Stop Letting AI Think for You | Dr. Vivienne Ming

33:15So when we ran this experiment, we need to figure out how do we do what you're kind of talking about? Say you want to put a business plan together or write a screenplay or what have you. And you have a jerk like me saying, no, no, no, no. Think for yourself. How do you find value out of this stuff? Well, I'm kind of suggesting maybe you aren't getting as much value out of it as you think. You're letting it do your thinking for you. Because here's one thing that the broad research in the space is already showing, which is, guess what? It turns out everyone else using that same tool is now having that exact same idea because they're all getting the same answers.

33:54So among scientists, for example, scientists using AI, they publish quicker, they get more citations, they reach tenure faster. The fields of science that are getting more heavy AI use. Everyone's converging to the same questions. Their papers are reading the same. So helping scientists hurting science. Helping you in the moment maybe get that business plan out. But now you're writing the same business plan everybody else is writing. So for me, what I do and what we see in this experiment is, how can we do this? How can we have an experiment to test not just what human intelligence or artificial intelligence can do, but let's call it hybrid intelligence, humans and machines together?

34:38Can they do something that's qualitatively different and ideally better than either of them alone? And how would we measure it? How about we make them predictions of the future? That way the AI can't simply memorize the answer. And we pull all these questions off of Polymarket, which is getting a lot of press right now. So we pulled these non-crazy, so it's not celebrity questions and it's not sports. So we pull these questions off of prediction markets where people betting, let's be honest about it, on future outcomes. And what we find is humans are terrible at this. So we grab three people. We give them an hour to make 10 predictions about something they know nothing about.

35:20What's the price of oil going to be in six months? Right now, people may know the price of oil. But that's pretty unusual. That was not true when I did this last summer. And they're above chance. Well above chance. But still, terrible. Ask an open source model like Gwern or Llama to do this. Not way better than the best humans. Ask a bleeding edge model like the best GPT. Better still. give a human and a model together and suddenly they're about 60 of them are about as good as the model was all by itself because they just say what will the price of oil be in six months and then they submit that answer they're the automators they're the ones we see no brain activity they're on their way to early dementia uh i know that's a bold claim but again this it's not a hard one because there's a lot of evidence that if you don't think you're looking at a bad future um the next group are the validators they think at least i know the answer gemini tell me why i'm right gemini gives them all this tells them why they're right tells you why you're handsome tells you they smell great and then they submit their answer and they do worse than the ais by themselves.

36:39I have now described 90 to 95 % of all participants in my experiment. That's terrible. Why are we there if AIs are giving better answers all by themselves, even the cheap little open source ones? Then we have this last group that I call the cyborgs. I wish I could say I coined the term because I'm a sci-fi nerd and I went to grad school interviewing, telling people I wanted to build cyborgs. They scooted away from me. It's kind of the term that's become the term of art in this space of looking at human computer interactions. When we look at the cyborgs in their hour transcripts, interacting three humans together, and even just an open source AI, what we find is, one, it's not even clue who made the decision.

37:30Was it the humans? Was it the machine? the humans explore and the machines pull back to the data and the humans explore and the machines pull back they go through a subtle cycles of this i call this the machines are the master of the well posed the questions for which we already know the answer even if those answers are esoteric or involve complex formulas if this problem already has a known answer that is no longer a job for a human being. But then there's the ill post. My job. I don't even know what the question is. Day one, let's start trying to figure out what the question is. Humans are terrible at this.

38:10We are terrible with uncertainty. We are terrible with the unknown. We are the only game in town. We have, for now, a unique ability to explore the unknown. And interestingly, the kind of mistakes we make and the kind of mistakes even the best AI makes are different kinds of mistakes. When you respect that and you bring those two things together, hey, Claude, I just wrote this chapter. It's titled How to Robot Proof Yourself. It's a set of recommendations, actions you can actually do to help. You know what, Claude? Today, you are my nemesis. You are my lifelong enemy. You have found every mistake I've ever made and pointed it out in excruciating detail to the public.

39:08Take this new chapter and tell me why I'm wrong and what I can do about it. And then the cool thing about AI is then I can go and say, hey, Gemini, you are my next bored audience that doesn't know why any of this applies to them. Read through this chapter. Imagine I'm giving it as a talk. Tell me why it isn't connecting and what I could do to really make this click. and then the flip side is then as a human i have to one take that feedback learn and change and two in that learning no you know what it was it was right about points two and three but it's wrong about five and six i i don't think i'd understood what i was saying so here's the cool thing that came out of this research that we did what predicts hybrid intelligence because it turns out those best teams those cyborgs were better than polymarket the smartest thing on this planet is not the latest version of gpt it's not terence tau as brilliant as he is it is a reasonably smart small group of people in combination even with a small open source AI.

40:35So it turned out the AI benchmarks didn't predict much of anything. And I have an important point about that. It was almost entirely predicted by human capital. And traditional things like G and IQ, not so much, but fluid intelligence, working memory span. So being smart helped, but just as predictive was measures of curiosity, intellectual humility, and perspective taking or theory of mind. Humans' ability to understand other humans predicted their ability to get the most out of AI. To recognize when the brilliance and eloquence of Claude is a sham. Yeah. When GPT is giving you an undeniable answer and you still have to say, I don't think we're on the same page.

