Inflection’s CEO on how the startup has pivoted in the past year

Once among the highest-flying AI startups, Inflection has spent the last several months shifting its focus to more down-to-earth tech.

Around a year ago, the startup’s path was forever altered when Microsoft inked a deal to license Inflection’s models and hire much of its staff, including co-founders Mustafa Suleyman and Karén Simonyan. Since then, CEO Sean White, who joined to replace Suleyman around the same time, has charted a new course focused more on practical enterprise use cases and less on chasing superintelligence.

White is zeroing in on a few key advantages—the conversational or emotional intelligence of its models, their ability to run on-premises or in private clouds, and its institutional know-how.

Tech Brew spoke with White about Inflection’s new mission, why he sees the AGI race as a distraction, and what he thinks of the hype around agents.

This conversation has been edited for length and clarity.

Tell me about how Inflection has pivoted since the Microsoft deal.

I want to spend a moment on this because I want to clarify a couple things. We’re an independent company. All Microsoft did is they licensed our code and our intellectual property. They did not get access to any of the user information—there was no acquisition. They had to make a copy…Now, Mustafa did leave. He went to Microsoft to become CEO of Microsoft AI and I was brought in to replace him. And that was the point where we started the pivot. And maybe a little bit about that: I joined for several reasons. My PhD is in computer science; I have always been, just in general, interested in how we use technology to make the human experience better. And in my mind, [Inflection is] a public benefit corp that is trying to do the same thing, focused on AI, and one of the seven companies that is a leader in building these frontier models.

And it is weird. With my academic background, you would normally say, “And here are the papers for all these different things.” A lot of the growth around this is really in what almost feels like alchemical texts, which is really the knowledge passed down to get these large models to converge. And so it’s still, like, we can converge hundreds of billions of parameters. Google obviously can at DeepMind. Microsoft can now, because they licensed our code. But Apple can’t, Netflix can’t, Amazon can’t. And that’s worth keeping in mind. The ecosystem is changing so quickly, but there’s a lot of important know-how.

So when I joined, I hired a new leadership team. I brought on some folks that I knew and trusted or who were exceptionally talented…We kept Pi [Inflection’s chatbot] going. Part of that was, in my mind, the execution of the public benefit corp, but it was also that it was a great example of one of the characters that we have, which was this really, really high conversational intelligence. Sometimes we’ll talk about it as emotional intelligence. There are lots of versions of that and where that comes into play, but it meant that it was and is still best in class. If everybody’s racing toward the IQ, there was a place that we were also racing toward, and had spent a lot of money and time and effort on a unique data set built over two years to create. So then we opened up an API so that commercial users could start building things and testing it out and getting a feel for it, and started running these voice-of-the-customer meetings…There were a couple of commonalities. One, we felt like for a large number of them, they cared about the conversational intelligence and fine-tuning to their corporation. Two, some subset of them really cared about the aspect of their data never leaving and going to another system, particularly regulated industries.

In what kinds of enterprise use cases is this sort of conversational intelligence most important?

I’ll give you a couple of examples. For the API, they are just starting to experiment. Sometimes that is what you might think of, like, someone is building a system to talk to a spreadsheet or we have people who are looking at it for wellness...We have hundreds of people using the API paying for that quality. And then there are the folks who want it on-prem and private and have the conversational intelligence. One of the most recent ones—I can’t tell you the name of the company—but they are using it for wealth management, where, one, they cannot let that data leave their network. It has to stay there. And two, it’s really important for them to actually have a system that has a high level of conversational intelligence and EQ.

You might say “Sean, why does that even matter?” It comes from a couple of things. One, they are using the system to help them in what they say and do and interact. We bought two companies, one called BoostKPI, one called Jelled.ai. BoostKPI helped us accelerate our work on doing data analysis; we could have done a lot with the data itself and data warehouse, and Jelled.ai, same thing, except around things like email, accounting, all of the sort of organizational systems. But none of that matters if you’re just getting like a book report, as opposed to the Socratic dialogue that lets you actually get something done, learn something, understand something. And so the wealth management example of this, we’ve had folks who are using and experimenting with internal support or internal data access, and some you just need to have private access to very large-scale data sets.

How does being structured as a public benefit corporation play a role in your new mission?

I was around for the first internet revolution. I did a startup then as well, and you get to choose how you steer the industry, and that matters both for the industry and for society. The public benefit corp means that we can make choices that are aligned with this being better for society, which oftentimes I think is better for the business. I’ll use the example of—I feel bad, not intentionally—but Character.ai had some really bad incidents. That was bad for people, also really bad for the AI industry. And so the public benefit corp, and we have support from the board, lets us make choices that may on the surface look like they are giving Pi away for free. We do that, and there are people who every day send us messages saying, “Oh, this is great. It helps me do something.” But we also learn about how people are using these systems at work.

Trust is this thing that takes so long to earn, and you can lose it in seconds. And I think the public benefit corp means that it also gives people a way to at least ask us, “Why should we trust you? How do we know you’re not going to just make some decision and abandon us?” And I like that structure. It means that we get to have a discussion.

You’ve said in other interviews that Inflection is no longer focused on building next-generation frontier models. Does that race not fit with this new enterprise-focused mission?

Actually, our CTO, Vibhu Mattal, just put a post up about a new family of models that we have. And so the point I was trying to make is, if there are folks who are trying to build these 500,000-GPU systems, million-GPU systems, we’re not going to go into that race…We are still generating new system models that have more parameters than our previous models, that are state-of-the-art. We’ve been using a mixture of experts for a while…that have competitive scores…But I didn’t want to get into this “we’re racing into AGI.” I think it’s a distraction for us. I’m much more interested in, “How do we build things to augment people and augment organizations,” as opposed to, “How do we make AGI?” Literally, it’s a design criterion–—which one do you want to focus on?

What do you think of all of the hype right now around AI agents? Is that something you’re thinking about?

My team hates me sometimes, because when someone says, “Oh, it’s agentic,” I say, “Really? What does that mean?” Even in the late ’80s, early ’90s, people were working on these agentic systems. In that case, what it meant was like one system talks to another system and they exchange some stuff; they go away. And right now, that feels like it can mean anything from “I looked at all the pixels, and I watched what you were doing…I see what you’re doing, and now I’m going to mirror that for you and kind of help you do it.” And I agree, I think that’s interesting. Or it might be, “I have a whole set of tasks that are all piled together, and I’m gonna put it in a cron job.” And it might also mean, “Oh, well, I have a whole series of systems that represent me and a whole series of systems that represent you, and they’re gonna talk to each other, and they’re gonna do their thing.” Those are three models of agentic AI today.

And I think for us, what we have done has been pretty practical. So we have research work…around agentic AI, the various flavors of it. We also announced a partnership with UiPath.

And part of this was that reading pixels is fantastic. It’s really cool. It works 40% of the time. That will improve the same way that speech recognition improved. There’ll be some disruptions. But right now, we need something that works. And so I think of it as a stepping stone where we use some of the systems like UiPath, where the LLM is the probabilistic part, and UiPath is the deterministic part, and we build from there.