#330 Harnessing AI to Help Humanity with Professor Sandy Pentland, HAI Fellow at Stanford, Co-founder of MIT Media Lab

Generative AI is transforming industries at an unprecedented pace. But as AI changes how you or.

Your team work, one thing is clear. Your skills also need to evolve.

At datacamp, we offer everything you or.

Your team need to adapt and thrive with AI.

Whether it's business users looking to get.

The Most out of ChatGPT and Copilot, or developers and data scientists looking to fine tune models.

You can learn the entire AI skills spectrum on Datacamp. Power your AI transformation today.

Start learning@datacamp.com.

The main thing, I think, is to keep in mind that ideas and stories are valuable. Learning about things, hearing the other person's story. Relationships that work best right across communities are ones that are essentially people trading, like in the sense of commerce. We don't have to agree about anything else, but we're helping each other. And if you ask people about folks where they have that sort of trading relationship, you find a great deal of trust and a great deal of value. You know, you need to be a trader in stories that are relevant to what you're doing. And if you can do that, that's good for you, good for them, good for your community. Everyone's very concerned about polarization. The people on the other side are crazy. And so we did a couple of big experiments. These are like 30,000 people statistically balanced for all sorts of things. And it turns out that most of this sort of people being polarized and disagreeing with each other is that they're just not familiar with what the other side actually thinks. There's so many loud voices, the people who may in fact be genuinely crazy, that dominate and just saturate all of the airwaves and the social feeds and stuff that you miss the fact that the vast majority of people don't believe that even on the other side.

Welcome to Data Framed. This is Richie. Most data and AI projects are team sports. In the good case of collaboration called the wisdom of crowds, we can achieve things as a group that no individual could do alone. Unfortunately, there's also the bad case called the stupidity of crowds, where groups perform worse than individuals. Today we're going to look at how storytelling and collaboration and culture can help us get to the wisdom state rather than stupidity. Beyond this, AI is the latest wave of technology changing how we communicate. And I want to know the implications of this. Today we have a very special guest, Alex Sandy Pentland, a faculty fellow at Stanford High and Toshiba professor at mit. Sandy is one of the most cited computer scientists in the world. He's a pioneer in computational social science, organizational Engineering, wearable computing, modern biometrics, and image recognition. In 2012, Forbes named Sandy one of the seven most powerful data scientists in the world. Currently acts as faculty lead for digital platforms and society research at the Stanford Digital Economy Lab. Previously, he was the founder of the MIT Media Lab. Sandy's also written three books, most recently Shared Wisdom, which forms the basis of today's conversation. So let's find out Sandy's opinions. Hi, Sandy. Welcome to the show.

Glad to be here. Let's have some fun.

Absolutely. I'm definitely looking forward to some fun. So I guess today we're going to be talking a lot about storytelling. So to begin with, can you tell me, do you have, like, a favorite story about data or AI?

Yeah, but first let me talk about stories because, like, this is about data. What are we talking about? Stories. Right. So stories are like theories or facts, but modest because, you know, you look at the data and you come up with a story and. And you maybe think it's a fact, but three months from now you might find out it's not a fact. So I like to use the word story because it's an explanation of data, but you are not like 100% committed to it. So that's why we're talking about stories. We're just trying to remember that we're wrong sometimes. Okay. And here's an example of.

Hi, Sandy, welcome to the show.

Glad to be here. Let's have some fun.

Absolutely. I'm definitely looking forward to some fun. So I guess today we're going to be talking a lot about storytelling. So to begin with, can you tell me, do you have, like, a favorite story about data or AI?

Yeah, but first let me talk about stories because, like, this is about data. What are we talking about? Stories. Right. So stories are like theories or facts, but modest because, you know, you look at the data and you come up with a story and you maybe think it's a fact, but three months from now you might find out it's not a fact. So I like to use the word story because it's an explanation of data, but you're not like 100% committed to it. So that's why we're talking about stories. We're just trying to remember that we're wrong sometimes. Okay? And here's an example of a funny story. This is a Yann Lecun story. So here's a little puzzle for your LLM. A man and a sheep are at a side of a river, and there's a boat with room for a man and A sheep. How do they get to the other side of the river? And this is such a stupid question that nobody actually wrote it on the entire Internet until the summer when he posted this. And so the LLMs couldn't answer it because it wasn't in their database. Basically it's like, well, that's crazy. I mean, a five year old would know what to do with this, but not your brilliant LLM, right?

Yeah, that's true. I mean normally you have like a wolf in there and you've got to work out who goes one at a time. So it's kind of an anti riddle, right?

And that's what they did. They went crazy trying to think about all. But no, no, this, this question is stupid. It's obvious and they couldn't get it.

Okay, So I do like the idea that there's still skill for human thinking in there because yeah, often LLMs do get things wrong. I think one, one of the recurring themes in your book is all about how technology changes the way we tell stories. So I guess talk us through, like how is AI impacting storytelling?

