Exclusive: Inflection AI reveals new exec team and plan

Join us in returning to NYC on June 5th to collaborate with executive leaders in exploring comprehensive methods for auditing AI models regarding bias, performance, and ethical compliance across diverse organizations. Find out how you can attend here.

When Mustafa Sulyman, a leading figure in artificial intelligence, left Inflection AI in March to head Microsoft’s AI division, the tech world buzzed about Microsoft’s gain in the ongoing AI talent war. 

Less discussed, however, was the impact of this departure on Inflection AI. With Sulyman and approximately 70 team members moving to Microsoft, the question loomed: What would become of Inflection AI, which had secured $1.525 billion in funding and spent a good chunk of it to develop an emotionally empathetic personal AI assistant named Pi?

In an exclusive interview last week with VentureBeat, Inflection AI unveiled its new leadership team, composed of seasoned Silicon Valley veterans.

Sean White, the newly appointed CEO, has a background in user experience and augmented reality, most recently leading Mozilla’s R&D.

VB Event

The AI Impact Tour: The AI Audit

Join us as we return to NYC on June 5th to engage with top executive leaders, delving into strategies for auditing AI models to ensure fairness, optimal performance, and ethical compliance across diverse organizations. Secure your attendance for this exclusive invite-only event.

Request an invite

Vibhu Mittal, now chief technology officer, previously worked on early forms of generative AI as a graduate student, and later helped build Google Translate.

Ted Shelton, the new chief operating officer, comes from Bain & Co, where he consulted for enterprise companies on AI applications.

Ian McCarthy, leading product, brings experience from Microsoft, Sony, Yahoo, and LinkedIn.

Despite the Suleyman exodus, co-founder Reid Hoffman, a venture capitalist best known for founding LinkedIn, has opted not to disband but to double down. He’s aided by a warchest of less than the reported $650 million that Microsoft paid Inflection in a deal to take Suleyman. Hoffman wouldn’t say how much the company has left, but in an interview with VentureBeat said the company has “real money. It’s well funded for 18 months.”

Today, the company announces a refocused mission: To double down on its leadership in the realm of emotional intelligence within AI, at a time when industry giants like OpenAI, Microsoft and Google prioritize cognitive prowess. And its business plan is to build empathetic chatbots for other businesses – bots that can adapt to and reflect business’ unique communicative needs.

Hoffman provides a recent example of how Pi stood out for him versus other leading models like ChatGPT and Gemini.

When a friend lost a loved companion cat, Hoffman said he first asked the leading traditional models what he should do to console his friend, and they all responded roughly the same way, with a list: for example, getting a friend flowers, or offering to help them with daily things.

But Pi responded differently: “That must be really hard for your friend, and because you’re a friend, you care about that,” Hoffman recalls it saying. “But you know your friend. What way would you think your friend would most want you to be present for him?” In other words, Hoffman said, Pi knows the list that everyone else does, but it responds with less of a Wikipedia listing, focusing instead on the “emotional fabric” around the question.

The new team says its commitment to what it calls “EQ”, or emotional quotient – something the team says is a crucial but often overlooked component — will set it apart in a market clamoring for genuinely helpful chatbot interactions.

Sean White, the newly appointed CEO, framed the pivot during an interview in the Silicon Valley conference room at Greylock Ventures, where Hoffman is a partner. The leading LLMs, he says, “have lots of knowledge, they can have lots of IQ, but they’re not necessarily actively listening.” 

Team background

In my interview with White, he’s joined around a conference table by the three other new leaders of the company, all of whom Hoffman played a role in recruiting. They’re all seasoned Silicon Valley veterans, in many cases having worked with each other at previous startups. 

The throughline of the team, they say, is a deep experience and alignment around working on user experience. And they’re joined by a vision, they say, to make technology that not only understands and delivers what users are wanting – but makes users feel better about the experience afterward.

The team talks about each of their accomplishments, and why they’ll help with the pivot.

CEO White, for one, invented the phrase “situated visualization,” which refers to the idea of changing visual information so that it shows up as you’d expect as you move or your physical environment changes. He’s also spent considerable time investing in the area of neuroscience, giving him an appreciation of the role EQ plays in dialogue.

Mittal, meanwhile, worked on early generative AI starting 35 years ago before it was popular. Shelton’s work for enterprise at Bain included not just advising them on which apps to AI build, but also how to overcome challenges in deployment.

And McCarthy, who was most recently at Microsoft for two years, helped shape the company’s new AI commerce products before and after the launch of ChatGPT. 

Focus on EQ

The team acknowledges that EQ is not a very well researched domain in computer science, largely because the field has focused so much on IQ.

The lack of a widely accepted benchmark for emotional intelligence levels in GPT-4 class LLMs emphasizes the uniqueness of Inflection’s focus, Shelton says. The company plans to announce an initiative to establish an objective industry benchmark, he said. One exploration of “empathetic conversation” by a UC Berkeley professor last year found that Pi excelled against Claude, Chat-GPT, and Google’s leading model at the time, Bard.

