Reid Hoffman & Inflection AI’s Sean White on designing AI that makes us better humans

0:02Hey, folks. Jeff Berman here. We have something very special for you today. A fireside chat between longtime innovators and friends, both of each other's and of mine, our very own Reid Hoffman and Inflection CEO, Dr. Sean White. This conversation was presented by Inflection and recorded live at the 2025 Masters of Scale Summit. I hope you enjoy it.

0:33One of the things we're running into that's going to have a serious problem are going to be mistaking AI for friends, right? How do we get psychologically misled by the magic of these tools, devices, agents, companions, et cetera? So what do you think some of the really important things are relative to human psychology and interaction that should be design principles about this agentic universe that we're rushing into? I love this topic because I think it's really important, right? And I'm sure that every one of you out here has at some point used your first chatbot and felt, wow, that feels human-like.

1:13That feels really real, that sort of magical experience. And I think we want to keep that, by the way. I think that is actually part of it being a really good interface. At the same time, you want it to be transparent, right? And so you shouldn't feel bad about this. Your brains are built to recognize people in things. There are whole psychological conditions where you'll see faces in trees and other objects. There's a field of study called captology that was part of human-computer interaction as well that BJ Fogg had sort of created, which was noticing that we applied some of the same social rules to computers that we did to individuals, right?

1:54If this computer was really kind of critical, sometimes we think that's smarter. And so we think, ah, that computer is smarter than this other one. And so in terms of principles, the number one is transparency. That is always knowing that there is an AI present where the AI is present, not pretending that it is human or sentient or some other thing, right? Or conscious. Or conscious. That's right. There are some other design principles in here, though, that you want to balance. Making sure that you have a good interface, right? So you and I have empathy as humans, right? And you want those systems to act that way.

2:30but if it starts leading down a path where it's clear that the human is actually going down a bad path, an unhealthy path, being clear on that as well. That's a design decision. And so maybe a last one to think about is actually starting off by thinking about this as pro-social. Well, I think it's interesting to have a goalpost. Everything starts in design principles by saying, what are you trying to do? If you try and aim for AGI, you're going to build AGI. If you try and build great tools for humans, you will build great tools for humans. And so say also a little bit about what already some of the pro-social things that we've seen coming out of even the agents as they exist today.

3:13One example is around just helping people. With Pi, I've gotten emails and messages, letters around this. someone will reach out and say, hey, I was having a really hard time. My mom died and I couldn't reach a therapist and Pi helped me through it. But the important part that's after that is to also then re-engage with other people. That's a large part of how we think about the training, right? The data sets that we use for doing that. Also thinking about this not as a transactional thing, but a relational thing. The other place where it's been pretty interesting has been in mediation. This actually was a journalist and we had this great conversation.

3:55And then at the end, she said, you know, I use Pi. I said, oh, that's interesting. What do you use Pi for? My husband and I use Pi to mediate sometimes when we're talking. And it makes sense, right? The last example, and one of the things I think you should keep an eye out for, last year, we actually had run some experiments. We were running Pi with WhatsApp. So that's groups of people and individuals, right? Because, you know, in the end, I like people. If you're here, you probably like people. And so figuring out ways to help encourage that. And there must be some stories for you too. For sure.

4:29So I was part of the crew that persuaded Bill Gates that the GPT revolution was real. And what we had done is we'd gotten a challenge from him to say, I'll believe you have general purpose functional technology if without special training you can pass the AP bio exam at a five so we arranged a whole meeting we brought a U.S. one of the bio olympiads with us who she was a gold medalist like very very smart to cross check ask the questions all the rest and that was all persuasive but then it was the first time the bill that I saw him change his mind in real time because I was looking over at him and he was looking at this.

