AI Should Make Us Better - Dr. Vivienne Ming

Speaker A: ACAST powers the world's best podcasts. Here's a show that we recommend.

Speaker B: A photographer in Texas earns an extra $47,000 a year shooting Star wars cosplay portraits. A teacher in Maryland turns her woodworking hobby into a five figure side income without leaving the classroom. And a couple in Pennsylvania will rent you backyard chickens for the season so you can try the egg laying life without commitment. My name is Chris Guillabo. I'm the host of side Hustle School. I share these kinds of stories every single day in detail with full transparency about the numbers. The point isn't just to inspire you, it's to show you what's possible. Proof that ordinary people are quietly building extra income in surprising ways, including a few ideas you can borrow less than 10 minutes a day, every day. Subscribe or follow side Hustle School wherever you get your podcasts or find us directly @sidehustleschool.com.

Speaker A: acast helps creators launch, grow and monetize their podcasts everywhere. Acast.com.

Speaker C: Hello and welcome to Don't Stop Us Now AI Edition. I'm Claire Hatton.

Speaker D: And I'm Greta Thomas. AI has been described as the most transformative technology since the harnessing of electricity, and we are here to keep you in the loop for what you need to know and do in order to stay relevant in this fast changing world.

Speaker C: Each episode we bring you leading experts and together we explore how you can stay ahead of the curve and in the know for what skills will be valued in the future, how jobs may change, and how industries will evolve.

Speaker D: So why not subscribe to stay in the loop and without further ado, hit. Enjoy this week's episode. Hello and welcome to this week's episode of Don't Stop Us Now AI Edition. If you've ever had a nagging feeling that the way you and most people are using AI right now might actually be doing some harm, I guess this week has the research to back that nagging feeling up.

Speaker C: Dr. Vivian Ming is a theoretical neuroscientist, entrepreneur and author who calls herself a professional mad scientist. And she means it. She co founded socos Labs, a think tank dedicated to the future of human potential, and has spent nearly 30 years working with AI long before it became a household word. She's founded six startups, served as chief scientist at two others, and her latest book, Robot When Machines have All the Answers, Build Better People, draws on all of that experience to make a provocative argument. The vast majority of us are using AI the wrong way.

Speaker D: Yeah, and I think she's pretty Right. The stakes, as Vivian lays out in our conversation, are, uh, higher than you might think. She's talking about cognitive decline, risk of dementia, dementia itself, and about the difference between AI that makes us sharper versus AI that can hollow us out.

Speaker C: In this episode, you'll hear why having an army of robot minions is not the goal and what to aim for instead. The skills that predict who thrives with AI Why the best decisions are the ones neither human nor AI could reach alone. Vivian's nemesis, prompt. And why it makes her work sharper, and what Vivian would say directly to the founders of the big AI companies.

Speaker D: And that's pretty interesting. Claire and I were pretty blown away by our conversation with Vivian, and we think you will be too. The span and particularly the purpose of her work and her expertise in AI made us think about and frame AI very differently. As ah, Vivian says, it's not about the fear of AI replacing us. It's about something more fundamental. AI Defining us and whether we'll make that a positive or a negative thing. So, without further ado, enjoy this thought provoking conversation with the unique, funny and whip smart Dr. Vivian Ming.

Speaker C: Vivian Ming, welcome to Don't Stop Us Now AI Edition.

Speaker A: It's a, uh, pleasure to be here. I'm looking forward to the conversation.

Speaker C: Yeah, we are too. Now, a question we always like to ask our, uh, guests is imagine you're at a dinner party. You've sat next to somebody that you've never met before, and they say, vivienne, what do you do? How would you generally answer that?

Speaker A: I mean, the first thing I'd have to do is ask myself, how have I ended up at a dinner party? What sort of kidnapping scenario has led me? Uh, autistic badgers look at me with pity, uh, for my lack of social instincts. But I spend a lot of time on stages. It does so happen that people put me in sort of work dinner party environments. Although that also means people often know who I am ahead of time. So what the hell am I going to say? I'm a professional mad scientist. People bring me fascinating problems. Uh, Dr. Ming, my daughter has 500 seizures a day. Please save her life. Uh, my son can't enter REM sleep. It's like an episode of House or Our Country. We did everything the World bank says our education and our international PISA test scores go up every year. No one's hiring our citizens. Please help. I have the best job in the whole world. If I think my team and I have any unique insights to offer, I pay for everything and whatever we invent, we give away. So I call myself a professional mad scientist because if we knew it would work, it wouldn't be science. If anyone else was doing it, it wouldn't be mad. And I do the professional bit just because otherwise I'd feel like I was just playing.

