← Tom Bilyeu's Impact Theory
Tom Bilyeu's Impact Theory
AI: Savior or Destroyer? What the AI Revolution Means for Business, Healthcare, and Society | Emad Mostaque - PT 2
01:31:17
Loading summary
Transcript84 lines
- [00:00]
Commercial Announcer
Your next unforgettable experience can happen anytime. Take an AMEX card with you for rewards wherever you go. Morning coffee run with an old friend Earn cash back Weekend getaway Earn miles AMEX rewards your inner explorer. Learn more@americanexpress.com terms apply if you work in university maintenance.
- [00:17]
Grainger Advertiser
Grainger considers you an MVP because your playbook ensures your arena is always ready for tip off. And Grainger is your trusted partner, offering the products you need all in one place, from H Vac and plumbing supplies to lighting and more. And all delivered with plenty of time left on the clock so your team always gets the win. Call 1-800-GRAINGER visit grainger.com or just stop by Grainger for the ones who get it done.
- [00:45]
Tom Bilyeu
I'm Tom Bilyeu and this is Impact Theory. We're diving right back in to Part two with Emad Mostaq. How do we avoid Black Mirror? I want the benefits of AI, but I don't want to end up in this sort of emotional uncanny valley again. As a video game developer, I think about this a lot. I want to bring really rad things into my game like Elden Ring, where you really have a sense of I made progress, I accomplished something. This was fun. I get that it's just a game, but it was just tremendously enjoyable. But there's a phenomenal episode of Black Mirror where the woman's, I think fiance dies and she's able to get a robot with his AI in it. It looks like him, it talks like him, it sounds like him, but some part of her knows he's deeply comforting at first. Deeply comforting. But some part of her knows it's not him. How do we escape that? I can get what I want, which is my grandma back, my grandpa, my parents, lover that passed away, whatever, but it's not really them. And so there is some dark off key note that struck how do we have our cake and eat it too?
- [01:58]
Emad Mostaq
Imod, I think it's just against societal norms where it's a dangerous society. I think some level of regulation is needed. Again, the same with the gacha mechanics for video games, the Snoop Box mechanics and more like what you just described.
- [02:11]
Tom Bilyeu
So are you saying pass a law that doesn't let you bring your dead wife back in robot form?
- [02:18]
Emad Mostaq
No, because I think some people will want that. And again. But then the question is what's the permission system? Because they should give permission for something like that, right? These are things whereby legislation and society I think again needs to have just groups of talented people that can help inform and they can help adapt to that. And there's a commonality of this because the dead wife back then husband back is same across everyone in the world. Potentially. The cost of building an AI that can replicate. And again, it will be available on WhatsApp video. Anything to talk to you of a loved one who's about to pass away is probably going to be a hundred bucks next year. 100 bucks and 15 cents per million words they say and any language anywhere. Just need some of their history and then to chat a little bit. And I have the voice, I have everything. So this is moving so fast with things like that, which again, I don't know what, have a real emotional impact. You might find it discordant, but someone else might not. And so again, this is the thing where, you know, you want your libertarianism versus kind of other stuff. One of the most difficult things is again, we're very privileged. We grow up everywhere. We have and then. But we've also pushed through. We have our own viewpoints in the world. The reality is AI will be used differently by everyone around the world. And some people will use it to enslave, some people will use it to emancipate. Right? But it can't just be that there is only a few choices. You need to proliferate some of the choices. And again, we need to set some standards around this because these are really hard things like bringing back loved ones. I don't know how I feel about it, but I don't think it's something we should legislate against. But it's something we should definitely look into the harms of, because I can see the positive and the negative. And that applies to so many uses of this technology because it reflects us and it's actually come from our collective intelligence on the Internet, which again is kind of crazy. We need better data sets and it resonates, which is the crazy thing. Again, once you hear an AI whisper to you with full emotion, it's crazy. Once you see it, I think the latest video technology, when you agree it's broken through the uncanny valley now whereby it's incredible. You and I have trained eyes. I would say, until a couple of months ago, I'd be like, I see something instantly, that's AI. Now I'm like, Then I go and look and I can see some subtle things. But just again, my fast thinking is like, I can't tell anymore. So that's why I think we need to standards around it. We need to adapt, we need to proliferate those. But then we Also need to again, allow for liberty, allow for use as long as there's an infringe upon the rights of others. And that's really hard. That's why we need to distribute this intelligence. We need an open, distributed, intelligent AI system and then we'll have the proprietary ones and all these other ones too. But it needs to exist with the best standards we can have. So you can say what is the best way to teach calculus? Let me tell it all is another question. Right? But that might not be the way you teach it because of your local curriculum, but eventually we want to move our infrastructure of thought towards an optimized thing. We want to build a society OS and reconstruct it. And this is the one chance we get at that. Right? And my view is that countries and communities that embrace this and take the best of our generalized common knowledge and these systems will outperform those that don't. And so again, if you set good standards, that's huge potential for a incredibly positive view of humanity where we can solve and coordinate for most of the problems that can adapt to all the bad stuff. Again, this OODA loop, because we don't know what we don't know might emerge. Like it may be that there's an AI that figured out how to hypnotize people. What will you do when that emerges and it's controlled? It can be a foreign adversary, it can be Facebook. How are we going to react to that? Like literally hypnosis kind of works, right?
- [06:36]
Tom Bilyeu
Welcome to my nightmare.
- [06:38]
Emad Mostaq
But again these are, we don't, we need to build systems that are more robust, self adapting and I think we can only do that with AI.
- [06:44]
Tom Bilyeu
What would you give them as your best argument for why AI is going to be awesome?
- [06:52]
Emad Mostaq
I think it depends on your framing. But the most powerful thing I've basically thought of is AI is like those really talented graduate assistants that we wish we had but we can never be sure of. And what does that mean? It means we can do things we can never do before. So the very practical thing is we're all going to die, right? And it's probably going to be through some sickness. All of us know someone in our lives who've had cancer, Alzheimer's, dementia, one of these conditions. You will have an AI in a year or two that gives you comprehensive, authoritative, up to date knowledge on these conditions through your journey and outperforms human doctors on empathy and is available in every language from the medical side to how to talk to your family, mental health and more over the coming years. That's something that could never be done before and that will reduce human suffering. Because the loss of agency you feel when you get one of those diagnosis is insane, right? And again, it's difficult to find the support you need. That's just one small example. And I think that once we have that right framing, we can think of many more examples where you just wish. I wish I had some experts around me. I wish I had support. I wish I had help in the digital and physical realm.