5:10It's okay, but what question do you want to ask? And he said, well, and I think it's because it happened with a friend of his. He said, say you had a friend whose treasured pet just died. How would you try to comfort them? And so we asked that of GBD-4. And GBD-4 gave a pitch perfect answer. It's like, well, actually, in fact, you probably know your friend better than I do. So the specific thing for your friend, I'm not sure. but here's some things you could consider it's just you go be with them you could offer to make their life a little easier you could tell them that you're aware of how much their pet meant to them and how much it was part of the fabric of their life and it had an incomplete both practical but also cognitive and emotional range and i literally at that moment i saw bill's face change right it was like that's great right because it showed all forms of human context awareness it was like i'm not going to tell you what to do you're going to be a participant in this and that's part of the whole agency thing i'm going to give you a pallet of things that you might consider i'm ready to continue doing that conversation and that design principle for these agents with exactly as you say, which is the focus here is how do you help people be better human beings in community and with themselves?

6:33Yes, I think it's incredibly important to bend that arc. And, you know, I'm a fan of the Socratic conversation. Tell me more about that. Exactly. Well, so if you assume that everything should be single shot, it would be, Reed, tell me story. And I get the story and then that's it, right? Once upon a time, there was a Dr. Sean. Yeah, that's right. But oftentimes, and you and I have done this many times before, trying to understand something, try and get it some truth or tell some joke, it's a back and forth. And so the Socratic aspect of that, the conversational intelligence goes along with the EQ, and it also goes along, I mean, I know you think a lot about not just the transactional, but the relational.

7:15Yeah, and that's part of the reason why it's an absolute mistake. Like I literally correct people that I talk to and say, oh, yeah, like, you know, Claude's my best friend. And I'm like, no, friendship is a bi-directional relationship. Part of what is actually essential in friendship is I show up going, oh, I've had a really bad day. I want to tell Sean about my bad day and sit down and I hear something about your day. Oh, Sean's had a much worse day. We're going to talk about Sean's day, right? And that is actually not just a question of kind of like load balancing. else. It's the way that by being friends to each other, when you allow your friend to help you, you're actually not just getting the help, but you're enriching the friendship, you're enriching the friend, it's extending the friend's capabilities, and that's part of the bi-directional thing.

8:00That's right. But we're a very long way away from AI friends. Yes, and rightly so, right? I think that the more interesting ways for us to connect get better. I can ask you, it goes a a little bit to the human-computer interaction aspect, a lot of that gets grounded in philosophy. I know you don't like talking about philosophy, but, you know. There are no limits to language. Yes, exactly. So what's your sense for how that drives not just what we have today, but what we should have in five years and ten years? Well, so here's one of the things to kind of think about this at some philosophical depth.

8:39So one, the evolution of tools changes our epistemology. Just think like microscope, telescope, et cetera. Our world changes through tools. Think about how AI is going to change our epistemology. So philosophers tend to have this mistaken conception. No, it's just pure thought. You sit there with your pipe in a library and it's pure thought. It's like, no, no, no. Our actual conception of the world is we're homotechnics. It's through the technology, through the conceptions of even what we think of, quote-unquote logically possible. So the notion of what is our ontology, what is our epistemology is going to radically change and evolve.

9:14Then think about, okay, if we learn new things about language, because language evolves too as a tool and it's one that constitutes who we are, that will also evolve what our concept of self is, our concept of the world, our concept of how we understand each other, what kinds of things there are. And so it's one of the things on the more subtle or philosophical side. And so the notion of, oh, this language works in a new, different way, like we've talked about music as a language over 20 years, something, like all of that changes the way we think, and then we'll change philosophy. Like one of the primal mistakes I think philosophers make is they think everything that is philosophically knowable is philosophically knowable right here just by thinking.

10:02And that is not the case. Right. We sort of co-evolve with our tools. I mean, this is one of the reasons why I care so much about how we develop AI, not just us, but the entire ecosystem, right, is we will co-evolve with these systems. I mean, the same way we co-evolved with language. Yes. And we co-evolved with writing. And that's one of the reasons to have design principles on the interaction with the agents. Like Sean at Stanford taught classes in human computer design. And actually, in fact, those are even more important now. Then it was like, okay, do you get the affordances right? Now it's, is my epistemology of how I think about the world and how I think about myself and how am I evolving that?