So the book, and I think a good way to introduce this is the question of like, you know, so we'd like to have good cultural change and all that. And the Enlightenment comes up all the time. It's like, oh, in the 1600s they had it right, Right. Well what did they do? Because there was already science and there was already all these things. But what they didn't have until the late 1500s was a postal system. They couldn't send letters back and forth to debate things with each other and nobody was keeping count. And the Kings at that time made the postal system available to common people who were discussing questions of why is the world this way? And the names, you know, most like Leibniz and so forth, would write one or two letters every single day their whole life to different scientists debating things. And so the King set up a center, the Academy of Science, this is in France, to keep track of this. And the people who wrote lots of letters and were really popular made, you know, got an appointment by the king. And that was really cool back then because now you had a good income and you got the hoi polloi and stuff like that that caused the alignment. Now what can we do today? Well, we got people, we've got electronic things. Can AI help us connect to each other better the way letters back then connected to people? And I think it can because you can ask, well, for instance, who in your company is doing the same stuff as I am? Let's get together or who knows about this or things of that sort. And also like we've built some systems, go to a pitch here, go to deliberation, IO, it's all open source, et cetera. Just use it. What it is is a way of people actually communicating to each other, deciding things without all the loud voices and BS that you get all the time. Right. That's what it's designed to do. And so, you know. And it uses AI to help to do that. Yeah, I think we can. What we shouldn't be doing is, you know, just abdicating and saying. Just say what's the answer to the AI? Because that's like answering, you know, asking the guy down the street, he'll give you an answer. You may not like it, it may not be right. You should really try and get into it more and talk to the guy down the street, but also think about it yourself and other people and stuff like that. So that's the basic idea of the book. Let's use AI to help ourselves be smarter.

Okay. So I would say it's really interesting that you putting the postal system as like the big driver of the enlightenment. I think a lot of people, when you think of like technology enabling things, it's like the printing press because then you could disseminate information one way. But this is about like debating stuff, sharing ideas and having backwards and forwards conversations.

Yeah, people think about the printing press, but you know what the printing press was mostly used for? It was used for pamphlets and so like luther. Right. The 95 theses caused 30 years of war. So it was like all the political BS you see on social media today, these like unregulated, not really a debate, just people spouting off. That was the printing press for the first 30, 40 years. Yeah, they did some Bibles on the side, but you know, we've had social media before and it wasn't good.

That's fascinating. And yeah, so maybe like the first wave of technology is kind of the bad wave where it starts off with like, I guess politicization and misinformation, things like that. That's the first wave of stuff. And then you get the useful stuff afterwards perhaps. So I want to talk about your sort of, you've got like a three step process for like how stories are used to make decisions. Can you talk me through what are those three steps?

Yeah, so you know, the way we think comes from evolution and hundreds of thousands of years of people who didn't do it right got eaten by the lions. So we have a particular way of doing things. And the title of the book is called Shared Wisdom. And that's sort of the key. What we do is we try to build a culture, a set of things that our work, our buds, all believe. And we do that by sharing stories about, oh, Jimmy did this and got eaten by the lion, or Mary did that and didn't get sick and it was good. We tell these things, we end up with an understanding of the world through our stories. The stories have some facts behind them and some outcomes and things like that. But let's be modest again. It's not like a fair sample. It's not like all those things that you need for science with a big S. Right. And so we build this shared wisdom. And when we have that and something comes up, then we among ourselves can say, oh, that's like this. And when they did, X, Y happened. And we can choose them and everyone will be on board because we all sort of understand the range of things that we do. So there's this shared sort of ideas about what you can do. There's shared, then actions that people take to get to a good place, and we can have collected action. All work together. Right. If we have that sort of shared set of things we believe that we've heard in our stories. And then for you yourself, this is the third type of story is, you know, you sometimes don't want to do what everybody else does. So you have to look at the culture, you have to look at what, you know, what other people are doing, what works for them, and you pick something for yourself. So there's stories for you, there's stories for action, and then there's this sort of shared worldview. And that's all constructed by listening to stories of other sorts of things that have happened and trying to estimate what works and what doesn't work and what is likely to be able to. Good for you. And so those are the three types of stories.

Okay, that's fascinating. And I guess, yeah, the idea is if you all or your peers or have a kind of shared culture that's going to decide on, like, what is a good idea to do, and you can decide whether or not you want to accept that.

Well, then you can get people to work together. Right. Because if everybody goes off their own way, the lion is going to eat you. All right, I'm sorry. It's just the way it works, right?

Yeah, I suppose a lot of the time, I mean, there's a sort of idea that you should be like a lion, not a Sheep and do your own thing. But actually doing like the sheep strategy is often successful. You want to do what other people do along with no.

The reason we have villages and the reason we put walls around the villages was because of the lions. And we all had to agree to do that. And it was good.

And then you asked the AI like what should I do? And it's not sure because there was no line in the story or something.