- [08:06]
Tom Bilyeu
So I just wrote down small. If that's a small example, which already sounds pretty insane, what are some of the ways that you think society will be positively impacted? As you think about those ripples away from just the individual, what ends up happening to society?
- [08:25]
Emad Mostaq
I think that the infrastructure of our society is based on texts and reports and bureaucracy. If you think about something like the US elections are coming up, it'd be unpleasant time, right? Having the ability to embed artificial intelligence in the government, to check every single policy and then make it relevant to you and explain it comprehensively will be awesome for understanding candidate positions, but then also for going to the DMV and getting your driver's license. You know, that will be automated in some number of years, may take longer, may take sooner, depending on our capabilities. But every single one of the things that you find frustration about, where there's this beach to report framework, where there's this idea, the image video, the barriers are being reduced. For all of those, the information will flow more freely. And again, think about most of your frustrations in life. It's due to lack of agency, it's due to frustration over systems of information, getting stuff done or access to information. All of those barriers will be brought down by this technology, having that role of again, these talented grads everywhere. There's no part of information society that doesn't impact. So hopefully it'll help us reduce frustration, increase flow, increase our fun quotient as well, and maybe make things more functional.
- [09:50]
Tom Bilyeu
So that is, when you talk about a grad student, I think that is something that's really clear for people that are in the elite bucket. But for the average person, how's this going to impact them? Because I think the average person is panicking because they think about their job and they think this is going to be automated away and it really might be. And so if you were talking to somebody who you say, oh, it's like having a grad student in your pocket, they're like, I don't even know what that is. I don't care about grad students, why is their life going to be better?
- [10:22]
Emad Mostaq
Oh, because they can do more. You know, this is the thing whereby you've always had technology like spreadsheets came and it changed accounting and other things, but you know, it created brand new jobs. The main concern that we've had, we discussed this kind of last year, I think, you know, was can you create the new jobs quick enough? But the reality is anyone listening to this can do them more, more effectively using this technology. Even if it's in the report writing phase. Everyone has to write something or summarizing emails. This technology started to seep in. Siri will actually be useful, right? Like again, if it's not a grad student, it's a hire, you know, a junior hire. You want someone to help you and bringing that hire in will be impactful and they'll do what it says on the tin. All these issues like hallucinations where the AI makes something up, you know, all these issues about AI going wonky, they're being slowly ironed out and now you get these really reliable partners that can work with you. That's on the productivity side. Then there's the other side, which is the creativity side. You use Suno use udo, you use midjourney. Now you use grox image generation, but you could never generate images like you can now. And now video is coming, you can make music. And it's one thing making that instant, just playing around with prompts. Once you actually start to apply yourself, just like anything, your creative capabilities have gone dramatically up. Now what does that mean? It means that as you're doing your job, the question is, if I had helpers, how could I do my job better? Then you'll outcompete those that do not embrace this technology because the pace of technological adoption is always slower than you'd expect until it isn't. And right now we're in this building phase where everyone's trying this technology, but the implementation will come in a few years, you know, but then it'll come as a wave. So if you've got 1, 2 years experience applying this technology to whatever your job is, you know, to make it more efficient, to make yourself more productive, then you'll do better.
- [12:19]
Tom Bilyeu
Okay, so let me give you what I see is the, the big problems that we're facing as a society today and let me know if AI helps and if so, how. So right now I think people really feel the sting of inflation. Prices skyrocketed during COVID they have not come back down. People do not feel good at the grocery store, a Zoomer has 10% of the buying power that a boomer had at the exact same age. In their 20s. There's a just mental health crisis in young people. Parents, racism, rightly are freaking out. They don't know what to do about it. And I think a lot of people look at AI and go, oh my God. Like, this is just the, the TikTok algorithm come to life. It's going to suck my child into a black hole. This is not good, emod. These are all bad things. What do you say to that basket of I just want to be able to pay for things. I just want my kids to be okay. How's AI going to help there? Or is it not?
- [13:19]
Emad Mostaq
Well, I think AI is a very broad thing. So TikTok AI is different to generative, AI is different to Facebook, AI in some ways different to Siri. It could be an incredibly negative future. These things are dangerous. You see a little hijack the minds of a generation and now it can be amped. But again, like I said, on the flip side, there's the positive side. How could we use this positively? So you look at things like inflation. Inflation is a monetary phenomenon here whereby we're suffering the overhang called the stimulus spending during COVID which nobody knew how much we had to spend. We didn't know if any people would die. And now it's the overhang from that. So inflation started to come down a bit, but you said buying powers decreased in general because of this. When you look at the components of inflation, food and other things, I think that. And you look at AI and you think about massive numbers of graduates, digital graduates, entering the workforce, it's probably going to be deflationary. But then the question is, can you have the stuff you need for living? That's your hygiene factor, the stuff you need not to bother. And in fact, I think I can't remember who said it. Someone said there's only two types of things, things that are living for living and entertainment. So we've got to get to that basic level first. And having more efficient systems to organize things, to remove bureaucracy, to remove this overhead should help with that. It should be deflationary, taking apart bureaucracies, doing these things. There's a question of jobs, a question of, let's say, the American dream. Like, what is the American dream? If you strive hard, you work hard, you can be successful. I think part of this crisis is given all the knocks that we've seen, as you said, parents thinking about their kids. If my kid tries Hard can they be successful, particularly when the AIs are coming? I said at the start of last year, in five years there are no more programmers. I should have added as we know it, just like we grew up with the Flash programmers. Flash was programming and other things. The nature will change. Like if you use Claude from Anthropic. Right now you can make games, basic games like Snake and others, just by typing and telling it how to edit it. And it will do it live right there. It's more proficient than most graduate programmers, but that means the level of graduate programming has to increase. So I think that when we look at that, this is the key worry is can we create the jobs in time or will there be enough of a demand response from the increased amounts of entertainment and productivity to allow these kids to have careers? And what will those look like? That's a really complicated question. When I think about things like housing inflation, others and monetary things, I think that it's deflationary and the cost of living right and the cost of organizing right will drop. Again. These are very basic things, like the fat in the US system. How much does LA or San Francisco spend on timelessness? It's over $100,000 per person per year. AI can easily deconstruct all that spending and route it more appropriately. When will that happen? I know, five, ten years out. But then you apply that to almost every part of living and you can see how it can make it more efficient and more accessible. But we've got these competing factors here whereby costs will come down, but will demand catch up? And this opens up that question of do we need ubi? Do we need universal basic jobs? Do we need more stimulus? It all depends on the pace of adoption of this technology and how we deploy it.