10:49How does that change? Ooh, epistemology is affordance. Yes, exactly. That's very complex. Yes, yes. Yeah, mind bending. Yeah, no, good. So part of the other thing, by the way, so I've given a couple of speeches in the last few years in Italy. I had been thinking a lot and intensely about a theme that goes on here in Silicon Valley, which is Silicon Valley's the next renaissance. And in some ways, that's very deep and very true. And some things that people don't even fully recognize, which is the renaissance was on a trade route. It had a massive elevation of both science and technology as part of what it was doing.

11:25but the thing that we need to recapture more and add more to what we do and as part of the reason why I think the Masters of Scale people do such a great job with this event is the humanism of it the humanity of it and I think that we only would get those laurels of being another important renaissance moment if we bring that humanism back in and that's where philosophy becomes very important because part of it is not just technological space of possibility, technological space of creation and building, but also what's your theory of current humanity and what's your theory of what journey we should be on, what we should be enabled to do with ourselves and with each other.

12:08And by the way, for those of you who've read Super Agency, I don't think it can be pre-planned. I think it's an iterative deployment. You go out, you learn something, you iterate. You go out, you learn something, you iterate. The frequent mistake is to think, oh, it's almost like the philosophical mistake. I can figure it all out just sitting here in advance. And almost nothing serious works that way. Yeah, that's right. One of the things that I have enjoyed about the push and the acceleration around AI is it's forcing us to ask, what does it mean to be human? Because I think we've been a little lazy about what it means to be human.

12:39It will evolve what it is to be human. And we have a question.

12:45My name is Ellie Barmasal from Paranaut. And a lot of my work has been thinking about what are the technologies and brands that are going to enable human flourishing over a century-long lifespan. And I think a lot of AI doomers say that AI is going to be the end of humanity. But I'm curious about another question. When you think about the current state of the human condition, what are the things that make us most radically human that we stand to lose without AI? Great question. I'll do three in very different directions. So one is obviously one of the amazing elevations of the human condition is spreading literacy.

13:24And it only became possible because of printing press and books and whatnot. And I think AI is obviously enabled by that, but it's the most amazing learning technology we've invented in human history. And the hope that it's not crazy to think we might be able to, within a small number of years, literally get to call it 80 % plus of the 8 billion people, over 8 billion people, and enable literacy, education, et cetera, to what we currently think of as an everyday Western wealthy world standard. And that is a huge evolution. So that's one. Two is the ability for AI to facilitate medicine is amazing.

14:16And one of the things I keep telling people is to say, if you get a serious medical concern, if you're not using one or more frontier agents as a second opinion, you're making a desperately bad mistake. I literally have had, my friend was hiking with his cousin in the Adirondacks. They had an issue. They went to a hospital. The hospital said, it's fine. Here's some antibiotics. I'm like, I'm not quite sure. Check ChatGBT. ChatGBT said, go to another hospital. And went to another hospital and that hospital checked and said, oh, I'm glad you're here. If you were two hours later, you'd be dead. Right?

14:50So the second opinion, I'd say still be cautious about first opinion unless you have no doctors. Right? But second opinion is super important. And then the third is, obviously, one of the reasons why I put such energy into the optimism case is not because I'm blindly optimistic. I actually think it's rational. But, like, for example, we obviously run more and more at risk of serious pandemics, whether they're man-made or natural or anything else. AI is the only thing I can really think of as a good defense against that. Everyone tends to go to, oh my God, AI is terrible because of cybersecurity risk or enabling a bioterrorist, which would be terrible.