Speaker C: But it sounds like you actually feel like you are in many ways, because it sounds like it's really fun.

Speaker A: It's a very expensive job to sort, uh, of grant oneself. But I've had the good fortune that at least the second half of my life has been pretty wonderful and this is how I want to spend it.

Speaker D: Do you fund all of the incredible work with the speaking and the books? Um, and we'll come to your latest books shortly.

Speaker A: So all of my projects start as philanthropic projects. They exist simply because I thought this is a thing that should happen. Sometimes those produce a science paper every so often, uh, like during lockdown. We worked on questions of depression and suicidality. And it led to a company, uh, that has developed an epigenetic test for postpartum depression. So things come out of that, as my co founders of my companies would attest. I never found a business plan I couldn't make worse with a little heart. But I've had the dumb luck of occasionally being successful enough that people have taken note of some of this work. But yeah, I get up on a stage and I give a talk for the truly unconscionable sums of money people will pay me to do that. That often will fund an entire project.

Speaker C: Wow. Well, that's, I mean, it's great to hear that you can do that. And, you know, the reason that they pay you huge amounts of money is because you've been, you've spent a lifetime doing some incredible research, being on the frontier of technology. And I, uh, know in your latest book, Robot Proof, when machines have all the answers build better people, you have done a heap of research into the human brain and how it interacts with AI. How would you briefly summarize what most typical gen AI users are doing today and the consequence it has on their brains?

Speaker A: So here's the most important thing I can say just to set the field. All of my answers will be messy and complicated, because human beings are messy and complicated. I sometimes describe myself as an AI realist. I built my first model nearly 30 years ago. So this is long before agentic models or the Nobel Prizes recently. And then it was real. It was a scientific curiosity. In fact, my field of theoretical neuroscience, we collectively earned one of those Nobel Prizes in John Hopfield's award and, uh, his co awardee was Geoff Hinton. We were all building models to understand people. Can we. If we took a whole bunch of images and we built a big enough neural network, would it tell us something about how the brain sees images? And it does. Or my work was in hearing. If we got enough sound recordings of people having long conversations in every language, would it tell us something about the human auditory system? And it does. Really cool. Fascinating, but very niche scientists. And then Jeff took some of those models and stacked them on top of each other. And pretty soon it was able to tell dog breeds apart in pictures better than humans could. Now we're in a whole new world, particularly with large language models. And then, you know, agentic AI is an extension of that modeling concept. So for the nerds that are listening, I use the term machine learning to talk about the underlying computational and scientific modeling. AI to me is a product. I think it's more than reasonable. AI models are intelligent. I know a lot of people bitch and complain that they don't like that. We call it artificial intelligence. Have a conversation with Gemini is probably the worst conversationalist. But even have a conversation with a small open source model and tell me that it's not intelligent. I'm not saying it understands what it's saying, but it can have a conversation. Some colleagues of mine at UC San Diego ran the classic Turing Test, and I think they really genuinely did it. Not one of these thought leader statements. They ran a robust example of the Turing Test. People chatted with a real human and an AI. They had to figure out which was which. And GPT passed the Turing test. I wrote a blog post saying actually humans failed the Turing Test because a true Turing Test pass would be 50 50. You can't tell the difference. Humans thought GPT was the human 75% of the time. So clearly we knew one from the other, but we got confused of which is which. When you look at these and you're trying to understand what's happening to our brains, again, we are messy. Different people are producing different results. We can go into some detail later. Right now I just want to say what happens to your brain is what you put into it. If I'm giving you sort of a short answer, if you engage with agentic models shallowly, give me the answer. Write this marketing post for me, right? Do this spreadsheet analysis. Then. What I see in my research is this measure of cognitive engagement that is very robust, called the gamma band. Activity goes way down. Like the difference between working on a Hard math problem and just watching tv. Enough of a drop that it actually worries me about things like early cognitive decline, dementia rates, Alzheimer's rates. I mean, shocking. So if you are shallowly engaging with AI, to quote the cartoon Futurama a robot, uh, experience this moment of existential angst for me, then I virtually guarantee you that if you're not harming your brain, you're certainly not helping it. This isn't just my research. For example, doctors doing colonoscopies, AI assisted colonoscopies. The more AI assistants they rely on, the worse they're getting at the colonoscopies over time when they don't use that assistant. And Anthropic's own research on clog code shows a similar messy story, which is for a large percentage of clog code users, but not all of them. Their conceptual understanding of coding gets worse over time because it's doing everything. So if you are really engaging shallowly, you don't fully understand how to do your job. You're using it to save, uh, time and cut corners. You're essentially using it. The way Sam Altman is pitching it, it will do all the boring work. You get to do the fun stuff. I say, on balance, you're harming yourself. Now that's not everybody by a long shot. And in my experiment and in some really notable research by a group at mit, we see what we call cyborgs. Yes, I'm a sci fi nerd, but in this case it was the Harvard guys that came up with this term. But uh, in my research, the cyborgs don't let the AIs do the boring part. They do the boring stuff with the AI. They do the creative, high level stuff with the AI, the hypothesis generation, the creative ideation, the low level data analysis. They are all in at all levels. You know, maybe the model's running some of the underlying statistical analysis, but they're saying, run this one now, run that one now. Show me this distribution. So now they understand what the data's saying, not what the AI is telling them. They're saying, all right, I've got this hypothesis. Gemini, Claude, tell me why I'm wrong. Claude or Gemini gives them an answer and they say, okay, I don't think we're on the same page. Like their spidey sense start tingling, they push back and say, let's make certain you're answering the question I thought you were answering because I don't think I'm hearing what I was expecting. In other words, I domain expertise matters. So this idea that AI will do all of this stuff for you. It can in theory. I mean what these models benchmark now is astonishing. But I will tell you from my research and kind of just my framing here, I want to build cyborgs. I mean, you get it from the title of my book, right? It isn't that I don't believe in what AI can do, but I believe more in what we can do together. And I think the mistake is this kind of human in the loop. The machine does all the work and then you kind of approve it. And what we see, and many have seen, is that's actually not good for humans, nor does it produce the best results that AIs are capable of.