Yeah, well, the AI actually, I mean, this is one of the real weaknesses of AI. It has no context. This is the stuff that everybody says is in AI. It's not the stuff for where you work or where you live or you as a particular person. It's not personalized, it's not contextualized. And so it gives the average answer. It's like walking up to the random person on the street and asking a question. You'll get an answer. It may not be something that's useful to you or maybe anti useful. You really want something that's good for you and your community so that you can get things done.

I guess the key to this having is having a culture where you are sharing ideas. Like how do you set that up? Like how do you get this story sharing culture?

First of all, understand this business of we build a shared culture by telling stories. This is good, right? And then we pick actions from the things that we've seen work. We pick our individual actions from the things we've heard. You need to really understand that it's not logic and deduction because we don't know all the facts. So that makes deduction not work. And it's not the boss saying jump because then who knows why that's going to happen. And actually we'll jump in different directions. So if you understand that, then you say, well look, I had better go out and collect some good stories. And there's been a lot of classic studies. Like there's a Bell Labs one, they took the highest performing people, the ones that were innovative, the best engineers, and they asked, well, what's different about these people? And what was different is they had buddies in different areas, different technical areas that they talked to regularly so they could sort of keep up with what's happening and what's coming down the line. And so they had this network of stories that they had gotten from their buds. It was just, you know, have coffee together sort of things. And that wealth of stories gave them greater insight about what's likely to happen. So that's the type of mental model to keep in mind. You are also Asking about bosses. Bosses usually come in and say, here, we're going to do this. Well, you know, the problem is then nobody really understands why we're doing this, or they have different theories about it, or they may disagree with it completely. You need to have a discussion with people, not just to get them to bought in, but to educate them. So a meeting where the boss says we're going to do this is a bad meeting. Then the actual work happens out in the hall afterwards where people say, what is this? Right? And then they have these theories, these stories about what the boss is doing and why he's crazy and it'll never work. And that's not going to be a good team for executing something. I'm sorry, right. You want people to sort of understand the situation and see that this is actually a plausible thing. And that requires actually talking to people.

Oh man. You're doing context engineering on your employees. You give them enough information so they can make the good decisions.

Context engineering, I like that. Huh? I'll use that.

Okay, that's cool. No, I do like the idea that if you want people to do things then you're gonna have have some kind of backwards and forwards situation. So a lot of, I guess work success or maybe like cultural success is going to be about enabling people to discuss stuff rather than just it being one way flows of information. Is that about right?

Yeah, absolutely.

Okay.

An interesting thing is what I see people doing about this. Right? So I was on a panel with the CTOs of several of the leading companies and they were talking about hard to get data for AI and blah, blah, blah. But everybody had built what they called AI buddies for every employee. So you know, all those manuals that you don't read and the newsletters that you know, like they just throw them in the trap. They put all those into a customized LLM for each employee. So now the employee has an LLM that knows all of the what's going on and what's the rhythm and you know, stuff like that. And it makes it easier for people to get up to speed. It makes it easier to find people that you want to talk to and learn from. It just, you know, nobody ever reads a manual. I'm sorry, how do you even produce these things? But if you have something that is active and so that's the difference, the manual is now active. Through the LLM, you can actually use the information that's in there. Right. And that's actually pretty good.

I do like that. I mean there's always so many like notices from HR or like this other team that you only vaguely care about, they've done like a weekly or monthly like update on their statistics or whatever. Like their success is like, well, I only kind of care. I'm probably not going to read like 20 pages of what you've been doing this last month, but maybe I'll ask like one question for the Keyla highlights. So I do like that idea of automating to find stuff out better.

So automating communication. You're helping people find the people to talk to there. And then this deliberation IO thing I mentioned is about how do you get people to talk together better without the screaming and yelling and loud voices and stuff like that.

Not having yelling matches with your colleagues. That seems like a very good idea. I mean, I guess outside of work as well, like having ways to chat in a productive way.

Well, we did an interesting thing. We helped Washington D.C. decide what sort of AI should be in city government. There's a big town hall thing, but we did the online part of it. And do you know what people in Washington D.C. wanted? Now these are not like the brainy people, the powerful people. These are the people with two jobs and three kids. Right? So, so, so what they wanted. Can you want to guess?

I. I have no idea. I don't know. Politicians pick the will of.

They wanted an AI for themselves to help them deal with all the crazy stuff that government does. Right. They wanted somebody that would help them fill out forms and notify them when things are happening and, you know, just keep them on course because it's too hard. It's too crazy difficult. And that sounds like this AI buddy type thing, but it's a little book. It's aimed at government and keeping you square with the government. I think it's brilliant. None of us tech guys thought about that. None of the government guys think about it. It was a person with two jobs and three kids that thought about that. So that's an example of using AI to be able to find things that are actually just great ideas and not hard to do.