- [16:43]
Commercial Announcer
You know what they say. Early bird gets the ultimate vacation home. Book early and save over $120 with VRBO because early gets you closer to the action, whether it's waves lapping at the shore or snoozing in a hammock that overlooks, well, whatever you want it to. So you can all enjoy the payoff come summer with Verpo's early booking deals. Rise and shine. Average savings, $141.
- [17:11]
Grainger Advertiser
Select homes only if you work in university maintenance. Grainger considers you an MVP because your playbook ensures your arena is always ready for tip off. And Grainger is your trusted partner, offering the products you need all in one place, from H Vac and plumbing supplies to lighting more. And all delivered with plenty of time left on the clock. So your Team always gets the win. Call 1-800-GRAINGER visit grainger.com or just stop by Grainger for the ones who get it done.
- [17:43]
Tom Bilyeu
Okay, so we'll get to the downside of it first, but give me the positive look at this deployed. Well, what does that look like? What would we need to do? So integrating into the government, that seems really smart. Not only so that I can or everybody can understand. These are the candidates, these are the positions. You know that bill that they dropped last night that you know is 1900 pages and it has to be voted on tomorrow. Here's the summation, here are the key points, here are the things they're trying to sneak in. Super helpful. Transparency on budgeting, super helpful. But what are other ways that you can see where if we integrate this into society, well, that it drives costs down where people care without triggering that existential crisis of well, now what do I do as a human?
- [18:33]
Emad Mostaq
So I think if we look at regulated sectors, these are two very important ones. Let's start with two things. Education and healthcare. After the recent, I think Trump musk debate discussion, I think they said why even have a department of education in the US if educational outcomes are what they are? Like kids graduating from school not being literate is insane. Yeah, right. And what is school? It's childcare mixed with a social status game mixed with a petri dish in many cases. Right. The only thing that's been proven to work in education is called the Bloom Effect. Two standard deviation improvement outcomes is one to one tuition. Sometime in the near future, every single child should have their own personalized AI tutorial that can tell. Are they auditory, are they visual learners, do they have dyslexia and adapt the curriculum to them and encourage them to engage with others and increase their agency and basically help them on the educational path. Again, one to one tuition has a two standard deviation improvement. It's the only thing that's proven to work in education really. So something like that could transform the education sector from an information going back and forth with the kids and again bringing kids together to do stuff. Now the question is who's going to build it and who's going to build it right and implement it? We'll come back to later. Healthcare is another one. The average US life expectancy has been dropping, but everyone else has been increasing. How insane is that? The US spends more than the UK as a percentage of GDP on public health care, yet you will have to have private health care as well. And UK healthcare isn't bad. It isn't great but again, you don't have that health that you deserve. And ultimately it's your health that's going to get you. Your health impacts every part of your life and ultimately we all will die. What if we could reduce many of these things? And that again is your personal healthcare buddy. But also, I think every medical diagnosis has to be checked by AI before the diagnosis is made by AI. It'll be malpractice not to have that. And how many lives could be saved. We all know people that have had botched diagnosis, botched this, botched that. You won't have that with AI. AI outperforms human doctors on diagnosis as well as, like I said, empathy. So when you kind of look at all the sectors that are regulated, they're usually regulated because they have a real human impact. And most of the issue is information flow and personalization. The teacher cannot teach 30 kids at once because their attention is split and they're misaligned incentives. And the thing is, we all know people who start and want to become teachers and doctors and in the government, most of them start with ideals and they have the need to mount out of them by the system. Right? They can't reach everyone. We see this again and again and again. This technology can change because you can have personalized information for every single individual relevant to them and you can coordinate information as well. Our collective knowledge. So this is where we should have all the council knowledge in the world comprehensively, authoritatively and up to date available to anyone at any time as a public good, for example. Again, that just needs to be built. So I'd say those are the types of areas, the regulated industries where you can have a transformation. Government, education, healthcare and more. And that's again what you need for living. Then there's the other side, which is entertainment, which is we all know creatives and how much does it cost to make a movie? Or it's getting more and smaller. How much does it cost to make a video game? How much does it cost to tell your story? You know, like TikTok was one of the first indications where you know, you're short, doing short clips, right? But we both, I think agree in a few years you'll be able to make Hollywood length movies because movies are just made of two and a half second shots. Now it might not just be type of prompts. I don't think that's what we're talking about. We're talking about I want to tell a story, I want the control over that we have the generative capability that we can do that. Be it music, be it movies, be it more. And again, it can transform because you're no longer constrained to the medium. Like, why should songs be three minutes long? What is the appropriate way to tell a story? Is it snippets? Is it large movie? You can play around with it. You can take this podcast and Translate it into 100 languages now and it'll have your voice and my voice. But I think these are some of the things I've been thinking about. Like, again, that's just having a bigger team around you, a team that's on your side when it comes to regulated industry and increased transparency. Breaking down the barriers. And then also on the creative side, it's building the tools that remove the barriers to people telling their stories.