Speaker D: Yeah. And so if I was to paraphrase what you've kind of just been talking about, the superficial, potentially harmful way of engaging AI is to delegate in a kind of almost like a one shot. The way to get the most out of AI, uh, the way to get a better end result potentially and to become a cyborg is to use it as a thought partner, bounce the ball multiple times, have multiple, multiple interactions on different facets of the plane that you of the problem that you're working on and use your own domain expertise to challenge and to direct that conversation. Is that a kind of a fair sort of um, summary?

Speaker A: I think that's really good. Instead of thinking of agents as an army of robotic minions that will carry out your every whim, which I think honestly you look at the valuations that anthropic and ah, OpenAI are expecting. Hard to deny that part of that pricing is the assumption that these robots do everything. But I'm not saying this just to be precious about humanity. The best results in my research come out of the cyborgs, not out of the autonomous AIs and frequently obviously not out of humans. It's this collaborative thing. So I'm going to push though even a little further than what you're going for, which is, yes, think of it as a collaboration. It's very natural. This is a separate entity from you. But I'm going to lean even harder into the cyborg. Clearly the white cane people with, you know, visual impairments use is literally not a part of them. But very quickly as they're using it, they go from feeling the cane in their hand to feeling the things the tip of the cane is touching. Their brain begins representing that that's uh, a really natural thing for our brains to do. You use AI for a little bit, you will quickly, if you're not careful, begin to internalize its capabilities as yours, causing it you to be really poor at, uh, judging your own capabilities when you don't have AI, in fact even get this just with Google searches. So I'm saying this as a warning, but also it's something to lean into. How is you making AI better? How is AI making you better? How are you challenging each other to become better? And ultimately the AI as a separate thing kind of vanishes and you're really just leveraging the way it expands. Like our working memory is this classic. One of the most famous papers in all of psychology is the magic number seven plus or minus two. You know, it's this not entirely accurate, but this was the 1950s idea that we can remember about seven things at any given time. It is with all things much messier story. But the spirit of it is true. There's a very limited number of things that we can attend to that we can memorize that at any given time that we can really compute on. AI doesn't have those kinds of limitations. Its cognition is different than ours. So it isn't like it has the same kind of working memory that we do. But the parallels are there. It can keep track of masses of data in ways that we simply cannot. I'm really sort of looking for the best of what it can do, and the best of what we can do turn out to be complementary to each other. And the kinds of errors it is prone to make and the kinds we are prone to make turn out to also be complementary to each other. So there's a degree to which I want the ultimate expression of agentic AI is the agents just kind of disappear. And we are just thinking and it's expanding what we can attend to, what we can keep in mind at any given time. I'm not saying this is coming tomorrow. It's just, you know, as someone that actually works in neurotechnologies, this is the sort of dream vision I have is the end state for us. We're thinking about all of this and the AI is never doing for us what we could have been doing for ourselves.