EQ in AI means the system can perceive the mood of its users and respond in ways that feel considerate and helpful, they said. Businesses want chatbots that can do this, but haven’t figured out how: Right now you call up businesses, and you get a random rotary of people who don’t know you,” says Shelton. “What if you call up Hilton, or whatever hotel chain you belong to, and it remembers you? They say, ‘Hey Matt, great to hear from you. Last time you booked us in Palm Springs. You’re headed there again!’”

And what if, further, businesses could give their AI agents a specific tone and personality they can use across communication channels? One financial company Inflection is working with said it takes weeks to train a new human contact center agent on things like brand values and how to talk to customers, Shelton said.

Inflection AI’s technological innovations

Inflection AI’s advantage is that it has worked on EQ for the past two years. Its half a dozen models are trained on large datasets of emotional conversations between real people, which teach it how to respond to people expressing deeply personal issues. The company then additionally refines its models with what it calls “empathetic fine-tuning,” which customizes its personality. This fine-tuning is almost at the scale of pre-training itself, says Mittal. 

Because it’s part of the fine-tuning, the personality customization actually changes the model’s so-called “weights,” thus ensuring personality becomes stable in the model, White said. “An LLM personality generated through context and prompts alone will frequently get confused about which personality it is,” he said. 

The approach contrasts to competing leading bots, he Mittal says, that have focused on size of pre-trained data. These bots used the same training data that everybody else does, which is the corpus on the web. But for EQ, that data is much harder to come by, he says. “There’s no label corpus around it,” said Mittal. “And we have access to hundreds of 1000s of fine-tuning examples that we can use to try and improve any model that we build.”

To be sure, the outgoing team had spent huge amounts of money to build foundation models that were strong in IQ too. Before Suleyman’s departure in March, the company released its Inflection 2.5 model and claimed it was “more than 94 percent the average performance of GPT-4,” OpenAI’s then top-of-the-line AI model, at accomplishing IQ-related tasks. 

Inflection AI said in March that its latest model performs at more than 94 percent the level of GPT-4

“One of the reasons we are so good is we have the GPT-4 equivalent IQ,” says White, “but we have the best EQ out there…We have the platform, the framework for not just doing the foundation level training, but also the fine-tuning.”

Another aspect of Pi is that it offers a voice module, and there too, the conversation is characterized by a supportive tone, ensuring a more empathetic interaction. Shelton said that when he talks with users of Pi, they often say they use it to problem solve, with several saying they used it in the car while driving. 

The importance of memory

Pi’s personal touch is what sold White on joining as CEO. He recalls a recent trip to visit family, when he asked Pi if there were Muffaletta sandwiches in town and where to get the best one. When he arrived back in San Francisco, he opened up the Pi app to ask it about something different, and was surprised when Pi asked him if he’d been able to try the Muffaletta sandwich. “There was something crazy and magic about that,” White says. 

That magic, of course, has known technical underpinnings. Part of that is memory, since Pi can remember at least 100 turns of conversations. It remembers who you are and the most important things about you. It asks follow up questions, in order to fully understand what you are asking of it. The app builds a hierarchy around your perceived needs, so that it doesn’t get distracted with singular backs and forths. 

These memory and emotion features are something that competitors are catching up with. OpenAI company last month upgraded its memory features in ChatGPT. And OpenAI’s new GPT-4o model introduced more emotion. 

While OpenAI and other chatbot leaders are likely to make improvements in their EQ offerings, Inflection’s Mittal says he thinks the company has at least a year before the others catch up. That’s why Inflection is pivoting to help businesses build personalized chatbots – because competing chatbot companies like OpenAI and Google are racing to serve the mass market, but not focused on customization. 

The company has a warchest, but needs to hire

The company does not need funding, White says, because Greylock is fully behind the company. It does need to hire talent, he said. It has a team of about 12 people now, but is expanding its engineering team, focusing on fine-tuning and platform engineering expertise. 

The app still enjoys about 6 million unique users a month, and over a million a day, and so continues to learn from the consumer interactions within the app. White also says that he has a commitment from several members of the departed team that built the core systems that they will continue to help with maintenance through the transition.

Going forward, the company is centering its efforts around a building an AI studio, to work with brands to build their own assistants. Inflection sees two main areas to serve businesses, via its APIs: empathetic customer support agents, and bots that serve internal employees, for example to answer questions for manager or employee needs. And third, it plans to license emphatic agent technology to platform companies that serve chatbot technologies to other companies.

While Inflection AI’s seasoned team champions the crucial role of empathy in AI, and appears to have a head, the real test lies in their execution. With its funding and experienced leadership, Inflection AI has as good a chance as any at going after a huge, yet fiercely competitive, B2B market.

Source link