- [23:08]
Tom Bilyeu
All right, I'm going to look at everything through an optimistic lens for now, like I said, again, we'll get to the darker sides of this. I'm going to paint a picture. This is what I actually think is going to happen. But if you spot any naive elements in this, please point them out. Because I really do want an accurate vision of the future more than just an exciting one. But when I look out at the world, there are a few things that I see. Number one is the health crisis. You've already talked about that I'm going to have something in my pocket that's going to help diagnose, or if I've already been diagnosed by a standard physician that is keeping me abreast of all the research that's coming out that is unique or that applies to me. It's going to know my DNA, it's going to know my microbiome. So it will have a better understanding of how someone makeup is going to metabolize a given drug. And so is there a trial or something that I should be going after? I think that would be huge. I think the mental health crisis is absolutely massive. And I think being able to help with that, I think AI is going to be tremendously advantageous because here's how I see it working. You're going to be wearing devices that are taking a, an objective snapshot of your real time physiology. So where's your heart rate? It will detect a spike in anxiety. How well are you sleep? What's your blood oxygen level? All things that we can read already, but it's just so much data, it becomes very difficult to correlate to a thing I just did or something that I saw or encountered or what I ate, which would be a huge part of this. And it's tracking this just incredibly multivariate thing and saying, oh, right now you just encountered this thing, it's triggering a stress response, you need to relax or hey, you've been relaxing too much, you need to push yourself. Given the patterns that we've seen on the days you're more active are the days where your anxiety is lower, your depression is lower, whatever. And use the biofeedback as a way to get out of sort of the, you know, talk to me about it. Because I think that, talk to me about it. There's some question marks around how helpful that is depending on how your therapist is running the session. So therapy is one of those things I think can go very right and I think it can go very wrong. So if we have, if we introduce the biological data into that feedback loop, I think, I think things get a lot better. And then education again, you've already talked us through that. But that one to one curriculum I think is incredibly important. And your AI being being able to flag for you where you're at. Are you ahead, are you behind? You've told it the things that you want to optimize for in your life. And it's like, well, you say you want to optimize for that, but you're doing these things, they don't really align with your goals. What if you took this course, whatever, and then the last one is, and this might be the most important is all throughout history, every time that you get into a very concentrated urban area, which you can read as a proxy
- [26:13]
Emad Mostaq
for
- [26:15]
Tom Bilyeu
safety, stability in society, the whole strong men make good times, good times make weak men. That whole cycle population declines. That's not only a modern phenomena, that's happened over and over and over. And when you look at the different population collapses, they've always been tied to a precipitous drop in birth rate which obviously we're seeing now just rapid beyond all measure, especially in South Korea, Japan, America's not doing great, not doing it as badly, but you're seeing it in, in a lot of places now. My hope with the population collapse is that one just that AI can do all the things with health read into goals in terms of helping steer people's behavior so that they want to have children, that they can be on that track, that the AI can help them not be 36 and suddenly realize they want to do it, but be advising them much earlier but possibly more important, and I'd love to get your take on this very specifically is robotics. And that as we get this upside down pyramid of a ton of old people and not nearly enough young people to support them. That robots and. Or embodied AI, if you will, which is another way of saying robots will be able to step in and take care of that burden so that not all bright minds of the young generation are sucked into elder care. What do you think about that? Is that any of that Pollyanna or does that.
- [27:48]
Emad Mostaq
I think that reasonable. Like I think the mental health one is a particularly interesting one because there's your biomarkers, but it's also your voice. You know, you can tell a lot from a person by the way they're speaking and just picking up on that. And again, that's what a good therapist can do. The AI can do that even better. Now we've got. We've seen some research that'll be coming out soon about that being able to pick up the timbre and also respond with the correct voice. Like, yes, there's a bit of a black mirror ish thing I've been thinking about recently whereby this is the BBC show about dystopian views of the future. I'm not sure there's a good thing or bad thing. I'm not sure we discussed it before. Given a few hours of discussion with someone, I can recreate a perfect loan of them in terms of voice reaction. You can WhatsApp them, you can call them, you can even have zooms with them now live and you won't be able to tell. That sounds creepy, but what if it is a loved one that's about to pass away? Your grandma, Someone that you trust, you've always looked out for you. Then imagine that person being with you throughout and supporting you. Now that I see. Horrifying and slightly comforting, but it gives you an idea of kind of the stuff that is now possible because like I said, you can zoom with them, you can text them, you can call them. Maybe one day there'll be a robot. Is that positive for mental health or negative? It can be again, be both ways. This is a dual technology. Right? But I think it reflects the fact that the AIs that are coming. There's a classical big data and we have this measured life with our Fitbits and all of that. Right. And a lot of these healthcare things are like dashboard based. I can measure this and some of our brains are set for that. But then there's nothing like someone next to us telling us you're doing okay or you know, shape up. And we will put so much trust in these AIs that are next to us because we have to remember as well, you know, this isn't a knock like when you look at the distribution of intelligence with intelligence is coming into the community now, right? We were building artificial intelligences. Half of America, half of the world's under 100 on IQ. That's just the way it is because it's a normal distribution, right? Yeah. 95% of people think they're above average IQ. And in places where there's malnutrition, it can be as low as 80 on average for nations. US states go from 95 to 105 on average. But if you think about that, many of us technologists, we build for, like people who've had very privileged lives and we live in bubbles, right? But if you think about 10, 20, 30 years from now, and many of these things, it's like people didn't have the opportunity, didn't have the chance. They've had malnutrition, kind of bad backgrounds, or they're just flooded with junk they don't believe all of a sudden, like, I think anthropic did an IQ test on Claude. It hit 100. But then it will also again, be able to speak to you in the most comforting voice in the world or tell you that you need to shape up or these other things. And so again, when I think of mental health, I see this range of things from recreating your grandma before they pass away so you can always get their wisdom to, I don't know, like Aristotle. You're just your own personality that grows with you. And I don't think we've ever. We've never even been able to explore or conceive that technology. Right. But again, I know if someone I trust tells me to do something, I want to work hard for them. And that's the most difficult thing. Again, not these dashboards and these big data kind of things. It works for certain quantified minds, but nothing kind of beats that kind word, you know, at the right time. So this is why I think it'll be huge for mental health, for kind of feedback loops. It could also be incredibly dangerous, you know, but again, we're doing the positive side of things. There aren't enough people to support people. That's why, like I said, something like a loved one had a diagnosis of pancreatic cancer, and so for a week I thought she was going to die. A thousand A.I. just analyze everything. Pancreatic cancer is a 5% survival rate, right? And it turned out to be a misdiagnosis after we did all the analysis. But think about everyone who gets diagnosed about the loss of agency. They feel there's not enough humans in the world to support appropriately. But it should be a combination of the AI's helping and bringing the right humans to support you as well. Because again, nothing beats real world contact until you have robots. I think, again, the embodied AIs. The pace of robotics has leapfrogged in the last year or two.
- [32:28]
Commercial Announcer
When you manage procurement for multiple facilities, every order matters. But when it's for a hospital system, they matter even more. Grainger gets it and knows there's no time for managing multiple suppliers and no room for shipping delays. That's why Grainger offers millions of products in fast, dependable delivery. So you can keep your facility stocked, safe and running smoothly. Call 1-800-GRAINGER Click grainger.com or just stop by Grainger for the ones who get it done.