Speaker D: So I'm intrigued about the, um, the vision impaired person with the white cane in your ideal scenario of how things play out, just as that person with the cane, uh, their brain is kind of adapted to treat the end of the cane that is finding their way as almost part of them. You're also thinking that that's how we will kind of regard AI when we're working to our strengths and working with it correctly.

Speaker A: And it's Evolving in my book, uh, where I kind of jokingly title it Artificial General Luck, where these systems are modeling us, but instead of giving us what we need, what we want at every moment, they're giving us what we need in that moment. And the best version of that, the model itself disappears. We're not even really thinking about it, but it's just right there. When we reach out for that piece of information, that knowledge, uh, that idea, these things are there. There's big risks with what I'm dreaming of here, but nonetheless I'm really thinking about that because right now in my recent, uh, experiment that I wrote about in the Wall Street Journal, the vast majority of people are in this worrisome group that we call automators, where they're simply using the AI to produce everything.

Speaker D: Ah.

Speaker A: And a lot of it's really worrying and a not unreasonable estimate is that the number of cyborgs sort of naturally occurring out there right now is probably low single digits of our labor force. And so that idea of people interacting with these systems is most people are asking for the easy answer. And part of what I'm kind of getting at here is I don't want to be pedantic, uh, or prescriptive for people, but these models, I think need to be benchmarked against the impact they have on us rather than their ability to, to autonomously answer questions. That is important. They need to be able to do that. But at the same time we should be benchmarking. So I'm working on a hybrid intelligence benchmark. So you know, which of these actually increases the creativity and innovation of a hybrid team the most, rather than just the autonomous capabilities. But I'm, um, also thankful these benchmarks should say things like which of these models has the lowest all cause mortality of any user base? Which of these models has the highest income and wage accumulation or the largest friendship networks? All of these are measurable positive life outcomes. When I study people, I'm looking at these broad life outcomes, even when I'm looking at very proximal things in people's lives. And I think we need that same kind of wild eyed mad scientist approach to AI. And I'm intentionally being a little provocative in some of my metaphors here because I want us to think differently about it than, yes, it's a bunch of robotic minions. It'll do whatever you tell it and it'll do it better than you can, so you shouldn't bother.

Speaker C: I think it's such a optimistic way of actually thinking about what AI could do for us. But it is also feels to me that it's just so far away from what we're hearing out of all the

Speaker D: labs and the current, well and the

Speaker C: current trajectory and the, and the trajectory is no end. I mean it's just, you know, uh, I sit on a couple of boards and you know that automation pieces is like Ka Ching. You know that's, that's, it's a very easy sell.

Speaker A: Understand why that is what people are selling. My very first startup was an ed tech startup. In 2008 I did a startup using AI in education. There's no Coursera, there's just Khan Academy is just a guy on YouTube doing these free videos. And I will say I met him once during that era and he tried to make a startup. No one would fund it. That's why they're ah, a nonprofit because no one could see the value add. We got the same thing. We'd go in, I'd do my demo obviously long before long large language models but we're still doing a lot of natural language analysis and it was really exciting stuff. And they would respond, wow, you can read students minds. Because that was as much as they understood AI back then. Bluntly, that's as much as most VCs understand it today. Unfortunately, I think that's as much as a lot of computer scientists understand it today. Here's $2 million if you do financial fraud detection. So to get funded we had to throw out our whole vision of working in education. Uh, my wife is an education researcher and I had to essentially just give them the company. So that was the experience of founding an AI driven startup in 2008. Wow. And let's be clear, that startup was a failure. Nobody wanted to use our product, nobody wanted to fund what we were working on because AI just, it felt like magic. Do it somewhere where we know we could all get rich, sure, but what you guys want to do, no, the only thing that I can say changed it is it was a uh, huge and amazing success in launching my career. Bill Gates wanted to know what I had to say about education because he heard about what I was working on. And Inc Magazine said I was one of 10 women M to watch in tech because of that company that no one wanted to fund and no one would use our product. So I know it sounds dreamy to think how do I use this in a way to build better people and change everything that everyone's doing? And boy did I learn the hard way. Not everyone wants to use that product.