41:32I think maybe you didn't understand what I was asking about.

41:35Johnny:Johnny, I didn't realize how much a mattress could change my nights until I switched to Lisa. That soft sink in moment when you lay down, but the support is there, you wake up feeling aligned. Sleep is everything. You can eat right, you can exercise, but if you're not sleeping well, none of it matters. Lisa has a lineup of mattresses tailored to how you sleep. Each one is designed with specific sleep positions and feel preferences in mind. From night one, you'd feel the difference. Premium materials, full body support, no matter how you sleep. Lisa mattresses are designed and assembled in the USA.

42:10Johnny:Free shipping, easy returns, 120-night sleep trial. New York Times Wirecutter named them the best hybrid mattress. West Elm features them as their go-to mattress partner. They also donate thousands of mattresses each year to those in need and partner with CleanHub to remove plastic waste from our oceans. Go to Lisa.com for 30 % off mattresses, plus get an extra$50 off with promo code CHARM exclusively for our listeners. That's L-E-E-S-A dot com. Promo code CHARM for 30 % off mattresses, plus an extra$50 off. Support our show and let them know we sent you after checkout. Lisa.com, promo code charm.

42:54Johnny:A thoughtful wardrobe comes down to pieces that mix well and last. That's where Quince shines. Premium fabrics, considered design, everyday essentials that just feel effortless to wear. Lightweight cashmere sweaters, short-sleeved Mongolian cashmere polos, linen bottoms, tees, and 100 % Pima cotton. Versatile pieces that make your wardrobe work season to season. Quince works directly with top factories and cuts out the middlemen. You're not paying for brand markup or fancy retail stores. Just quality clothing. The cashmere is 100 % Mongolian. Same stuff luxury brands use. The Pima cotton is long staple.

43:36Stays soft. Doesn't pill. European jersey linen is breathable and lightweight. Everything is built to hold up.

43:42Johnny:I gotta say, these linen shorts have become my go-to. They don't wrinkle like cheap linen does. They work with everything and they didn't cost a fortune either. Stop overcomplicating your wardrobe. You don't need a closet full of options. You need the pieces that work. Right now, go to quince.com slash charm for free shipping and 365 day returns. That's a full year to build your wardrobe and love it. Now available in Canada too. Go to quince.com slash charm for free shipping and 365 day returns. Quince.com slash charm. Well, I think in this study made headlines around Stanford looking at the accuracy of what we're getting in response to AI and how important it is for AI to agree with us to get us to keep using it.

44:34Johnny:So the race is not for the AI that they give us to be smarter. They're doing that internally for their models to sell elsewhere. But the race for the AI that's public is to get us to use it more. Therefore, we become a subscriber. We can actually pay for all the research and development they're doing for these other models. So they have to tune it in a way that keeps us coming back. And a lot of that is agreeing with us, telling us our terrible decisions are right, making inaccuracies, speaking eloquently about things that it actually doesn't have information about but will satisfy us. So having that critical ability to question what AI is giving us, it sounds like, moves us closer to the cyborg, moves us closer to robot proofing.

45:11It is huge. I have a couple of just complete quirk, random things about my later life career here. One is that recommendation to work on that project. The principal investigator in that was a famous neuroscientist, computational neuroscientist called Terry Sanofsky. And his advisor is a guy named John Hopfield. So my original PI, my original advisor, who was also the advisor to my grad school advisor, he shared this story from John Hopfield, which was science is a story. And that story should be an hour long. Well, John won the Nobel Prize in physics for his early work in AI. So here's one thing.

45:59like literally on the flip of a coin and then a random recommendation i so maybe i never win that novo prize but i ended up working in this novo prize lineage on artificial intelligence machine learning what a weird thing but the other is my first company um was using ai in education which I will tell you in 2008, what did VCs not want to fund? They did not want to fund, so my wife and I started this company, two married women doing an education company. But they loved AI, thought it was magic. You can read students' minds. Here's$2 million if you do financial fraud detection. And I've got the perfect CEO for you.

46:46By the way, I was the CEO learning that I was being fired in the moment of doing this. Taking the money. But we, as people say, pivoted. Terrible decision in the end. But we pivoted and got into unstructured log file analysis for social gaming, where we learned how Zynga builds its social games. And they start with these behavioral scientists, these behavioral psychologists. They don't start with the game idea at all. It was reward schedule, reward variability, all this stuff out of behavioral psychology. Where did they hire these people from? Here is Casino. Which is so deeply wrong. I don't just mean it in a kind of weird moral sense.

47:35It's wrong because it misses the value point of these truly powerful models. Because here's the last twist in our experiment. We only get this tiny percentage of people to actually exhibit this cyborg behavior and do this cool stuff and beat the market. Naive people that know nothing about the predictions, doing better than the AIs alone, doing better than Polymarket. well having identified the human qualities then we could play the long-term game that's what education should be about that's what hiring should be about that's what management should be about it's not about whether you know what the caudate nucleus is or whether you know how to code up these models cloud code will write the code for you it is about do you know how to explore the unknown.

