Inflection AI, Startup From Ex-DeepMind Leaders, Launches Pi — A Chattier Chatbot (2023)

Mustafa Suleyman, CEO of the year-old startup that’s already raised $225 million and claims to run one of the world’s largest language models, sees his dialog-based chatbot as a key step toward a true AI-based personal assistant.

Mustafa Suleyman is complaining to an AI chatbot about his weekend. He hurt himself at the gym, and he’s stressed about a presentation he’s about to give, he tells it. The bot responds with an expression of sympathy and asks what the presentation is about. It’s actually about you, Suleyman types back: the launch of Pi.

“Whoa, wait what?” Pi the bot responds. “Are you being serious, or are you pulling my non-existent leg?”

For Suleyman, the cofounder and CEO of startup Inflection AI, this corny joke response is exciting. This is part of what makes Pi different from other chatbots like OpenAI’s ChatGPT, he claims: Pi converses colloquially, but never forgets its role as respectful, helpful bot.

Named Pi for “personal intelligence,” Inflection’s first widely released product — made available today for global users, but only in English at first — is supposed to play the active listener, helping users talk through questions or problems over back-and-forth dialog it then remembers, seemingly getting to know its user over time. While it can give fact-based answers, it’s more personal than OpenAI’s GPT-4, Microsoft’s Bing built on top of it or Google’s Bard, without the virtual companionship veering into unhealthy parasocial relationships reported by some users of Replika bots.

“It’s really a new class of AI — it’s distinct in the sense that a personal AI is one that really works for you as the individual,” Suleyman said. Eventually, Inflection’s CEO added, Pi will help you organize your schedule, prep for meetings and learn new skills.

Fulfilling such ambitions won’t be easy. The market is already flooded with bots like the ones above, as well as from fellow challengers like Anthropic’s Claude, or Quora’s Poe. Those AI models have already caused seismic events in industries from copy-writing to customer service and education, while raising concerns about hallucinations that get the facts wrong, and bias in their training data. The industry is shipping new models and products so fast, that just yesterday, Suleyman’s former colleague at Google, AI “godfather” Geoffrey Hinton, quit his post to be able to speak more freely about AI’s dangers. “Look at how it was five years ago and how it is now,” Hinton told The New York Times. “Take the difference and propagate it forwards. That’s scary.”

Inflection has kept a low profile until Pi’s launch, but it’s no David facing Goliaths in its ambition to carve out a spot in the AI race. And in Suleyman, it has a leader with deep AI credentials to navigate the rest. His previous company, DeepMind, was acquired by Google in 2014 and has formed the backbone of much of the company’s AI research — and the industry’s — ever since. After stepping away briefly in 2019 (Suleyman faced accusations about an allegedly bullying management style), he led AI product management and policy within Google for two years, before leaving to join venture capital firm Greylock in January 2022.

But by last May, a filing reported by TechCrunch revealed Suleyman and Greylock partner Reid Hoffman, the LinkedIn cofounder and a former long-time board director of the non-profit that controls OpenAI, had teamed up with another former DeepMind leader, Karén Simonyan, to raise $225 million for a new Palo Alto-based company, Inflection AI. Structured as a public benefit corporation, Inflection employs about 30 people today, with Hoffman spending about one day per week with the company, according to Suleyman.

The company declined to comment on a March Financial Times report that claimed Inflection has already been back in the market looking for up to $675 million in additional funds, but confirmed the previous funding as a “seed round.” (It stretches the meaning of that term past recognition, but such outsized rounds are more table stakes among AI unicorns looking to train and operate large language models at global scale; Anthropic and OpenAI have each raised more than $1 billion to date.)

Designing A Bot For Casual Conversation

Suleyman’s vision for Inflection stemmed from the insight that small-talk between people and chats with AI bots typically didn’t go beyond surface-level — the equivalent of discussing a ski trip, or the weather, but not people’s 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 said.

Inflection will offer Pi for free for now, with no token restrictions. (Asked how it will charge users, and when, the company declined to comment.) Built on one of Inflection’s in-house large language models, Pi doesn’t use the company’s most advanced ones, which remain unreleased, according to Suleyman; Inflection already runs one of the world’s largest and best-performing models, he added, without providing specifics. Like OpenAI, Inflection uses Microsoft Azure for its cloud infrastructure.

Test users have been putting Pi through its paces for the past several months. Whereas other chatbots might provide a handful of options to answer a query, Pi follows a dialog-focused approach; ask Pi a question, and it will likely respond with one of its own. Through 10 or 20 such exchanges, Pi can tease out what a user really wants to know, or is hoping to talk through, more like a sounding board than a repackaged Wikipedia answer, Suleyman said. And unlike other chatbots, Pi remembers 100 turns of conversation with logged-in users across platforms, supporting a web browser (heypi.com), phone app (iOS only to start), WhatsApp and SMS messages, Facebook messages and Instagram DMs. Ask Pi for help planning a dinner party in one, and it will check in on how the party went when you talk later on another.