Speaker C: Uh, what I'm really fascinated about Is I think most people and most companies are down the, Going down the path of the automator. How do you think practically individuals. Let's start, let's take individuals to start with. What do you think they could do personally to move themselves more towards this cyborg? I know there's one other group that you talk about as well as, um, the validator. But how could they move to the cyborg?

Speaker A: You know, here's a couple of things that really has resonated both in my own research, in my own practice. Let's start there from a parenting perspective. But I think this also tells you something about running organizations in general. This great study, I think it needs to get replicated. But the initial findings are really provocative. They ran a controlled experiment in which they taught teachers not to praise answers, right answers, but to praise good questions. Within just a few weeks, the kids in the treated classes were asking more questions. They were more engaged then. They scored higher on standardized metrics of curiosity. And when new, uh, science curriculum material was introduced without any additional prompting, they engaged more deeply with it than the students in the control condition. We sometimes think, especially a lot of leaders in techie industries kind of think you got it or you don't. But the research is pretty clear that things that really matter, that resonate in my research that I write about in the book, like curiosity, are genuinely changeable qualities here. I'm talking about a study with children. But I guarantee you, praising smart questions in the workplace will get you very similar effects. Why is this important? Because I don't want to go so deep into this experiment we ran, but we had people make predictions about the future, either with or without AI support. Without the AIs tested alone were so much better than the human beings at this that it really makes you question, why would you have humans involved at all? And then when they had an AI to work with. This is where I get my finding. The vast majority of even very bright UC Berkeley students were just automators. They just did whatever the AI told them to. Some were. This other group you just referred to, validators, sufficient to say they made things worse. They at least they thought about the questions, but they used the AI to simply validate the things they already believed. But the cyborgs, their predictions were a couple of really fascinating things. They were better than the best humans and better than the best AIs. So even humans collaborating with an open source llama model did better than state of the art at the time, which was Gemini 3, which had just come out. And I used in the experiment. So naive kids, uh, having to make 10 predictions in one hour were doing not just better than the state of the art model. At that time. I took the questions off of Polymarket and they were doing nearly as well as the professionals with millions of dollars on the line at polymarket. So super exciting when you look at it through that framing. But when we analyze what predicted cyborgs, so I'm trying to coin a term here, hybrid intelligence. The best of what? Human, natural and artificial, emergent. You get hybrid. And most of it's not that impressive because people are automators or they're validators. But in these cyborgs, you couldn't tell who made the prediction. Was it the machine? Was it the human? You analyze the transcript, they're coming up with decisions neither would have made alone. But what predicts that behavior? Not the benchmark. Like I just said, the small open source models, they did nearly as well as when they had a state of the art model. So what did predict it? Curiosity, intellectual humility, fluid intelligence and a social skill perspective. Taking your ability to understand other people predicted success. Neat that I'm not the first person to see that in the data. And you know why? Curiosity. If the AI gave you a good enough answer, did you keep looking for a better one? Why intellectual humility? If the model told you you were wrong, did you ignore it? Did you just sort of crumble and then do whatever it said afterwards? Or did you say, okay, so not that, but let me think a little more deeply. Given what you just told me, I want to explore a little bit more. So what we saw is the humans would explore and then the model would sort of say, ah, ah, but what about the data? And the humans would say, okay, let's explore over here. And then the model would sort of pull them back. So the humans would explore the ill posed problem space where we don't even know what the right questions are. And the models were really specialists in the well posed. We understand the question, we know how to get the right answer. The best results came when the humans injected that exploratory, ill posed information into their collective decision making. Again, that's kind of why I'm saying like that distinction between the model and the person starts to disappear a little bit, um, because they're making decisions neither would have made alone. So that study with the kids where they're fostering curiosity by praising questions rather than answers, why am I interested in that? Because curiosity is both a predictor of positive long term life outcomes and people's Success using agentic AI, and you could run through a list of things that we know. Some things that didn't pop in my experiment because maybe the time frames were too short, but, like resilience, I, uh, also suspect that that's a big predictor. So how do you develop these in people? Well, if you want to get more resilient, you have to experience failure. No failure. No resilience. Because all of these qualities, unlike teaching you how to factorize a polynomial or how to code up a neural network, those are just hard knowledges. I can give you a lecture about them now. You know the facts, and in theory, you could do it, too. You want to get more resilient, you have to experience it. You have to experience failure and then find success out the other side. You want to get more creative. We need to retrain your dopaminergic system to give you dopamine signals. Not that the thing you need to do right now is get on Instagram, but to give you the signals that the thing you need to do right now is explore a question is to make a mistake so that now you know what you don't know. And in these amazing, um, students and the cyborgs, they're willing to be wrong, because being wrong becomes a learning signal. And they have learned to treat that as a positive rather than a negative. So this is something. I started doing this with my kids. I actually, this is fighting fire with fire. At the end of every week, I wrote a little program, but you could probably just do it as, uh, a prompt where I have Claude review their interactions with all of their AI tools. And then it writes a little report. Here's where you just did what the AI said without seeming to think about it. Here's where it gave you some good advice, and it seemed like you just ignored it. Here's where it gave you an idea, and you took it somewhere better. So the good, the bad, the positive, the ugly, and it surfaces at all. This is, you know, this is sort of metacognition, but it's a metacognitive augmentation.