48:29You have these deeper foundational qualities. And you also have to know the facts, it turns out, but that's a secondary thing. But how do we get more people? What about the AI side? If all these AI benchmarks that people are so proud of and are entirely building their models around aren't driving the superhuman behavior, then why do we care about them so much? So I built a model that would do disastrously bad on these benchmarks. Because what we did was we took an open source model and then we fine-tuned it to never give answers. Doesn't matter what you do, it is Socrates. It's nothing but context and questions.

49:13The participants in this experiment hated using it. It was so frustrating. What's the price of oil going to be in six months? That's a really interesting question. Have you thought about what the factors that could lead to shifts? Are there some things and, you know, here are historical trends that might interest you. They were so upset to not be able to get an answer. Twice as many people, instead of 10%, 20 % exhibited cyborg behavior and showed superhuman performance. So this weird thing that shouldn't work. The AI gives you less information, not more. cognitively it forced people to think it still gave them information it just didn't give them the answer it forced them to think their gamma went up which hey is great for the long term um but more importantly they beat the models by themselves just with this little tweak We should be building models that aren't benchmarked on how good a model does on its own.

50:24We should be models that are benchmarked on how good they make us. And nobody is right now.

50:32Johnny:Well, I think unfortunately, the truth is the companies are betting on the diminish of human capital. They're shorting human capital in this equation and they're going long on the tech. So naturally, the benchmarks and the tech outperforming humans is the only thing that matters. You're 100 % right. A lot of people clearly have this vision of a sort of human-free future. And I'm telling you, whatever problems a universal basic income solves, it doesn't solve anything about human meaning or value. And again, I'm a hard number of scientists. So what I mean is it doesn't solve the problems that drive people to have amazing life outcomes.

51:18It might pay for your rent and allow you to buy some soil and grain, but only the non-humankind because it doesn't taste good. um but all of the things when i do my research and we look at what predicts positive life outcomes 122 million people when i was the chief scientist one of the first companies doing and hiring and it is resilience and a sense of purpose it's working memory span and perspective taking all these things i've been mentioning already the things that make you good at using ai make you robot proof and make you robot proof they predict that you'll live longer they predict that you'll be happier you'll earn more money um it's not where you went to school it's not the facts that you know it's not the things you claim as your skills on your resume it's these deeper qualities

52:14Johnny:but back to that obama administration conversation this is not a six-week upskilling exercise like what you're talking about building resilience perspective taking this ability to have the curiosity to question the answers given to us and not cognitively offload and shortcut our way to moving forward, being more productive. Well, that's a hard problem to solve for humans. That's not embedded in our education system. That's not embedded in value right now as a society. The productive output of us in our job is valued in the dollars and cents that the companies are making. And it's not to say they're evil.

52:49Johnny:But when you have Sam Altman come out and say, hey, AI takes energy, humans take energy. Well, it actually is less energy to operate the AI. You know, it does start to question, well, what is even robot proofing if that is the ultimate destination? If that's the ultimate destination, I think one of the things I try to get at in robot proof is that this is a choice and that the automation choice is a short-term one. long term, even if you just want to think in pure productivity perspective, long term productivity goes up when we're thinking about cyborgs. You know, the way I put it is stop trying to sell me robots.

53:38I want to be a cyborg. I don't want to buy C-3PO. I want to be a mentat. This is my sci-fi nerd cred. But that's no one wants to sell me that. Well, they want to sell me the automated experience because I get that dopamine rush of solving the problem, but I'm not learning how to do anything. So this I mentioned, I admire Anthropic. They released a report recently that showed, you know, that maybe only Anthropic or OpenAI could have released, which is people using cloud code, developers using cloud code. were faster and better when they were using it. And other people have published research on this to date.

54:30But they ran an experiment. They looked at your ability, your knowledge, complex and basic knowledge of coding and computational theory and all of these things before and after a big coding task. and some people got access to cloud code and some don't. And the people that had to code on their own simply learned dramatically more. And so I really admire that Anthropic was open, that it made them better when they were using it. But they became worse at the very thing that was supposedly their job. Again, if you see no future for humanity, then maybe you're not worried about that but i flip it around and say right now cyborgs are the smartest thing on the planet and they will only get smarter as ais get smarter which is one of the beauty of that vision of it so we should be working incredibly hard to both fill the human capital pipeline for that don't settle for five percent right 10 20 hey i'm a i'm a dreamer but i'm also a realist i we're certainly in this generation of workforce we're not bringing everyone along for this ride i wish i could if i could wave a magic wand we would so we need to do something else like take all of the amazing surplus this is producing and how about instead of spending a whole bunch of money bombing things in various places, we spend it rebuilding every bridge in the United States.

56:15We kind of need those bridges. How about rebuilding the energy infrastructure? We really need that to give a whole generation of people who are simply never going to join this coming economy, a job with decent pay and dignity that's truly building for the future and then bust our asses to get their kids ready for this new world.