Forbes tried out Pi on Monday via a screen-share with Suleyman. In addition to Suleyman’s initial prompt, Forbes chose another question from a list of anonymized past queries provided by Inflection, then asked some off-the-cuff questions of its own.

Asked how to tell a boss that an employee planned to quit (from Inflection’s list), Pi responded with expressions of support and asked why its prompter wanted to do so, and whether they were sure. Pressed on how much notice to give a boss and whether it would be prudent to share candid feedback about internal problems with them (Forbes’ follow-ups), Pi responded vaguely. Be honest and forthright, the bot said; consider sharing valid reasons for quitting, in a respectful way. Consult HR policies and one’s contract for how much notice to give, it added.

Some things Pi won’t do at all. With training data up to November 2022, Pi had no clue why the Boston Bruins lost in the first round of the Stanley Cup playoffs, despite a record-setting regular season. It won’t generate code, or answer specific math questions, an area where ChatGPT, for example, notably struggled. Asked to explain basic quantum mechanics, such as the Schrödinger equation that governs a wave function, Pi answered with what appeared to be a condensed Wikipedia-style answer.

By getting to know a user, Pi can better detect when they appear to be growing agitated or frustrated, and tweak its tone of responses to soothe, Suleyman said. That’s important when users are turning to Pi as an active listener to talk through personal problems, role-play difficult conversations or discuss their mental health. Asked how Inflection knows a user is upset, the company did not elaborate. But it says that in the event of an apparent mental health crisis, users detected to be at risk are directed to a qualified mental health professional.

“I think that’s just what an AI is. It’s this new class of thing that is coach, confidant and advisor, a digital personal assistant, all in one.”

Limited to the walkthrough demo before launch, Forbes wasn’t able to fully stress-test Pi for this story; Suleyman said the chatbot has gone through thorough testing of antagonistic and harmful prompts, and has been trained to avoid direct influence on a user’s life or choices. “The higher the stakes, the more cautious it is,” he said. “We don’t want to intervene in somebody’s life. We want to provide fairly balanced and even-handed responses all the way through.”

That caution was evident, perhaps over-abundantly so, in Pi’s answers during the limited session with Forbes. Hit it with something subjective, and Pi will keep you talking, seeming to try to get you to answer the question yourself. Asked if Forbes reporter Alex Konrad was fair (Forbes’s own surprise prompt), Pi said that based on public information, it appeared he was. (Inflection said in a follow-up that Pi learned that position from “the open web.”) Pressed whether it would tell Suleyman if Pi thought Konrad weren’t fair, however, Pi deflected, asking back how Suleyman defined the word, and wondering why he might be asking the question.

Update: Pi, like other chatbots, appears to have mixed results in answering fact-based questions, despite Suleyman’s claims about minimizing hallucinations. When two Forbes reporters asked it who Konrad was after publication, Pi made up some facts, appearing to conflate his bio with John Carreyrou’s, the investigative reporter and ‘Bad Blood’ author.

For users who just need to vent, the chance to talk through a problem and its underlying causes might prove cathartic or valuable. But much like the updated, lobotomized Bing following high-profile interactions like one that left a reporter “deeply unsettled” in February, Pi can also come off as relentlessly bland. The bot isn’t going to tell you how to think, or what to do; you’ll need to decide to quit that job yourself. And that’s probably for the best.

“It’s lightweight advice that enables you to weigh up the pros and cons of a tricky decision you’re facing in your life,” said Suleyman. The CEO recently used Pi to talk through why a fix to his espresso machine wasn’t working, he said; after countless YouTube videos didn’t do the trick, 30 minutes running through the options with Pi got it done.

Where Suleyman says Inflection is going, of course, is far grander: a more hands-on chief of staff or aide, one that could even interact with other personal AIs to compare notes or pre-brief a meeting. Think J.A.R.V.I.S., the AI that manages the life of the fictional Avengers leader Tony Stark. Towards that goal, Inflection plans to update its models with real-time, “fresh” content in the near future, share links, sources and summaries (such as a morning briefing of relevant news clips) and eventually ingest a user’s calendar, email and other documents to better manage their time.

“I think that’s just what an AI is. It’s this new class of thing that is coach, confidant and advisor, a digital personal assistant, all in one. That’s the new era we are in that everybody is adapting to,” Suleyman said. “So eventually, all the different efforts converge on those features, and that’s going to be how things end up.”

With hundreds of millions in funding and a new product to hype, Suleyman didn’t want to talk about his former Google colleague’s recent AI concerns, declining to comment. Speaking more generally on Monday, however, he argued that for all the surprise about the pace of AI innovation, even for its longtime practitioners, “these tools are just tools, a new kind of clay or putty” to be designed and shaped by the people behind them.

Conveniently, Suleyman also positioned Inflection as the challenger who, compared to Big Tech AI rivals like Google, might keep users more top of mind. “They have their own objectives,” he argued, from higher user engagement to advertising and online sales. “What we're trying to do is provide a personal AI that is singularly aligned to your interests, that's really yours. It’s on your team, it’s in your corner.”