Speaker C: Yeah, lovely.

Speaker A: So you're using AI to help you think about your own thinking and your own interactions with AI. Cause otherwise, at the end of the week, do you remember everything you did on your phone or at a prompt on your computer? Surfacing what you were doing well, and making you think about how to do it better next time takes you in that direction.

Speaker C: Yeah, fascinating.

Speaker D: Yeah, that's really great to have that, um, that review at the end of the week. I Think I'll get a feedback categorization and I think um, you know, your sort of focus on curiosity and asking good questions is really interesting because um, you know, you hear a lot about all the skills we're going to need with AI include human relationship building, human uh, kind of judgment skills, critical thinking skills. I don't hear nearly as and curiosity. And I know from a personal thing it's a lot easier to be humble with AI than it is to be with another human. It's a private thing, like I can be wrong, but it's only with the AI and that's, you know, so it's easier to be humble actually.

Speaker A: I agree. I think there's some de risking here. It's interesting because our brains simultaneously probably treat it more like a real person than it ought to. It's so fluent. The social signals are there. Like I said about that Turing Test, we almost think it's more human than humans are. Uh, it's sort of a Blade Runner ah, riff here. But it also just doesn't trigger those same. There's no gaze following. This isn't some public humiliation. Yeah, spooky Android that's interacting with you. And so you can really have a sort of different relationship. Another interesting thing about these models and you know, facilitated by the fact that I'm saying again, please genuinely believe me, but we could nerd out and get into it if we really wanted to. These models are not conscious, they are not self aware. I am very interested in how we build models like that. But these models are not that. They can be intelligent and not be that. And I think maybe that's something that non cognitive scientists don't fully appreciate. So saying that, yeah, it kind of frees up some of the tenseness of this relationship. If you dispense with this idea that you're talking to another person but you are still talking to something intelligent, it can answer your questions and it really doesn't care because it's simply not capable of it. Now what you can do is say I want you to be a certain kind of person. So here's another thing. I do write about this in my book. For example, I write about the Nemesis prompt, which I used to write the book. So what I didn't do is let. I was using a lot of Gemini at the time. I didn't let Gemini write anything because frankly it's not a great writer. But what I did do is I would put in the work. And I don't love writing, I don't love coding, I don't love any of this stuff. But I would put in the work a brain dumping down into a chapter. And then I would say, hey, Gemini, you are my nemesis, my lifelong enemy. You, um, have found every mistake I've ever made and pointed it out to the world. Here's a draft of my new chapter. Read it in detail and explain to me why I'm wrong and what I can do to fix it. But what's really cool and different about modern, uh, uh, language models is two seconds later, I could spin up a next prompt and say, hey, now you're my next board audience that doesn't get it, doesn't know why this is meaningful to them. Here's my new chapter. Explain to me why this isn't connecting to you. Why is this not making any sense so you could role play so many different entities? I think there's a lot of really cool work, for example, being done on simulated market research before you go. Spend a whole bunch of money on focus groups, simulate it. I think you do need to, in the end, talk to some real people, but it can help you find the right questions. It can give you insights to go on test. And me, I'm a theoretical scientist. This is what we do. We simulate things so that the empiricists, those dummies that weren't as smart enough as us to get a job where you didn't have to stick wires in a brain. We give them these insights, and if we're lucky enough, they say, you know what? That is kind of cool. I'm going to run that experiment.

Speaker D: If you could have Sam Altman, who runs OpenAI, and Dario Amade, who runs Anthropic, which has Claude and the heads of Gemini, and let's throw in Grok and a couple of others, you could get them in the room and tell them, um, you really want them to focus. What is the one thing you would like to see these models or these companies doing now?