56:38Johnny:Well, I think I want to speak for myself, but then also probably a lot of people listening is, you know, we're in a pretty great job that we worked really hard at. We have sacrificed quite a bit for the thought of putting on a hard hat and building bridges. It's not really appealing to me in the world that I'm in. And I don't think it's appealing to many in our audience. And what I hear in this is, OK, five percent of people have this natural proclivity in this experimental setting to be the cyborg. And you've identified the skills it requires to be the cyborg. But the coming train is it's musical chairs.

57:11Johnny:There's only going to be so many slots available at companies to be the cyborg. So how do you get there faster? What can I be doing right now to move to that direction as those seats are diminishing? We're seeing even the top companies who are hitting share prices unimaginable and profitability through the roof. They're cutting teams now in anticipation that AI is going to continue to be the productivity engine of the company. So the number of cyborg roles may have been 20 ,000 yesterday. It's down to 10 ,000 today and might be down to a thousand in the next five years. All those companies trimming people are trying to trim down to the only right ones.

57:53And, you know, as someone that's been And in this economy for 30 years, this one that I'm saying we're headed into, here's, if you were a fan of the show House, you know what my day job is like. I get up in front of my employees and I say, how do we measure the impossible? or this kid can't enter REM sleep and his doctors at the Harvard-MIT health sciences facility can't figure out what's wrong with him and they've come to us. How are we going to fix this? Go, pitch me ideas. All of which are wrong. I mean, what are my junior employees that have never done this before going to figure it out? But I take the time to explain in detail.

58:48Ah, here's why that's not going to work. And then here's why that's going to work. That idea was so bad, I'm actually going to make fun of you, but we'll skip the psychological torture from House. but watch an episode of house that's what working with me that's our process is we take an idea and we brainstorm crazy ideas that could possibly be why this is happening and what we could do about it and again and again and again we're shooting these things down i need them to be productively wrong. Desperately need them to be productively wrong. I need us collectively to explore new spaces. Because unfortunately, right now, as information rates go online, information is so available, masses of it.

59:43And then if you think of it, the flip side is with AIs in your pocket, the complement to that is information costs are essentially zero. Human beings do this weird paradoxical thing. As information availability sky wrecks, if the flow of information goes up exponentially, we stop exploring. Like the science, those fields of science collapsing.

1:00:09Johnny:It's the exact opposite of what we need to be doing. Exactly. Humans have this ability to explore. That's what we found in our experiment. Exploration necessarily requires being wrong. Otherwise, you haven't explored anything. Somewhere out there is a better right answer, but we end up systematically demonstrably hurting around safe answers. AI, as automators and validators, just causes us to hurt around those safe answers. So those companies that are trimming to the bone to only have the right answers are trimming down all of their future potential. Spring starts at The Home Depot, and we are bringing the heat to your backyard this season.

1:00:55Fire up the flavor with our wide variety of grills for under$300, like the Next Grill 4-Burner Gas Grill that's perfect for hosting your spring cookout. Then set the scene and turn your outdoor space into the go-to spot the patio sets for every budget. Bring it this season with grills that deliver flavor and patios that set the vibe from The Home Depot. Start your spring with low prices guaranteed at The Home Depot. Exclusion Supply, see homedepot.com slash price match for details. Score more with the college-branded Venmo Debit Card and earn up to 5 % cash back with Venmo Stash. Got paid back?

1:01:27With the Venmo Debit Card, you can instantly access your balance and spend on what you want, like game day snacks, gear, tickets, and more. The more you do, the more cash back you can earn. Plus, there's no monthly fee or minimum balance. Sign up now at Venmo.com slash college card. The Venmo MasterCard is issued by the Bancorp Bank N.A. Select schools available. Venmo Stash terms and exclusions apply at Venmo.me slash stash terms. Max$100 cash back per month.

1:01:50Johnny:It's the right answers today. But as we know, today is not predictive of the future in any business venture. So to kill all of your exploration as a business, you kill all of that future potential. And AI is limited by our input as humans on the exploration. It's still in our head that we are these weird factory line like engines. and wouldn't it be great if we could just automate the most expensive part of that modern factory line, which is you. And I'm going to argue there's a whole bunch of stuff that I fully agree should be automated. I mean, I'm from Salinas, California, where John Steinbeck is fun.

1:02:32Human beings shouldn't be bent over in a field 16 hours just to pay their food for their families, as long as we can find something else for them to do. and there's a lot of things that go on in a modern company that are perfectly reasonable to automate.

1:02:46Johnny:So the leaders are wrong and especially the ones who are doing these massive layoffs and thinking in the here and now but many of our audience members are not in that position to make those cuts. They're on the receiving end of those cuts. So I've heard the skills that make me robot proof. I recognize now okay I need to start developing these skills and I think it might be helpful because many of us also feel a lot of pressure to be learning and implementing skills with AI augmentation. So how can I strengthen my curiosity with AI? How can I strengthen my perspective taking with AI? What can I do to develop resilience?

