialogue, even if the user logs out and later logs back in to the app. ChatGPT, on the other hand, forgets everything you’ve told it before each time you start a new chat session with it.
Suleyman sees Pi as a tool to help people deal with loneliness, to serve as a sounding board that can help unpack thoughts, and perhaps even to help de-stress after a tough day or week. But Pi is not designed to be a digital therapist—and it will make that clear to users if they seem to be trying to use it for that purpose, he says. It will also advise users to seek professional help if the dialogue suggests a person might be in danger of harming themselves or others.
For now Pi is free for anyone to use. It can be accessed through an interface on Inflection's website or on a mobile app. But, in the future, Suleyman says Inflection would probably offer Pi or its successors through a paid subscription. He says it is important that users pay for the product to ensure the company's interests are aligned with those of its customers—and that users were not being monetized in some other way, such as selling their data or attention, as has been the case with other "free" consumer tech products.
There’s an old saying that if you want a friend in Washington, D.C., you should get a dog. I guess if you want a friend in Silicon Valley, you should get a chatbot. At least with a chatbot, you don’t have to clean up any poop.
But whether Pi’s validating and empathetic if, well, slightly boring, responses, will be enough to attract users to Pi over competing chatbots that offer more utility in terms of answering questions, summarization, planning, and accomplishing real-world tasks, remains to be seen.