Yeah, that's a really wonderful idea. And certainly like, I mean, it's just an endless stream of laws and okay, as an example, like, I guess I actually was like US Immigration quite a lot because I'm British but live in America and the laws keep changing. It's like in between filing things like, oh, something's changed. Different forms need.

Yeah.

Hard to keep up. So I can certainly imagine, like if you're a lawyer trying to keep up with these laws, like very Very difficult, time consuming task.

Yeah, but the average person, I mean, you know, you're educated, smart, you've got some time. Imagine that, you know, you got two jobs and three kids. Right. And now you have to keep up with all this stuff. No way. Right. You need some help and you want it to be personalized help and you want it to be smart. And it turns out we can now build stuff like that pretty easily. Quickly too.

Yeah, that does sound incredibly useful. Okay, so I mean, since you brought up, like, politics here, I mean, I guess one of the big things is you talk about having, like, communicating with people within your own culture. How do you communicate with people who are in a different culture or different community?

The main thing, I think, is to keep in mind that ideas and stories are, are valuable. Right. So learning about things, hearing the other person's story, and they would be valuable. They would value your stories too. So the relationships that work best right across communities are ones that are essentially people trading, like in the sense of commerce, trading stories. I'll tell you about this thing. You tell me what's going on over there. Both of us come over better. We don't have to agree about anything else, but we're helping each other. And if you ask people about folks where they have that sort of trading relationship, you find a great deal of trust and a great deal of value. And so, you know, you need to be a, A, a trader in stories that are relevant to what you're doing. And if you can do that, that's good for you, good for them, good for your community.

Yeah, that seem. I like the idea of, like having a, a trade in stories. You tell each other different things to get different perspectives.

This is what you do.

I'm like, yeah, this is my job.

In a sense, right?

Yeah, yeah, actually, yeah. Do you want to make this concrete? Like, do you have examples of, like, when this can be beneficial to different parties and when you might want to do this.

So for instance, I gave the, at the opening here, I gave this example from Bell Labs where people had developed networks by trading stories of what other people in other departments were doing. They'd go and have coffee once a month or something like that. And it was valuable for both parties, so it was well worth doing it. And that's sort of the basis for a trusted sort of business friendship type of a thing. And the people who did that were uniformly much higher performance, much more innovative, much more sort of skilled in survival than the ones that didn't. And you know, you see that in all sorts We've done lots of experiments. I guess one I like that is salient to people is we set up an experiment where we had day traders and we had a couple hundred thousand day traders. Right. And if you can set it up where they're trading their strategies with each other, not exactly how much they spend, but I did this and it worked out that way. Everybody does lots better. Not just a little better, lots better. And so there you are. You make more money if you trade stories. That's pretty good. And we've even done that with like expert traders. We're talking about people that, you know, trade $10 million a day for their companies. Right. It turns out that if you can get them to trade stories, right, just, you know, make this proprietary data, but they can say, oh yeah, I think that's going to be long, or that's just bs, whatever, that sort of thing that helps them avoid disasters. They're much safer investors. Them have much fewer mistakes because they shared stories with other people.

That's fascinating. I do like the idea of like sharing stories in order to make more money, but. So is this like an averaging effect then because you sort of doing the consensus of what other, other people in the group are doing, or is this a different effect?

It's not actually an averaging effect at all. It's what you would think, but it's not what it is, is, you know, like if you're doing trading, let's say for, you know, trading strategies in money or something like that, it's hard to keep all the different possible strategies in mind. And people forget ones. They look only at a few. And so trading stories helps you remember. Oh yeah, you know, I could do that. That would actually be pretty good, wouldn't it? So it's this sort of remembering the things that aren't in your sort of immediate portfolio. And then the other one for the expert traders, the biggest effect was remember that things screw up. So it was a tail risk. It was, you know. Yeah, but Johnny did that 10 years ago and got wiped out because this thing that you forgot about actually happens sometimes. So it's keeping people contextualized to all the possibilities. So you're not just, you know, in, in your own personal echo chamber that you actually sort of know what could happen and what other people are doing. And then you make use of that. You make your personal decisions based on that.

Okay, yeah, certainly. I mean, with a complex idea like trading like investments, there's so many different possible strategies that, yeah, humans just can't keep all of them in Their brains at once. I like the idea that just exposing yourself to a few other ideas, I guess it's going to get rid of you or eliminate some, Minimize some of your own biases then, so. Yeah, okay.

And remind you that things sometimes really screw up and you don't want to be exposed to that.

Yeah, definitely. I like the idea stories to stop you doing stupid stuff. I, I, oh, we're back to, like, there we are.

Yeah, anti stupidity.

We're back to, like, how do you deal with the lions and the wolves and whatever.

So. Yeah, yeah, same, same thing. Don't go there. Sometimes people get eaten. Yeah, okay.

All right. So I guess the tricky part comes when you've got different people with different opinions on things and you need to make a decision. So how do you go about coming to a consensus when people disagree?