15:29But the defense part of that, AI is the only thing I know that could be the defense. All of those things are humanity level, like billions of people. I don't know what you would. So I have a long list, but I think of two that I wanted to highlight. One, we have shortened our attention span for our technology. Sorry, what was that? Shorten. Oh my God, you got me. Well done. Well done. But my sense is that these systems can actually help us to pay more attention and help us with executive function. And that's not just to offload the executive function, but to retrain us on where executive function is positive.

16:14Without AI, we could be going down a sort of very short five second attention rabbit hole where we I go back to the days when we were mostly looking at pre-attentive cues in case there was a lion that was about to eat our head or something like that. Kind of related to this, my hope is that it helps us to value some things that we have been devaluing. Education is a good example. And in large part, it is because I think this can help in tutoring not to replace teachers, but to give more support. I mean, anytime someone has one-on-one support, along with the guidance from a teacher, that seems like a win in my mind.

16:55Nice to meet you. Hey, Sean. Hey, Reed. The thing that's been most surprising about the foundation models is how quickly they are growing and their capability to do general things. So you talk about the biology exam, but also math Olympiad and others, and these aren't specialized models. So there's a lot of people here who probably aren't working on foundation models. And so I'm curious, like Sean, for inflection and Reed for you generally, how you give people advice on here are the interesting things to work on outside of the foundation models that aren't going to be in GPT-6 and 7 and instantly taken away from you.

17:28Yeah, that's a great question. The world has been eaten up by LLMs. But actually, AI, which was almost an embarrassing thing to say you worked on for years, is a really broad field. And so if I was giving a recommendation, it would be look at all of the different forms in which we use AI. And some of those are rule-based systems. Something super exciting to me is the combination of knowledge graphs together with LLMs, for instance, where you have things that are like, we're getting into this regime where we have to be comfortable with probabilistic systems and putting together deterministic with probabilistic systems.

18:03And this is complex. Like the life's work of folks like Michael Bernstein is human in the loop for those systems, the work that Vakash is doing at MIT, right? And so when you hear AI, if you're not familiar with it, realize that it's about 15 disciplines, one of them which has to do with machine learning and deep learning and LLMs, and you're way down in this little node, but there's a whole world to actually be adding in here. So there tends to be this thought, hey, I build a business off GPT Model X, and then X plus one blows my business out of the water because it's just better than that and it doesn't really matter.

18:40And we've already seen that with some of the businesses that were built off GPT-3, for example, when 4 came out. Now, if you're building businesses, working on projects, think about the LLMs will be available and evolving in multiple ways. That'll be great. Chart your product or your service strategy based on an expected model evolution that's not you just always training your own model. One angle of that is a scaffolding because one of the things that we already have with today's, like I could literally say everyone in this audience is not using LLMs to the extent that they could to empower themselves.

19:17Like we are underusing these things right now. It's one of the reasons I force myself to do queries constantly. Like I try to get a trigger where I go, I'm thinking about a problem. Okay, what's the prompt? Just to get into using it. because right now they are much more capable than you think. And the reason that that doesn't show through, and even though it's a very fast adopted product, it's the scaffolding, it's the affordances, it's how it's engaged with. Now, if you just build a thin little wrapper around it, the thin little wrapper is a thin little wrapper. Other people can copy it. It might get blown out of the water, et cetera, et cetera.

19:51So you have to think about that as part of your strategy. Now, you might also do things that Sean was gesturing out, which is like, well, actually my AI agent combined with a knowledge database that's rare and important. Could be something that's important. It could be medical. It could be something else. I actually know of people that are doing this kind of things even for like, call it engine design, which by the way, is super important for climate. So you have to think about that you have a product or business strategy that takes the evolution of these models, including the fact that the first model you may have built with may not actually be the absolute best model next year, next six months, 18 months.

20:28You may have access to that other model. You may not. How do you operate in that environment in terms of that? And that's how you think about it entrepreneurially.

21:08Thank you.