1:03:22One is a failure dictionary, a failure diary. And the goal isn't just write down everything you did that was wrong. Listen, everything I do that wrong plays out in my dreams. I do not need to write it down. I remember all of it. What you're making the diary about is how that led to success. So I mentioned something and you came back to it. Resilience. Hugely predictive of life outcomes. Hugely predictive of career outcomes. When you're exploring the unknown, you got to be resilient. And we define in psychological terms, define it as your probability of finding success after experiencing failure.

1:04:03So guess what is the only way you can become more resilient you have to experience failure in fact and process it in a meaningful

1:04:13Johnny:way genuinely process it not ruminate on it not be stuck by in fact this is one of the problems again of that shallow use that automator use of ai is it feel like you're coming up with the answer it feel like you're right which means so you got this part of your brain called the anterior cingulate cortex the acc and um it's funny when i was in grad school we used to call it the oh shit circuit because if you were doing like a brain imaging fmri um of someone doing some sort of a task and they like had to choose one thing or another and it was like time pressure and they're like they pressed b but they they instantly knew oh i was supposed to press a this thing would light that.

1:04:57Oh, shit, oh, shit, oh, shit, oh, shit. Well, the whole reason that thing lights up is because it's part of an extended circuit involving your medial prefrontal cortex and your amygdala, which involves emotions and fear, and your nucleus accumbens, which is your reward center. If you don't get error signals, just as like when an AI and neural network doesn't get error signals, you can't learn. No error signal, no learning. That's what essentially anthropics showed. When Cloud Code gave developers the answer, they didn't learn anything. So resilience, if you're not experiencing failure followed by success, and if you're not connecting the dots between the failure and success, you know, I pitched that idea.

1:05:44It wasn't right. The team didn't take it up. They didn't want to listen. But, you know, I wrote it down in my dictionary. I realized a couple of weeks later that that was why we shifted to this other. No one would have thought of that if I hadn't pitched that wrong idea. It led to the breakthrough. And now I remember, if you're leading a team, and I'm not saying you're the boss, you're a manager, are you rewarding productive failure? I mean, I don't even mean bonuses. I just mean attention. You know, you went out there and said something you knew probably wasn't going to be the idea, but it needed to be said for us to be able to move forward.

1:06:28And that was hugely worthwhile.

1:06:30Johnny:I think the problem for many of us right now and the way that we're interacting with AI is we are actively using AI to try to avoid failure. Exactly. To get the correct answer, skip over all of that nonsense of bouncing ideas that don't work and figure it out. and one, we're a bit too trusting of AI in that, oh, we avoided the failure, so I'm just gonna go ahead and do it and we're seeing this now in personal relationships. Again, that study from Stanford was really eye-opening around how many people are asking for therapeutic relief to difficult conversations, things in their lives and because the models are tuned to keep us talking to them, we're getting answers that are completely wrong, but we're trusting them, we're building up a different identity, sense of self around it and the world around us And then I also feel like many of us were coming at it with the idea, OK, we have a context of a million.

1:07:20Johnny:So I could just ask it a simple question. It's going to do all the different connections I need to get the right answer. And we're not actually testing it in a way that allows us, even if we're getting an answer from AI, to collect the failure in that and come back to AI. And in our own business, you know, we're using Claude a lot in marketing, but it's all experiments and tests. and we're coming back with the numbers and saying, hey, like you gave us this angle, this position, this marketing piece, it didn't actually work. Even though we fed you, here's all the followers, here's all the customers and here's what they shared on this quiz, we need to adjust course.

1:07:52Johnny:And we're learning through that process of going back, collecting the data, running the experiment and acting like they are co-scientists with us in the experiments that we're running. What it feels like is I'm interacting with my grad students. my grad students let's say at berkeley university college london they are brilliant truly brilliant i'm there because they know everything and they understand nothing and when i started interacting first with bard and then gemini and gpt and claude it felt exactly the same except here when i say they know everything i mean they know everything. The answer to every medical licensing board question, the answer to all the bar exams, the answer to my esoteric statistics problem.

1:08:43But it doesn't understand any of it.

1:08:47Johnny:That I think is really a huge belief shift that I would love the audience to take out of this. And you know, in my prior life before the podcast, I was a scientist in cancer biology. And I look back at my beliefs about the state of cancer and treatment back then. And it was all built on faulty observations that were held as gold standard, published in Nature, that were observed in the top labs by the smartest people in the world. And what they observed and what they predicted would then happen in the patients. You know, I've been out of the lab 20 years. We're not doing those treatments anymore.

1:09:19Johnny:That is all completely wrong. But guess what? That knowledge has been fed into LLMs, and that's exactly what they're acting on. And because they don't have any real sense making, AI is not operating in a lab. It's not like Gemini is taking all this information, then going and running more experiments, shifting what they're observing, starting to recognize, oh, well, we weren't measuring it correctly. We're basing it all off of all of human knowledge and a prediction engine based off of all of human knowledge up until this point. But if you ever worked in science, you'll recognize that if you talk to the grad students who then go on to get PhDs in the work that they do, it's going to disprove a lot of what you believed, what you thought was right about the world and how everything works.