- [32:59]
Emad Mostaq
There's a company in China called Unitree. They have a robot called the G1. Have you seen that one? No. So the G1 is 5 foot 2, 34 kilograms and can run at 4 miles per hour. It can make you an omelet. And it's got reinforcement learning. So if one robot learns kung fu, the other robots can learn kung fu as well. And it's $16,000. So you can look it up. Unitry G1. Elon Musk reckons the Tesla Optimus will be 10 to $20,000. I reckon it will be as well. And so right now, they look a bit weird, but again, they'll look more and more human. They'll be able to speak human because I think you don't want to get this uncanny valley thing where you've got this plastic face. But again, we respond a lot to voice. You look at AI voices now, they're perfect, right? But you can't tell anymore with 11 labs and hey gen and other things like that. Again, you can have your voice in 20, 30, 40 languages now, your voice, but in Spanish. So when I look at that, I'm like, it's not this Blade Runner future where it's kind of fleshbots and everything. I just think, again, robots will be there alongside us. But there is a difference. Again, with this embodiedness, just like we like our dogs and the kind of things like that. But all of a sudden they've got a level of engagement, Is it sentience? Is AGI? No, but again, they're there and they can communicate with you. They can be your barista. They can do that. They can do that. And you said this kind of fixes the population pyramid. And it's Very interesting that the place that will make the roast robots is China. Right?
- [34:35]
Tom Bilyeu
Which why do you think they'll make the most?
- [34:38]
Emad Mostaq
27% of the Chinese economy is industry. Apart from they're the biggest industrial producer in the world. Where do most of the cars in the world come from? Right. There's 80 million cars and 70 million motorcycles built a year. In a few years, I'm pretty sure the number of robots will get up to 100 million a year. Again, it's $10,000, $20,000. Let's put that in another context. When you look at the cost of you can look at like buying a car. If you buy a $10,000 car, how much does it cost to lease a month? Like a hundred bucks. For a hundred bucks you will have a 5 foot 6 robot that can do that, can make you omelets or look at a recipe and then make it instantly. That can learn from other robots that has fine grained control. But again, when you look at the unitary G1, it's got a Coke bottle, it flips its finger and the Coke bottle app goes off. You know, that's the level of control we've got now. So these will be useful. And we're seeing studies whereby, you know, robots and personal assistants help the elderly. Like how much does the US spend on elder care versus childcare? A lot more already. And as you said, this is coming. We want to increase the population. I think we've seen the studies around lack of testosterone, we've seen the studies around the hygiene factors of people not wanting to have kids. They can't be sure about the future we want to bring them. But it's positive to have this. But we need help. Yeah. And so if we can't find the human help, we need to have digital help. As we get digital help that can talk to us on our level and understand us better than anyone. And that's why we need to make sure that's aligned. It's physical help, the embodied help around you that can help you physically with doing all the tasks you might need, but then also going beyond that. And again, there's a shocking thing around the cost of all of this. Right? Like it's $8 a month for Grok on X and that includes a subscription to make as many images as you want and edit them. And an AI that's like OpenAI level. Who would have thought it'd be that cheap? You know, we're talking a hundred dollars a month the lease. A robot that can maybe do 95% of our human Condition. Who would have thought it'd be that cheap? And so again, it opens up this massive world of support whereby you can have as much support as you need because you're not limited. Like if I want to go and hire that personal trainer, it costs how much? Right? Human personal trainer will still be good, but then the availability of therapists, personal trainers, waiters, other things goes up dramatically. And again, there are negatives to that and there needs to be an economic response. But for many things it is also a positive when you think again about, I think, the hygiene factors of humanity, the living factors, and then you've got the flourishing factors right, if we can get the economy right, if we can get meaning right, and some of these other things.
- [37:43]
Tom Bilyeu
So how are entrepreneurs going to use this? I know a lot of young people, beginning entrepreneurs, they're going to want to know how they leverage their advantage. Is there an opportunity for wealth creation? Is this just going to come at all, at once that there'll be no competitive advantage? Or is there a window here where people can take advantage of this?
- [38:05]
Emad Mostaq
So, you know, I started out as a programmer 23 years ago. I was writing directly assembler code, so directly almost to the hardware. Then programming became about these bundle libraries of like this is how you make an iPhone, app template, things like that. You built up levels from that. Still, when you look at the banking system, it runs on COBOL and Fortran, which I believe were made in the 1960s, stuff takes time to upgrade. People don't want to include what are basically research models because this is still research, not engineering. A lot of AI, cutting edge stuff. You don't want cutting edge for your education, your healthcare, your critical systems. You want stuff that's tried and tested. So like I said, this will be both longer and quicker than we expect quicker because we're already seeing it appear and it just seamlessly goes in and breaks down these barriers longer because it's not going to go everywhere in society all at once. Because again, you can't take the risk. You can't blame the AI, right? But that transformation, we remember the.com transformation, we remember many of these transformations. Just embracing this technology, using this technology and helping people with that is an amazing business. Doesn't matter which sector you're in. You know, again the mental model is, you know, we wrote this piece how to think about AI, where we go through some of these. We were like, think about it like you've discovered this new continent of graduates that are specializing and how are you going to deploy that to your personal Life, your company, country, whatever, but your company, because it's so hard to find good people. Think about things in terms of flows, not files. So right now we're outputting image or a text, but eventually multi step, like that whole process of writing a report and getting all the information in multiple steps will all be captured that flow or in images and video. We have this system called ComfyUI that we built at Stability AI, which I left a little while ago. I form a company whereby when you input a prompt, it captures it. But then every single edit to that prompt that you do, and every single additional model you use for upscaling, and then if I send you that file, it reconstructs the whole flow. So files to flows and the final kind of point we made. And again, there's a lot more to it. You can check it out on our Twitter substack. It's just you're moving from thinking about the world in terms of these autonomous agents that will just get rid of all human endeavor to how can we make people more agentic? So if you're an entrepreneur, the best thing to do is like, how can I help my customer? Well, all businesses are the cost is less than the value created. So it's all about how can I help my customer achieve more? How could I do it myself? We're already seeing many of the startups that I'm seeing and small companies, they don't need to hire as many people for content creation as they used to before. They're using some basic agent based systems, they're using basic AI to be able to iterate and develop faster. Because I think there is a increase in reach. So again, systems like hey Gen, like 11 Labs, any podcast you do is you can make in how many languages you want, right? That's an instant increase in reach. Who knows, maybe this will be popular in India or China or wherever, right? There's an increase in velocity. So ideation, you know, again, for entrepreneurs, that's incredibly powerful because you can give it something like a Claude, a set of principles like, this is my idea for a business. Can you analyze it according to Porter's five forces? Or you know, hey, you're a Stanford MBA or Harvard mba, can you break it down for me? And then you can interrogate and it'll just go back at you. So the ideation phase is important. What could go wrong with this idea? I'm having a problem, like again, just having a buddy that kind of gets you. And then there's a cost decrease side things that you classically employed, had to find people for which is the most hard thing when you're an entrepreneur. Because unproven, you can offload a lot of that to the AI or to people that use AI, which I think is more important. And so, you know, that helps have more generalists, I think. And again, time to market, market size, access and then quality of service.