We did some big, everyone's very concerned about polarization. You know, the people on the other side, they're crazy. They, you know, eat babies, whatever. So some horrible way. And so we did a couple of big experiments. These are like 30,000 people statistically balanced for all sorts of things. And it turns out that most of this sort of like people being polarized and disagreeing with each other is that they're just not familiar with what the other side actually thinks. There's so many loud voices, the people who may in fact be genuinely crazy, you know, that dominate and just saturate all of the airwaves waves and the social feeds and stuff that you think about the other side as these loud voices. And you miss the fact that the vast majority of people don't believe that even on the other side. And so we did something where we had, and this is the deliberation IO thing again. People would put in comments, and then we'd just graph them along a particular dimension. And you could say, yeah, at the edges there are crazy people, but almost everybody on almost every issue clusters in the middle. So those people on the other side, they're actually pretty much like you. And sometimes it's surprising, sometimes you think they're liberal, but they're actually more conservative about some things and vice versa. And what that did is it changed people's opinions because they stopped paying so much attention to the loud voices, and they started paying attention to the fact that most people actually are pretty reasonable. And so I can be reasonable, too. And we saw something like a 30% reduction in these polarization measures, and it stuck for a couple weeks. People get caught up in the loud voices again. But it's amazing how just sort of having a picture of what's going on in the world helps you make much better decisions. And so that's what we're building. Right. And that's what you need to do. For instance, in companies too is typical meeting you get in there, you don't know what people think they're not about to say because they might be outliers. So all you hear is the loudmouths, including the boss. And you don't really know what most people are thinking. And if you did, you might realize that there's another path that everybody could get behind or that it's really not as weird as you think it is and so forth. So you need mechanisms and you can have these digital mechanisms that help people do that. So what we do is we have people putting comments for things. This is on your mobile phone. What do you think of this? Bang. Right. And then what the AI does is do summarization. It doesn't add anything of its own. No content from AIs. What it does do is it reflects, it says, oh, you know, most people think this and a few people think that. That's all it says. And now you say, oh really? I didn't know people thought that. And who are these guys who do this? Why do they think that? And so you might say, why do you think that? Right. So it gets to be a much more productive thing because it's a genuine communication stories among each other. And what the AI is helping you do is make sure everybody's story is heard and it gets out there in a non contentious way. And it works pretty well. We call it Socratic dialogue. You're just asking questions, that's all. I love that.

Just making sure that if you understand what other people are thinking, then you realize that most people are actually reasonable. I guess if you are hearing the loud voice a lot, it's like probably time to step away from social media just for a while and speak to regular people.

Well, it's dominated all the media today. The newspapers, the video, everything is dominated by people who are incendiary. You say crazy things and then you make more money because more people listen to you. And that feedback loop just destroys the ability to make decisions.

Absolutely. So yeah, we're trying to say too many incendiary things on the show.

Oh no, let's go for it. We'll get the best ratings ever, right?

Absolutely. Yeah, we'll go viral. So you mentioned meetings. This seemed like everyone spends so much time in meetings. Do you have any tips for like how you can make sure everyone Gets heard in meetings and all the appropriate stories are shared.

Well, that's one of the things with this, like, little digital app. Right? This deliberation. IOS is you don't have to have meetings. This is literally the thing. Your phone rings, here's the question. Yes, no, six. Whatever the answer is, that's it. And then after a couple minutes, you get to see a little graph about what everybody answered. There are other things like this, but the point is that for surveys, but this is iterative, so you say, oh, people believe that. Then you put another little comment in. And what it is is this discussion that is curated by the AI, doesn't add content. It just keeps you focused on the problem ahead of you. Doesn't allow you to curse and swear and do stuff. Right. But just. Just a little bit. Not a lot. But what it's doing is it's trying to keep people on the whole and not get crazy about particular comments and works.

Okay. I mean, that's the holy grail. There is, like, not having meetings, I think, like, for every company, like. Like no one's vomit. So. Yeah, yeah, I love it. Like, poll. Poll your colleagues, like, and, you know, summarize what's going on.

Yeah. Most meetings, you know, like, people always say, God, we could have done that in five minutes. Well, you could have done it in less. And that's the point is, is that, you know, if you have the right sort of tools. It's not like Slack. Right. Because Slack is. This is more like social media in the sense of being a continuum. This is commentary, it's deliberation. It's iterative. It has, you know, ongoing summaries, and that makes it all the difference.

I like that having some commentary as well. Just like someone gave this answer, like, you get a bit of the why there as well.

All right. And you can include all of the opinions, not just allowed one.

Nice. Yes. Okay. So when we teach data, storytelling data, one of the big things is, like, you've got to personalize stories for the audience, so it's appropriate. But the idea of, like, having a shared story, that seems like it's more of, like, a central thing. So, like, when do you want, like, a standard story versus a personalized story?