1:09:58That's just how science operates. As people say about science, our job is never to be right because that is literally fundamentally unachievable. Our job is only to be less wrong. And the purpose of science is for humanity to be less wrong over time. In psychology, there is this classic notion of model-based and model-free cognition. The model-free cognition, let's call it statistical learning, is just learning the patterns. There isn't really causality there per se. It's just what word comes next, for example. And although a modern agentic AI isn't a model, it is hundreds of massive models of different kinds, reinforcement learning and transformers and deep neural networks doing, but tons and tons of different kinds of machine learning under the hood producing this global thing.

1:10:58But at its heart, they are all model free. There's no model of the world going on here beyond the patterns that is recognized. and its patterns although the emergence of multimodal large multimodal models as opposed to large language models is a thing and that it is improved and then reasoning on top of that i use finger quotes for that because i am it is not reasoning like what how human reasoning is But it is superhuman. I mean, vastly superhuman in its model-free cognition. It's amazing. I am truly not underselling agentic AI to say it is only a model-free learner. Because it turns out all of human knowledge, right and wrong, fed into this thing, just predicting the next word in the sentence or the next pixel in the image is truly amazing.

1:12:07I mean, truly and transformative. that little extra component that model-based learning gives us and even gives your dog is small yet transformative we build models of the world it is a different part of our neural architecture the real story is again even more complex because these things interact with each other and whatever we were not built by an engineer um who is trying to keep this all nice and and well partitioned um our code base is a fucking mess um but in this case we have this more than this but again let's keep the story simple this other ability that it lacks reasoning models are kind of like okay reasoning is really an effortful deeply prefrontal model-based process involving what's called explaining away oh because this happened then that can't happen um we've substituted a weird kind of model free learning reasoning where we say, first, write a sentence.

1:13:27Then another model is going to look at that sentence and write its own sentence about why it might be wrong and how to improve it. Then another model is going to look at that. So we've done this sort of model-free reasoning, which actually turns out to dramatically improve the outputs of these systems. It's still not model-based in the way that most natural cognition is um so that subtle difference and others produces this sort of interestingly unique human value i mean in a funny way you could say ai doesn't need to be resilient because that just doesn't yeah it's just spitting out words hey could to break it to you to anyone who thinks differently but these are not conscious they're not aware we can make it 10 trillion parameters We can make it 100 trillion parameters.

1:14:20There is no theoretical reason to believe that they're ever going to wake up without additional things being layered on top of them. And it turns out causality is super computational expensive. Go cross town to UCLA, talk to Judea Pearl about it. We should be including that in our models. We're not. So this is our little nerd tour here. without that component what you are getting out of these systems is patterns it doesn't know what's behind those patterns there's no real causality going on in its reasoning process it's turning its own patterns into new patterns again what is amazing is how far that gets you it's also interesting is how much of what we do is exactly the same how much of speech is also just the next word coming out.

1:15:14But again, not underselling it, it's amazing. So I can feed it a genuinely complex problem and it can give me profound insights if that problem is within its patterns. If it's not, it can be giving you interesting insights for you to work with. But really, this is where that human model-based, uncertainty-based reasoning starts to really come in. Because the models, you can actually prompt them to introspect about their own uncertainty. Easy enough, you'll get better results when you're near the edges of their capabilities. You can really strongly prompt them to ask you questions. Do you truly understand what I'm asking you to do.

1:16:05But the default is, as you're saying, keep it easy, easy breezy, shallow, quick, make you feel like it's you were the brilliant one. And it was, I don't know, the number of times it tells me, oh, Vivian, that was that's you just had the crucial breakthrough. Now let's package it up. And I have to fight against that all the time. That's the enemy of these interactions.

1:16:28Johnny:Well, and also energy conservation is a big part of this. I know you can choose the model that's longer thinking and it puts more energy. But ultimately, the model defaults to the prediction for the next word number in software, whatever you want to call it, not at I'm going to throw every ounce of compute I have and all the energy in the world at this. Like, what is the shortest distance between that? 100 percent. The next word. Even to get back to what you were talking about earlier every run on these machines could be curing cancer or it could be your virtual hookup or yeah you could generate some silly little like there's both there's energy and compute time and i think if we just thought about what it meant to run that query You know, how meaningful is this to you?

1:17:27Whose life is better because this happened? But, you know, my other big, bold statement here, it's just a riff off of everything I've said so far, is simply that AI that is optimized only for autonomy to be that minion that does what you tell it. it is just a dead end for humanity. I mean, I mean that on so many levels. Like we exist. We got to figure out something to do with us. I have this really cool finding that us plus machines together, not literally wired together, but I'm happy to talk about that too. Us plus machines together is better than either separately. So why don't we lean into that?

1:18:16heart across humanity again i get it's not going to be everyone tomorrow but that should be our trajectory that unique thing that we bring isn't some vague fuzzy um you know self-help-y book kind of statement it's a grounded reality we have a unique form of cognition and here's the here's the cool thing. I'm a scientist. I am proud of what humanity has discovered. We are both right. Even the things we have discovered are uncertain. We don't know fully what they mean. But when people say AI is a black box, it immediately makes me think, so surely you must then think our brains are a black box and therefore no one should be a neuroscientist.