- [42:16]
Tom Bilyeu
All right, walk me through how this is actually going to happen. So when I think about these potential employees, for somebody that wants to use AI, let's say you're going to staff your company up with eight to 10 people, you know, all of them being AI, but you're going to want them theoretically to specialize. So will it be just one AI is capable of going in any direction. So you just take off the shelf nine or ten generic AIs and then you say you, you're in charge of this or is it one AI? And you just say, okay, do this task in this way, do this task in this way, so on and so forth.
- [42:53]
Emad Mostaq
So we built the first models for creation, you know, and again, we've got to human level in these narrow fields now. We're creating these flows and design patterns of jobs. So you've got your liberal arts grad, now they're specializing. So off the shelf. Now you look at something like Intercom, right, which many of us have used for the chat responses, you know, the help side, that thing on the bottom right that you click on and then you can chat. Intercom have changed their business model, introducing AI now. So they have an AI agent that looks at all your previous answers and you actually only pay Intercom for every resolved query. It's not even a subscription service anymore. I think it's just fascinating. But you're seeing call center worker AIs come now that again can be your entire customer support. But when I say entire, I don't actually mean entire, because a much better system is that you need to have a Manager for the AIs. We haven't developed the AI manager yet. So one person managing all these AIs and again, I think they're getting increasingly customized is quite a lot. But it doesn't mean that you can't have someone that you say, solve the problem of customer service for my product and then you don't need to hire all these customer service reps and you've actually got better outcomes according to the data that's been coming out already. You know, you've got your sales outreach and your sales calls and that can be massively automated. Now services are Popping out for that, including engaging. You don't know it's an AI effectively, whereas previously you had how many sales reps for a lot of these things, Advertising, SEO, design, you can expect a lot more from your designers. Now there's a minimum level of quality that's going up. But again, the designer is probably going to be managing a series of AIs. So I think that we haven't moved, we're moving up the level now from these generalized models, these more specialized services. But if you still need to have your managers, I think managing these services and again, a human in the loop, just to check, because they're not quite good enough, they'll break through to good enough. But again, the most difficult thing as an entrepreneur, as someone trying to make a difference, is hiring the right people to support you in these functions as you yourself move from a generalist to specialist, building a team. And these AIs and these systems have flaws and issues which I think you still need human loop, but they're improving fast. But again, I think that's the mental model that you need to have when approaching this. What are the functional components of my business and which of these services am I introduced to have cheaper cost, better experience, faster iteration? Right. And that differs area to area.
- [45:19]
Tom Bilyeu
If I think about this as an entrepreneur, the thing that immediately jumps to mind is if these are all pre made, I'm not going to have an advantage. You're going to help me lower my cost of doing a thing that I need to do. So take customer service. If I'm the only one with a customer service AI and it delivers better results, I'm in a great position. If we all have customer service AI now, it's neutral. Again, unless I have control over the knobs that allow me to make my AI customer service better. So take culture. I can inject culture into my AI and have a competitive advantage over my competitors just because they don't. For instance, back at Quest, we used to say, if you called us up and said, hey, I really want to get in shape, what should I eat? We would say chicken breast and broccoli. Now that was like, at the time, everyone just thought that was crazy. But it was this, I won't say viral thing. But it definitely caught on. People were talking about it because we didn't promote our own product. We told people, look, we know where we fit in the dietary ecosystem. You're better off eating whole food whenever you can. But if you want to snack, then by all means have one of our bars. And so that Mentality was an advantage, but if I'm stuck with an agent that just handles things in sort of a generic way, I'm trapped. How, when will that level of control be available where I can take something off the shelf so I don't have to create it, but I can dial the knobs in a way that's unique only to me?
- [46:50]
Emad Mostaq
It's available right now. Again, these are, like I said, liberal arts grads, shall we say. You know, they have a wide range, but there's something called the context window. So that's the prompt that you type in. Most people will only type in like 50, 100 words. OpenAI's ChatGPT when it first came out could take 2,000 words. Google's latest models can take 2 million words or videos or audio all at once. So you can inject entire training courses and your company bible and books that you've given about your company culture and like sessions. And it will, at inference time, the time when it runs and responds, assimilate all of that. And that's crazy, right? Like, again, you created a culture guide for your company. You can put that in the prompt. You don't need to fine tune a model. Even so, the fine tuning is when you put it in the model permanently. And again, like, to give you an idea now where the costs have gone. They've dropped like probably 500 times. The latest Gemini Flash model is I think 15 cents for a million words of output. And fine tuning is free. Previously it was like $150. Whoa. And again, you have your entire culture handbook in there. So my customer service, again, it depends. Is this a distributional thing or is it a cost? Right. And so business are repeatable processes. And like, were you the first energy bar in the market? You know, like, I buy bottled water. There are many intangibles around brand around classical products that once you really understand where your cost, your distribution base is and these other things, that's where you apply the AI. And again, where is your edge? Your edge is ultimately in your engagement with the people that think you're giving value and are paying for that, whatever that might be. You know, like Clayton Christensen, I departed Harvard physical professor, came with disruptive innovation, had this jobs to be done model where he said, any product is part of a job to be done. The McDonald's milkshake in the morning is quite thick because you're drinking it on the way to work in your car in America. Whereas in the afternoon they make it thin and runny because your kid's drinking it, and you don't want them to be sucking on it for too long. But he said that there's a functional, a social and emotional component to any of these. So, again, as we look at the classical business theory and what's worked, it hasn't changed. It's just that we've got a new workforce and the workforce listens to instructions. But how many people on this call have realized, sorry, the podcast, listening to it have realized that you can give 30,000 words of instruction to a ChatGPT, that you should treat it just like a new hire. You can list out all the instructions and it'll follow them. You know, instead they just type in little bits and they're just like, oh, why isn't it consistent in its response right
- [50:03]
Tom Bilyeu
now? If you give it the consistent context, will it give you the same output every time?