Speaker A: I'm putting my money where my mouth is in answering this question at my nonprofit, the Human Trust. Amongst many projects we're working on, one of them is on a hybrid intelligence index, a benchmark for how the models are impacting the humans using them. And right now, even benchmarks on alignment are about what the model does autonomously, all by itself. In theoretical scenarios. Nothing is about empirically driven research on how people use these systems. So my easy way of saying it is we need models that give us what we need, not what we want. And the sales pitch right now is efficiency, efficiency, efficiency, automate. As Much as you can, don't say it publicly, but downsize, downsize and figure out how much of this work can be done by a machine. I genuinely believe there's a lot of jobs that ought to be automated. I don't think humans should be doing boring, soul crushing work, but the benefits of this should be going to people being able to be explorers. Uh, we don't have a good benchmark for this, so this is one of the main things for me is please stop selling AI as a quick efficiency trick. Even the data out of Anthropic and others is showing it's senior level experienced people using these tools creatively that are getting all the productivity boosts. It's almost certainly hurting career development in young individuals because it's just doing it for them. So please, let's take a different look at this. How can Claude code, and I really respect Anthropic, um, for releasing a paper showing that people are not all of them, but some of them using clog code are getting worse at coding. How could clog code provide what you need to write the best code right now, but also the better code next year? And for a senior person, that's going to be a different answer than a junior person. How can Gemini collaborate with me on my scientific research such that I'm a better scientist and we together are producing a better paper than either of us who would have written alone. I get it. Much like it's hard to build a school system around these complicated traits like curiosity and resilience, it's hard to think through, uh, maybe on some level what these models might look like, because just being able to answer questions correctly, hey, that's a solid benchmark. I can just fine tune around that. But I need them to step up and do so much more. The reason, after nearly 15 years after I started writing my book that I finally decided to publish it was hearing Sam Altman get in front of the US Congress and say, imagine a world in which every kid has their own AI tutor. I guess Sam doesn't know that. We've researched this for 50 years. And there's a golden rule in AI in education is that if it ever gives students the answers, they never learn anything. Wow. Which has been replicated with LLMs and um, with GPT specifically. It literally hurts learning, not because it inevitably will, those cyborgs will make wonderful use of it, but because most people will unthinkingly use it that way against themselves. We can build better people and we can build better models. Doing both of those together is the right way forward.

Speaker C: Yeah, I really, I love that vision. I think it's uh, it's so compelling. Before we sort of wrap up the conversation, we really love to get sort of practical and down into sort of the use of tools and how our uh, guests use, you know, I know you're using probably very sophisticated tools that uh, perhaps our listeners wouldn't come across day to day. But in terms of the sort of the Gen AI LLMs, how do you personally use them in your day to day life?

Speaker A: Right now I sort of leaning into what I've been saying all along. Where do I have expertise but maybe not the reach. So I'm working on a very complex dynamical systems model and you know, I kind of came to math a little later and so I have a foundation in that. But that's neither the area where a lot of my theoretical work uh, is done nor do I have a big educational history in that space. So I know what I want these models to do and I understand what structurally they're going to look like. But if I had to do this completely on my own, it would take me forever. And so I, I have a little team. There's me, I've got a Claude prompt and almost everything I do, I have it turned all the way up. I rarely am using anything short of opus or 4.8 Gemini with full thinking, uh, all the way up. I know it eats up my account really fast, but I want it to be smart. This is the value add uh, that I'm looking for. I'm not looking for it to spit out a good enough social post for LinkedIn. I'm looking for it to help me craft a complex dynamical systems model that explains mathematically some of the things we've literally been talking about in this conversation. I'm using tools people have use every day, but I'm really pushing them to their limits, uh, so that I can push myself to my limits. And sometimes that means. Wait, stop. You're saying I should add this term in? I actually think you're probably right, but you've gone a little bit on me now. Walk me through why this is the right thing to do. Why is the standard use case, give me some citations so I can go read papers on it. So I'm really pushing myself here. But I'm also get, I'm the senior part of this collaboration. If I'm not understanding what it's telling me, then it's totally wasted. Um, when we write the paper together where I do the actual writing and it does this sort of aggressive loyal opposition response. I gotta understand this stuff or I'm not going to be able to provide answers to myself, much less to my peers. So that's one. But, uh, you know, a lot of what I do are models I build for myself, bespoke for a problem I'm trying to work on.