1:19:10No, we understand so much of how neural networks and AI works. We understand so much of how brain works and how the weather works and about, you know, cosmology and quantum mechanics. It's just that in discovering these things, we've discovered there's even more we don't know. And that sucks because it makes you feel like you've learned less, like somehow you're in the negative knowledge space, but you're not. All of that stuff we have learned for good is now free in your pocket. Everything we don't know is what's left to us. And the wonderful thing about that is what we know is massive and it's hard one and I'm proud of it, but it's finite.

1:20:04Everything we don't know there will always be don't know to explore.

1:20:11Johnny:That gets to the heart of the curiosity. And to wrap here, I just want to touch on the concept of meta learning because as this is sped up, right, we are expected to gain expertise faster than ever. So for many of us, again, going back to, I went to undergrad, four years, studying subjects, devoting all of my time to those subjects, not having a job for most of us outside of that or maybe doing all the other things, but to become an expert, We spent a lot of time doing it. And it feels for those of us who are now in these roles that are collapsing with AI, we have to learn faster than ever. We have to become experts at a much rapider pace than we've ever experienced.

1:20:48Johnny:We don't have the time to our advantage. It doesn't feel like. So what can we do and what is this concept of meta learning? How can we bring that in as we wrap the show? Because I think that's a really big part of the equation. so you know i mentioned when i was the the chief scientist of this company i had this massive data set and my job was to predict how good you are at a job you've never held at the same time i had a separate company working education we were looking at what predicted long this was our insane idea predict kids long-term life outcomes then generate tonight hey dad mom do you have 15 minutes free, here's the best way to spend those 15 minutes to improve those long-term life outcomes.

1:21:31Not because we had a perfect crystal ball, but to really focus on these qualities that we could see. Turns out that Ivy League school you went to is only at best a modest predictor of your long-term outcomes, career and life, once I know these other things about you. And we've already run through a number of these constructs, numeracy, literacy, working memory span, attention, analogical reasoning, self-assessment, resilience, purpose, all of this stuff. What really bounds them together? There's a sort of term of art right now. Some people call them durable skills. I like foundational skills because a really cool paper found expertise Peace is not predictive of career outcomes without foundational skills.

1:22:24If you have them, going to university pays off. If you have them, knowing esoteric science or coding or what have you skills pays off. Without them, it doesn't. You want to become a plumber, you don't need foundation skills. This is the claim of the paper. So I like that foundation skill term. I've been calling them meta-learning in the sense of it's your ability to learn how to learn. In 30 years, I don't know what human beings will need to know. What will be AI? What will be human? Will we have run out of the energy budget to run these AI models? And now we need to go know new stuff. Why guess?

1:23:01It's as though you just asked me, hey, Dr. Ming, what stock should I put all of my money into right now today and then never touch for 30 years? That's insane. Build your kids for uncertainty. Let them figure out what they should know when they need to know it. The problem is not even to the whole length of our economy. We're obsessed with things like reskilling and upskilling. Oh, your business in real estate economics, remodeling is going out. We're just going to reskill you into doing financial analysis of the energy markets. It turns out actually, no, even highly educated people do not rescale easily.

1:23:48We have this idea like out of the German post-war economy, we'll just take all of these naval workers and we'll switch them to go work for Mercedes and Volkswagen because it worked in Germany in 1950. Well, it did. As I just said, the kind of skills we're talking about there didn't require foundational skills. the kind of skills you're talking about i i don't want to go build bridges i don't want to be doing those other um you know modern tennessee valley authority jobs they require metal they require you being able to change and that is things like meta uncertainty i bet you've never heard of that one before but it turns out each of us vary in our ability to assess our own uncertainty or even other people's uncertainty about a problem, for that matter, and AIs.

1:24:43That is a skill, if you will, or a human capacity that not only is measurable, along with resilience and purpose and all these other things, it's measurable. It is predictive and almost certainly causally related to positive life outcomes, all of them, from health to income to happiness. but it's changeable and again we kind of opened here change isn't easy that is not it's not a six-week job recruiting program that's why people like the reskilling idea because then we just take a bunch of coal miners we promise that we're going to retrain you to be software developers and then you don't have to worry about this whole globalization thing boy did we not keep that promise because it turns out Google has zero interest in hiring a bunch of coal miners to be software developers, not because they aren't good people, but because nothing in their life laid that foundation.

1:25:43Johnny:And those skills, I mean, we're seeing AI adopt them faster than humans could. Even if we upskilled in these areas, it's now the AI is outracing every human. For years, despite that this was the future of work, every future of work policy paper from the original World Economic Forum to every major government to every consultancy, every company. And they all said something about soft skills. Hey, these aren't soft skills. I can measure them. They are real and they are life changing. But also, they'd start talking about soft skills and then they'd say, teach everyone how to code or teach everyone about AI.