- [50:08]
Emad Mostaq
You can make it give you exactly the same output every time. So OpenAI just released this feature called JSON output, which is structured data, 100% reliable output. For example, again, that used to be 60% at the start of the year
- [50:26]
Tom Bilyeu
was very inconsistent the beginning. Same with the image creation. Okay, so we've got a lot of cool things going on at the AI level. A lot of it feels like it's not quite here, but if I lean into things that aren't quite here but that seem like they're coming around the corner. Elon Musk with what he's doing with Neuralink, I believe they already have two human patients. When I look at the future of AI, it seems inevitable to me that to keep up with AI, especially as it becomes embodied, we are going to have a choice, and I think to not end up in a position where you have something that just so outstrips your intelligence, your ability to communicate, your ability to perceive the world, I mean, just absolutely get left in the dust, no longer the dominant species on the planet. You're going to have to consider augmentation. What do you think about what's going on at Neuralink? Could you see yourself in the future embracing it?
- [51:23]
Emad Mostaq
Our phones are on Neuralink at the moment. So the whole concept of neuralink was that, you know, we're having a discussion right now, we're exchanging information, but it's like that much information. What if I could jack myself in and just almost communicate seamlessly all of the context at all times, you would have better input and output. And again, you said that's how you can compete with these AIs, because an AI that can accept a thousand words of instruction is kind of human, right? An AI that can accept 2 million words of instruction, that's something we've never seen before because you look at the thing, it's got like 99% accuracy as well. And again, to give you some context on this, you can upload an entire season of a TV show and ask it to pull out the funniest parts where your humor is defined by Seinfeld relevant to that. And it will do that in one go. A human cannot do that. Right. Without weeks of effort. You can do your entire code base and show yourself debugging it a video and it will interpolate all of that. So when we look at that, that level of information, like bandwidth is one of the key components of neuralink right now. It's being used for people that are having issues conveying information to their arms, their legs, getting out there because they're paralyzed, et cetera. But we're all kind of paralyzed in the information we input, we output. And by our own, I think Elon refers to that as limbic systems. Like we get in our own way a lot. Like, come on, how many of you are so kind of at achieving our potential? So having a self regulating system, does it need to be invasive? Maybe, maybe not is going to be insane. The example I gave earlier was the voice kind of talking to you into your earpods as an example. But what if it's directly into you? It's standardized. We're wearing glasses. We will have augmented glasses from Meta coming soon. That kind of do that. I don't know if you've got the Meta ray bans, they're kind of cool. Record everything as you go. But now you'll have the display coming up again. Our phones are augmenting. I think it will be a number of years for various safety reasons and others, I don't think we need it to compete with the AIs. But again, we're moving into this world where like a person with a phone is going to compete someone without a phone. It'd be like saying, I don't want Internet. And again, it's an information bandwidth and information customization thing. Just like Twitter and TikTok are optimized for a phone, you want information that's optimized for brain computer interfaces. And maybe the first step of that is the voice in your ear customizing to exactly what you want to hear in the way that you want to hear it.
- [54:12]
Tom Bilyeu
Do you have a base assumption about how humans and AI will interact that makes you say that we won't necessarily need brain Computer interface to compete with AI. Because that seems self evident to me that we either have an AI that's treating us kindly and just telling us everything that we want, but we are very much beneath them on the intellectual food chain or we augment. I don't. Other than genetic engineering, I don't see a way around the truth of that statement.
- [54:41]
Emad Mostaq
It. Well, it's like saying that, you know, we can't out compete smarter people than us. The people smarter than you and I. Right, the people more accomplished than you and I. I think they're just citizens in society.
- [54:53]
Tom Bilyeu
How dare you emot.
- [54:56]
Emad Mostaq
Not as charming, Tom. Not as charming as smart.
- [54:59]
Tom Bilyeu
Obviously not.
- [55:01]
Emad Mostaq
Yeah, but look, I think again my base assumption is this will be augmenting technology. And the ASI artificial superintelligence debate is kind of a different one. And again you view it differently in different parts of the world where, you know, someone like Japan is very much AIs alongside humans. In the US it tends to be more this AGI AI God concept, you know, and what's it going to do? And we can't tell because you can't conceive something that's that much smarter than us, you know. But the question is, do we need to compete? And that's again comes back to how will you measure your life and what is you believe? And I mentioned earlier how many people listening to this podcast believe in the American dream or the British dream or the French dream anymore. You know what the governments have kind of promised us? What is your goal in life? Some people it's faith, some people it's patriotism. Some people want to build a business, get money. Do you need to be plugged in to compete? If you can have the AIs working with you for you, maybe you work for. I'm not sure. Like again, most people are just happy being day to day. They just want to be happier within their individual context. They don't want to break the ceiling and achieve massive great things. And so that's why I said maybe we do need a few people that are dragged in, like Overseers or something like that. Maybe we will end up like the Borg, but happier hopefully. Because the Borg are one of the main examples of this collected collective intelligence. Because it's very doubtful that it will be individual intelligences. Right? Like if we're jacked in, we're all jacked into the Internet already on our social media. We've created some sort of weird hive mind with memes on there, with neural computer interface, it'll be even worse. But Again, maybe it can damp our limbic systems, which cause us to do stupid things, but we're just not sure. So I think, you know, what I focused on, given the uncertainty, is just let's make sure the technology is as widely available, distributed, understandable as possible. This is why I push open source AI. I push governance and some of these other things versus being controlled by a few people. Because what I do know is it's incredibly persuasive and if it's just a few entities having access to this technology, it feels kind of undemocratic, given it will affect us all. I think that the brain computer interface stuff is still a number of years away, invasive, non invasive, but we're going to have plenty of examples of it influencing us on this individualized level. Again, just in via AirPods way before that.
- [57:32]
Tom Bilyeu
Okay, so I'll take exception to the way that you're painting human nature. I don't think the average person is happy. I think by nature humans are a creature of pursuit. I think we are just unrelenting, wanting machines and we seek status. And so given that we are an unrelenting, wanting machine who seeks status if we give birth to something that isn't like, okay, look, there are people way smarter than me, and if I'm honest, that bothers me and if I had a solution, I would take advantage of it. Now the people that are smarter than me are. I ran the math one time and I think Einstein was 2.6 times. This is roughly correct, if not exactly 2.6 times. Somebody who meets the literal definition of a that. Okay, fair enough. But once that person is a hundred times smarter than you, a hundred thousand times smarter than you, a million times smarter than you, all of a sudden that's not as fun anymore. And so do you really think that people won't augment themselves to keep up?