Well, I think one of the big things people forget, it's not either everybody or you. There's this thing in the middle called your community. So you have a work community, you have a family community, you have the guys you play baseball with, whatever. Right. All of those communities need to work together and do things together. And so they need to have a shared culture. I call it shared wisdom, because hopefully it's a good culture and you're trying to make it wise so that you make good decisions. You know, it's not like there's one truth. This is. This is one of the things that screws people up all the time. In fact, there's very little that you can really call truth. There's what we think is going on, and sometimes we're pretty good at it, but often, not particularly, as time goes on, we realize, oh, that really wasn't what was happening. It was this other thing. And so each community needs to have its sort of worldview. It's culture. And you make decisions that are about the community, using the culture. So there are things you would do in a town hall that you wouldn't do around the family dinner and you wouldn't do in the baseball game. There are different sets of behaviors, there are different sets of norms. And so you need to have story sharing within each community to help establish the norms, the wisdom, as it were, and therefore the actions that the community can take by sharing the rights, sharing stories, and having people sort of discuss it and react to it. I'm sorry, that's a long answer. But it's not just everything's uniform or it's all personalized. Actually, you are the compound of a lot of different communities, and you need to have things that are relevant to the community context.

Absolutely. Oh, we're back to context again. But yeah, I definitely agree that, like, the way I chat on a podcast is different from the way I chat to my friends and all my family. Like, there's lots of different ways of communicating in different circumstances. So it seems that the key to this is having different stories for different communities. So can you talk with it? How is technology going to make this easier?

Is a challenge for technology because the economics of it, the way people have built things, is either it's the same for everybody or it's just for you. But actually you want context. And so first, it's like the AI buddy thing I talked about within companies. That's like the company menus, the company newsletters, the company's manuals, et cetera. So it gives you context to the community that you're working in at the moment. Well, you'd like the same sort of thing for family, where you have the trips you've taken and you know things about the kids, report cards or whatever. And so when it gives an answer, it knows something to give you a contextually appropriate answer. An interesting Thing is, is that if you look at, say, for instance, Facebook groups, they mirror the physical places that people meet up, right? So people produce these community spaces without being encouraged. Right? And they do that so everybody in the baseball club can know when the next game is or the next show that people want to see, or everybody in the family can be reminded about the upcoming dentist appointment, whatever it is. Right? So we already do that. But now what we want is we want AIs that know about the group so that when you ask it something or it reminds you about something, it's appropriate to the group. And you want the AI to be something where you own it, not somebody else. So that data remains private. It's under the control of the group or the individual. And that's what we're building with this loyalagents.org thing with consumer Reports. Consumer Reports is a community that wants to not get ripped off by people selling junk, Right? And so they share experiences, stories about, I bought this blunder and it broke the next day. Right? So when you do things with Consumer Reports, it says, you know, 96% of people had a good experience with this blender, but 4% would like, set fire to the company and, you know, that sort of thing. And that's really what you want. You don't want it to be influenced by advertising or things like that. Actually, can I. Can I. Let me just do something that's really surprising to people, but you should think about it. So one of the things that's likely to happen with all this agentic AI, so people will have, you know, AI agents. Everybody out in Silicon Valley is busy doing this now. Most of them, the company owns the data, not you. Let's just leave that aside for the moment. But one of the things it does is it gets rid of advertising. Because if you say, I want to buy a blender, the agent goes out and finds blenders that are good for you. It's not paying for clicks and stuff like that anymore. It's something where a lot of this stuff just happens automatically. So, like versus ChatGPT can now buy things for you. Right? You can tell it, oh, look, you know, I need milk again. And it'll find the best place to get milk, right? You don't have to spend your time on it. People are going to love this. In China, they have something called WeChat that's a little like that already. And everybody is addicted to it. It's one of the biggest, most profitable businesses in China. And so that's what we're seeing. It Gets rid of the advertising community because it's the agents that are finding things, not people like doom scrolling and stuff like that. So that's really interesting. And then the purchasing decisions are either automatic or the AI recommends something that's small. But if it's a big thing, they have to be much more like an advisor and they have to be trusted. And there needs to be somebody who, who rates how honest all these things are and how good they are. And that's what we're hoping to do with Consumer Reports is actually have a rating system where if somebody says, oh yes, this is a really good car, you say, well yeah, accept that. And that's what Consumer Reports has always done. So we figured that we're a good partner and we'll be trying it out on a couple million people near you very soon. So. Loyalagents.org There we are, there's the pitch.

So yeah, I do like the idea of agents being able to like perform these tax like shopping tasks. It does get interesting. So yeah, trusting an AI agent with your credit card, it's like, okay, maybe you like to buy milk, not to buy a car. Like it's not having my. I'm definitely giving it a budget limit.

Or drugs or from wherever or you know that Prince in Nigeria really doesn't need the loan of a thousand dollars. Yeah, okay. You can see where it's going.