1:26:24those aren't actually valuable skills. So being able to learn how to learn quickly. And deploy. You know, here's the thing. I made this joke way at the top if you keep it in the cut of me saying math, what am I going to do with that? Only just sort of later become a hacky applied mathematician because that's what machine learning fundamentally is. um well the way i hack through it is not that i get out a a blank sheet of paper and a pencil because boy that's going to break me i crack open what mathematicians have been using long before lms came along i crack open mathematica it can solve any math equation any problem it can do the proofs, all the things.

1:27:14What it can't do is know why. What it can't do is find, essentially, I'm a math artist. I'm the one with the vision. I need an equation that's structured like this. I need some linear terms. These are nonlinear. It's going to be hard for me to model. What can I substitute in an approximation? Where's this going to be good enough? it's walking through step by step i'm orchestrating this process i've got i've got this paintbrush which is mathematica and i'm painting this mathematical picture that frankly i couldn't have done on my own yes foundationally i understand the math but i i never remember to carry the minus sign so it's just disaster in the end i can write lots of code and i have my life i love the writing code.

1:28:06So having clog code write a lot of code for me, I'm not a developer. I don't love that. I'm a computational scientist. What I'm doing is I am orchestrating this whole thing. But again, not the minions. I'm saying, make this bulletproof. Tell me why I'm wrong. The nemesis prompt I mentioned earlier. And the thing is, I'm going to say this bluntly, it's not a real person. doesn't care whether you're nice to it. By the way, if you say thank you or please, you are spending energy budget. Google doesn't actually want you to do that. Because they don't want to reply to that. They don't want to reply to that.

1:28:45So, I am prompting it. Tell me why I'm wrong. I have an idea for how to run this algorithm. Here are alternatives. Then I read through it. I don't like this section. Let's rework this. Let's try that. Again, in the end, you couldn't point at the result, whether it's code or book or math and say, this was Dr. Ming's part and this was the machine's part. It didn't do the boring stuff the way Sam Altman wants to pitch it to you. And I did the fun stuff. If you're not part of the boring stuff, then you don't understand it and you're not going to be able to do something creative and different. so uh go read those chapters the how to how to robot proof chapters every one of them has specific exercises to go through whether you're a leader whether it's about you frankly the parenting ones i just had someone say i think i'm going to use these on my employees too work through those also um there are tangible and i'll even end very concretely here i'm writing a newsletter right now i was working on it before i on the car on the way over here on curiosity one of those predictive factors for cyborgs and of course that's great but if curiosity if it's just in your genes what good does it do um so what i'm working on is research showing the curiosity is developed how do you do it in kids how do you do it in adults what's the actual neurological underpinnings.

1:30:20So we're putting together this newsletter. The free version will be out maybe

1:30:25Johnny:even this week, possibly next week. So you got me curious. What's the takeaway for those in the audience who want to develop curiosity? The takeaway is the cool one that I loved was about kids, but you can really reframe this. It was have them engage in question problems. I actually love this. I gave a keynote at Syed Business School at Oxford. Instead of having a business pitch competition, they had a problem pitch competition. I loved that. Just admit you don't know. So what they did with these kids is they ran a class. Like we said, this isn't short. We're in a class in which the kids were trained to ask questions.

1:31:07Don't give answers. Ask questions. Ask, ask, ask. Importantly, they were rewarded simple attention rewards. What an interesting question. How interesting you are. You don't compliment your kids and your employees on things that are under their control. Don't say you're so pretty or you're so smart. That's just an invariant trait. Oh, you really thought hard about that. And you notice this thing. Draw attention to what they did well. Award the curiosity. And the attention itself is a wonderful award, even in adults. So build an award. We were talking earlier. You need the nucleus accumbens to trigger that error signal.

1:31:50Interestingly, a question is its own error signal because what's the answer? And then you want it to drive the learning. Well, it's attached now to an award. The teacher was really interested in the thing that I asked. pretty soon the kids start attaching the exploration part of the process, the asking questions, the curiosity to the discovery of even better answers. And so simple things like that, but genuinely extended over time.

1:32:24Johnny:Thank you so much for joining us today. Where can our audience find out about the book and your newsletter? You can find the book on Barnes and Nobles and Amazon and all those other slowly, but hopefully not truly dying places where people buy books. It's genuinely funny. I would like to believe I took the Infinite Jest slash Discworld approach to footnotes. All the footnotes are. This episode is brought to you by Athletic Brewing Company. No matter how you do game day, on the couch, in the crowd, or manning the snack table, Athletic Brewing fits right in. With a full lineup of non-alcoholic beer styles, you can enjoy bold flavors all game long.

1:33:05No hangovers, no buzz, no subbing out for water in the second half. Stock the fridge for tip-off with a variety of non-alcoholic craft styles available at your local grocery store or online at athleticbrewing.com. Your beer, fit for all times. Our dirty jokes, but unfortunately with the dirty parts crossed out, Um, and, uh, if you want to go right to the heart or support local bookstores, you can go to my personal site, uh, Socos.org, S-O-C-O-S.org. You can get my free newsletter. If you're insane, you can get my paid newsletter, but that supports all my philanthropic work. So please do. And you can find the links to all sorts of local resources about the book.

1:33:48Thank you so much. It was a blast. Thank you. Thank you.