Speaker C: Right.

Speaker A: But we're also looking at, hey, every labor economist has an opinion about gender wage data. What accounts for the wage gap. And interesting enough, from the most progressive to the most conservative, the data all kinds of shows the same thing, which is wage gap is largely. They're not exclusively a function of how many hours worked. Women work fewer hours than men. Done. I'm like, I believe the data. Do you literally think women are from Venus? Like, why would smart, educated people make seemingly irrational choices? Why are they choosing to work fewer hours? So I didn't instas scratch. I built a web crawler, attached a bunch of deep neural networks. This again, was a few years ago. So this was older stuff. And I analyzed the faces and read the quarterly reports of 63,000 companies and came up with a novel finding about gender wage gap, where we found that young women work much more similar hours to men when there's tangible evidence, I. E. Women in senior leadership positions within their companies, otherwise identical. Women simply invest more hours when there's evidence that their hard work will pay off. Not super shocking, but no one was saying the seemingly obvious thing or had empirical evidence of it. That work took me three days to get my first results. Imagine the army of research assistants and the years of data collection it would have taken even 10 years prior to that. Now, with agentic models in theory, anybody could do this.

Speaker C: I think what you're saying is that you're basically using these models to really push outside the box, find creative, innovative ways of thinking.

Speaker A: What is the question nobody else is answering? Even if I don't think it's gonna take me to the right answer, the right answer is free in everybody's pocket in a very real, literal sense. So it's what I would say that's different than anyone else that provides value to the world. Because even if I'm wrong, at least now we know, we understand that space, that part of hypothesis space. Uh, now we can look elsewhere for the next best answer. So we need humans to be doing that, exploring.

Speaker D: It's, it's fascinating and I think, you know, taking away in particular sort of this curiosity and this sense of how do I be the thought leader and partner with AI what are kind of what's the main thing you're working on right now. Problem wise.

Speaker A: My new company, Possibility Sciences is building a hybrid intelligence system for innovation itself. Um, so we're exploring science and innovation writ large in a system like that. In a philanthropic project at ah, the Human Trust, we're both analyzing and building tools to promote political pluralism. In a pilot project for the run up to the 2027 French elections, one of my companies is working in neurotechnologies for Alzheimer's. Is there a role for agentic models in supporting people in cognitive decline so that they can be more of themselves for longer? Uh, rather than just giving them easy answers, push them to be better. I have a ton of projects at any given time that I'm working on, including my next set of my next books. And here's the real fun one. A screenplay.

Speaker C: Wow. How do you have time to work on a screenplay?

Speaker A: Uh, my other project is I'm working on a time machine. I'm trying to add an extra two hours to every day so I can do that.

Speaker D: You are Vivian. This has been such a fascinating discussion. If listeners want to find out more about you and your work and the companies and projects you've just mentioned and obviously you know, your books, we'll put all the links on the show notes. But just briefly, where should they go?

Speaker A: The simplest place, because it's sort of my hub for all of my companies and projects is socos Labs. So that's s o c o s dot org. You can find my newsletter there. If you really love what we can do, sign up for the paid newsletter. I like to believe you will find some truly fascinating things like the Director's Cut original long version of my book because it's rolling out chapter by chapter just to the subscribers. But really it's just a vote in favor of doing this kind of philanthropic research that just doing some good in the world is intrinsically worthwhile. So visit there and then from there you'll find links out to my various companies and my creative projects and everything else.

Speaker D: It's so great to hear someone like yourself working on the kinds of problems about building better humans and making sure we really thrive and stay healthy, well rounded individuals given momentum that is behind this freight train of AI for efficiency. So thank you for the work you do. Keep it up and we'll put all those great notes, um, and links on our ah, show notes page. And it's been really great to speak to you. Vivian Ming. Thank you very much for having us.

Speaker A: Really, it's a pleasure. Thank you for having me. Hey, it's Ryan Reynolds here from Mint Mobile. Now, I was looking for fun ways to tell you that Mint's offer of unlimited Premium Wireless for $15 a month is back. So I thought it would be fun if we made $15 bills, but it turns out that's very illegal. Uh, so there goes my big idea for the commercial. Give it a try@mintmobile.com switch upfront payment

Speaker D: $45 for three months, $90 for six months or $180 for 12 month plan required $15 per month equivalent taxes and fees Extra initial plan term only greater than 50 gigabytes. Me slow when network is busy. See terms.