Yeah, yeah, there's definitely some problems. So actually you mentioned that that would mean advertising like to humans is going away. That's going to break the business models for like a lot of websites, virtually the entire net.

Yeah, absolutely.

So are we going to see adverts targeted to agents then like something that's going to like attract the agent to go and make the purchase there for you. Do you think that's a possibility?

You're going to see things that may or may not be human readable, but what they do is they have standard statistics about things like the price and sort of stuff, the specifications. But they also try to highlight the advantages of this thing, which is a little like advertising. You see it in like if you go on Amazon they'll have things about like what's interesting about this one versus all the other ones. And then they'll have the specifications so the agent will look at that. But then what you want is you want something like Consumer Reports that says this one's bs, don't pay attention. Right. And for small things it'll probably just buy things automagically and for bigger things it will probably come back and you'd have this sort of advisory from the agent, and then you sort of look at it yourself and you can say things like, well, what are the advantages of this? And what are the. But these things just feed off of what's on the web. So you need to actually have human input from people who have bought this thing, from people who have worked with a company before. So it's a little like Yelp, a little like Consumer Reports, a little like a couple of other types of things. You need to have that input to make wise decisions. That's a shared wisdom, incidentally.

Right, Absolutely. I like that. It just seemed like if you're working in retail, Ben, you got to think very carefully about what happens once shopping becomes more agentic. How are you going to attract the agents to your site? I guess. Yeah, the. All the. All the information on your site needs to be machine readable, for sure.

Yeah. Instead of just pumping out ads, now you actually have to have things that look like advisors. It's like if you go in and you buy, if you invest in something, you go to Fidelity or something, you have a lot of money, they'll assign somebody to talk with you about it. So that's more the sort of character the companies are going to have to have is really as advisors. And they have to be honest at a certain level, otherwise they're going to get a bad rating. So that's really key is to have that rating of honesty that's accessible to everybody.

Okay. Yeah. So I guess, yeah. Telling stories about, like, what's good and what's bad, then.

Yeah, yeah, exactly. Exactly.

All right, super. So one more thing I'd like to talk from your book. So you made this analogy between shared wisdom and Daniel Kahneman's ideas around thinking fast and thinking slow. I thought it was very interesting. So do you want to explain how it works?

Yeah. So Kahneman, Nobel Prize winner, has this idea that human thinking has two pieces. There's a system one, which is very fast and automatic and. And things happen and you're not thinking it through. And system two is a more deliberative, complementative sort of thing that you're reasoning about stuff. Well, what. LLMs are really much more like system one. It's just like all the things they've heard and then if you push their button, they regurgitate. Okay. They're not really thinking about it, even if you ask them, why did you do this? Or what's the chain of thought? Or. No, it's just they're making up a story based on all the things, they're heard. Right. But what you can do is you can embed them in a loop where you ask them, well, why did you say that? And where's the evidence? And give me the evidence. And you can approach it where there's bunches of LLMs all trying to give answers, but with slightly different context and say, well, why do I want to believe you versus them? So, so those are the sorts of things that people do to winnow out the truth. And yes, you can build those. And that's one of the hot topics in Silicon Valley today, is how do you take this sort of system one LLM, you push it, it says things, it's not bad, but it's just like the common conversation. It's not good either. Right. How do you turn that into something that's contemplative and has some, some real wisdom in it? And it seems that the strategy that's most popular is we're going to have several little answers from the system one thing, and we're going to bang them together to see which ones we really believe. Oh, you mean like people do? Yeah, like that. And people could be part of the process too, of course. Right.

That's interesting. So, yeah, it sounds like getting to this system, to thinking the stuff where you are, I guess, reasoning about things. This is what happens with the deep reasoning tools, then quite a few of them around. So from OpenAI, they're beginning to be that way.

Yeah.

Yeah. Okay.

All right.

So, yeah. Do you have any more advice then on, like, how you go about getting better thinking from AI?

Well, I think that there's a number of things that are just sort of fundamentally wrong in there. Right. One is that you can get better thinking, but it's still not contextualized. Right. It doesn't know about your community, your goals, your values. It can learn a little bit about that stuff. But there needs to be this notion of conversation among people in the community to establish what are the community values, and that's the context you need. One can imagine systems that do this, but you'd like to keep that sort of something that you own and not OpenAI just to pick a company. And so that'll be a battle in the next coming few years is like, who owns all this stuff? Who owns the data? I think that there's other things which are, for instance, AIs are not physical. They don't have the history of being a kid, of going to high school. They know what people say about it, but that's really different than what actually it felt like. And so people have this context of their feeling and their experience and their sensory apparatus that LLMs don't have at the moment and are unlikely to have for quite a while. I mean, you can imagine eventually robots will be that way and so forth. And moreover, you know, we don't really want a AI making decisions for us, in part because it's not grounded in the human context. It's going to say things that are just sort of wrong for us, maybe subtly. That's the danger part of it. So we'd like to have things where people are still very central to the decision loop. And that's why we're building things like this deliberation. IO yes, there's AI in there, but the AI is helping the humans be contextual. Right? It's not helping. It's not giving them facts or telling them what to think. Same thing with loyal agents. It's not telling people what to think. It's giving them much more sort of an even view of what's out there so that they can learn and express their preferences. So it's not taking the human out of the loop. It's empowering the human to make more contextualized choices and be more in control of their life. And I think that's something that's critical to do now because we're at a sort of inflection point where it could go either way. And we need to push on the let's make things for us, not things for getting rid of us branch of that tree just so that when it grows, it grows in a way that we like it. And when I say grow, I mean when the technology grows, it grows in a way that actually complements us, not does things that we find oppressive.

Absolutely. So I love the idea of, like, using AI in order to make data and like, other people's stories more accessible. So you helping them make better decisions, and that's certainly obviously a much more preferable situation to using AI for eliminating us or oppressing us. Wonderful. All right, so before we finish, I mean, we talked a lot about shared wisdom. You've also got, like, decades of, like, very cool stuff you've been doing. I'd love to touch on a few of those things. So you ran the MIT Media Lab for decades. Can you talk us through, like, some of your favorite projects that came out of that?

Well, I was one of the first people at the Media Lab, and I was academic head for all, but I didn't run it for decades. Okay. No, no, no. I just. People get touchy about that. Right, context again. Right. And some of the projects that I think are most interesting. One of the projects we did is we built a very simple type of AI. It's a rule based AI, but it runs on mobile phones. And what it does is it gives health information, health advice to people in very poor areas. So it's a little company called Dumague and it helps 400 million people in rural areas stay healthy. It handles one out of every 30 births in the world. Right. Helping the mother and the baby do well. That's amazing. It's very simple. It's a way to personalize and contextualize the medical advice that comes out of the capital city. Okay, so same type of theme, but it's really, really sort of spread. Another one that I'm proud of there is I was involved in helping data be used for better governance, to make policy that worked. And when the Sustainable Development Goals were set up, I was one of the UN Secretary General's data revolutionaries who were supposed to help the Sustainable Development Goals be realizable because they were concrete. They could have data, you could assess. How are people doing? And that was the idea. It still is the idea. And the UN is actually now attempting to take the same ideas here and making AIs that are helpful for local people to be able to assess things and make it a much more sort of concrete type of thing. So I'm proud of that. That was a pretty good thing. And then the move here to Stanford is interesting because I can do things like deliberative democracy. Right. That's pretty good. How do you get people to talk to each other better or the loyal agents? How do you make AI serve you and not somebody else? That's a question that industry is not so concerned about, except from a legal liability point of view. So sir, we're doing things, we're trying to make the world a little better, one piece at a time.

Wonderful. I mean, I do love the use cases there. Talking about like healthcare, talking about sustainable development, talking about improving democracy. So these are things that have real social impact as well. It's not just like messing about with data and AI just for the sake of it. It's like got real impacts there. All right, so just finally I'm always interested in like new people to follow new new ideas.

Who are you?

Whose research or work are you most interested in at the moment?

Well, there's a guy by the name of Michael Bernstein here at Stanford as a social AI group and he's the one that's done, you Know what happens if you have lots of LLMs that think that they're villagers and they live in a village and they talk to each other. What happens to the social interactions? Or most recently as a book called Flash Teams, which is can you use AI to help put together better teams? And if you do that, what does that do to the corporate structure? And I have to say that that's a really interesting question is if you can actually put together teams on the fly. So this is the sort of social context of make them fit together, then a lot of the reasons to have corporations go away. So that means we get back maybe to guilds the way things were in the 1500s where you're this type of, you're a full stack engineer and they call you up to do something for some other project. But it's anybody in the world could call you up as opposed to you're supposed to show up in this building five days a week or three days a week, week, whatever it is, it just changes the structure of society dramatically. We ought to think about this before it gets off there and starts happening.

Yeah, I mean that's going to be a huge change to like the idea of work. I mean, I suppose you already see quite a few people who are like fractional or whatever and they'll work for several companies at once.

Fractional CEOs, fractional CFOs. You get things like upwork and fiverr and you know, yeah, they come in and they do a project and yeah, you can just do these things dynamically now. And so things that used to take you hiring a bunch of people and a bunch of capital and set it all up. No, now you can just do it and it takes a quarter of the time and quarter of the cost.

Okay, so yeah, the rise of the consultant or external employee then I guess that sounds fascinating.

What's the death of the consultants? Because the consultants don't have a good role in this. It's going to be a little bit more like a gig economy where people have particular credentials and they plug into different projects but they're not part of a like lifetime employment contract. So that's a change. We should think about that.

Yeah, definitely something to think about. Certainly. Yeah. Want to ponder on that. Wonderful. Thank you so much for your time, Sandy.

My pleasure. Good talking. Take care.