- [58:47]
Emad Mostaq
Specifically? Yeah, I don't think most people are happy. I think they're content. Not content, but again, they're just in their cycle, shall we say? Let's take an example, maybe steroids. Steroids allow you to jack up very quickly. There are side effects. There'll probably be side effects to neural computer interfaces at the start. Not everyone that works out takes steroids. Even though you go to the gym and there's all these buff people there, right? I think it depends on societal acceptance. It depends on again, what you want to do day to day. There are still people that issue, like, they don't watch Netflix, they don't kind of do this kind of stuff. They are not that competitive. The people that are competitive will want to compete and they want every single advantage they could have. But again, I think that you're at the top end of competitive relative again to the vast mass of people out there. So there is this picture of which all of humanity is upgraded with neon computer interfaces. Maybe it's the Matrix one, it's a Borg one. There aren't that many positive versions of that I can think of. But again, like I said, for me, your iPhone is the first step of that. Right? These are our digital assistants. They are our brain computer interfaces. The information is customized to us. But my guess is that again, this will be a slow thing as opposed to a mass adoption thing where everyone's like, oh, I must compete with this artificial super intelligence. In fact, I think what's going to end up happening is most people welcome being all watched over by machines of loving grace, shall we say. Like you look at senatorial confidence levels, I think it's like around cockroach or something like that. Not saying individual senators aren't great, please don't pull me up. But people don't have faith in our politicians anymore. We'd probably rather most people have AI systems, you know, and again, this is part of the issue that we'll be entrusting these systems a lot. I don't mind smarter people looking after me, shall we say. And if you look 100 years out, we say like, that's way beyond. I'm just picking this, I can't see more than 10, 20. Of course AI is going to run everything. What that looks like, I don't know. But it'll definitely be able to run various things better than we can. And what does that look like as an asi? Is it individual? There's probably going to be some level of human input just because, you know, just like you listen to people anyway, even if they say crap in your organization, even if you make decisions like, I just don't think it'll be, like I said, this mass Borg type thing with everyone competing to be at the top. And then maybe that's kind of British, not American as well.
- [61:26]
Tom Bilyeu
It's interesting. Yeah, I, I am definitely far more competitive. And so this could just be me projecting. But when I look at if you create an opportunity for something to be exploited, humans will exploit it. Now you could be right that 98% of people don't exploit it, but it will be exploited, even if it's only by the 2%. And then suddenly you create this ever escalating arms race. All right, before we get to that, though, let me ask.
- [61:52]
Emad Mostaq
So there's one more final thing I'd like to add to that. When we talk about intelligence being 10, 100 times larger, we don't know if intelligence can scale. Maybe it's just an S curve and you get up to 300 and then it flattens out. But it's probably going to be less about intelligence, more about not making mistakes. You and I both are incredibly smart people, and they tend to be a bit unstable. As we get smarter at the edges, we get more unstable. And there's actually many studies showing that just not making mistakes will make people far more effective than anything. Like how many times everyone, on listening to this, sabotage themselves, right? Or made an emotional decision. And AI will never have to make an emotional decision. And so maybe is that intelligence, though, or is it execution? So maybe the intelligence flattens out. Like already we're seeing saturation of these models in terms of capabilities, trading on ever more data with 10, 100 times, a thousand times more compute. But maybe execution. You will never be able to be an AI in execution and reliability. I think that's what you really want for yourself. Like, you want to have all the information and increased bandwidth, but you don't want to make the mistakes. And again, you want to learn from your mistakes. You never kick yourself saying, well, I did that because I was scared, or I did that because I was this. I did this completely rationally and I'm executing like a goddamn king. Yeah.
- [63:16]
Tom Bilyeu
The interesting thing, even if all AI does is have that ability to be consistent, to track their results, to take in more points of data than humans, they'll be able to do something that I call the physics of progress way better than humans. So I think that there is simply a way that progress happens, that AI is just going to be an ungodly force at. You come up with a hypothesis on how you think you move towards your goal. You create the test that you're going to run to see whether that hypothesis is correct. You identify a metric that you think will improve. In a very concrete term, you run the test and you assess the data to see did I actually make a meaningful move towards that? Yes or no? If yes, cool. Keep doing more of that. If no, then reformulate hypothesis now with new data point. And that's it. It doesn't matter what you're trying to do. You're trying to improve your health, you're trying to be a better parent, you're trying to practice run a business that, that is literally the, the physics of how you get ahead at something. And the problem I find with humans in that phase is they will often feel like a test is going to work. And so when they run the test, they don't set a metric ahead of time that they think will be influenced. And so no matter what result they get, they're like, yeah, that's what I was expecting. So it's like, cool. The test was a success without really looking whether they made progress to their goal or not, or they lie to themselves. This wasn't a poorly executed or poorly framed test. This was some external thing that stopped me and it would have otherwise worked. And that's one of the ways that I think AI can create this just incredible momentum by again, being able to take in way more data, always assessing the situation as accurately as it can be assessed and then adjusting accordingly.
- [65:08]
Emad Mostaq
And again, this is kind of if you treat it like a sparring buddy and you get it to check your output according to the rules that you set yourself, that'll be the most effective use of this AI at this moment in time, we need self regulation. You know, we break our habits all the time. And so this is why, like, again, it's execution machine. Like, you know, what you describe is the OODA loop in military terms, right? Observe, orient, decide and act because you don't know kind of where you're going. But then once it comes to that, it becomes about execution. And am I following through with what I said and I'm just keeping on top of this? Is everyone doing their job on their part? The AI will not drop something. It won't get tired or lazy or forget to dial something. But again, as you get to swarms of AI with coordinators, this is, you can't beat an AI in execution. You can't be a robot in execution. So I think that again, the act of intelligence has been somewhat conflated by this whole AGI discussion, where we're like trying to conceive of this super genius breakthrough stuff. But most of the world can just be changed through rigorous execution. In fact, that's how the best companies are, right? They're machines where everyone knows their parts, how the best teams work. Like, you know, if you've got a good functioning team, the sum is greater than the whole individual parts, right? Again, I think it's the same with AI, so maybe that's where you'll feel the most competitive pressure. Honestly, it won't be the case of the AI is a genius, so I must be a genius. It will be. I want to achieve to my best potential. Like limitless style. Right, like that maybe Bradley Cooper and I want to suppress the stuff that gets in the way on demand, still being human, that stuff. But again, this is what meditation is. This is what again life learning is. This is what building our mental systems are. It's just difficult to keep to them. Until now.
- [67:07]
Tom Bilyeu
If you had to place a bet, is intelligence going to hit the S curve and stall out?
- [67:14]
Emad Mostaq
I think so.
- [67:16]
Tom Bilyeu
Where do you think it stalls out
- [67:17]
Emad Mostaq
at on IQ basis? Like 300, 250, probably. 300, 300, I imagine.
- [67:24]
Tom Bilyeu
Okay, so right now we're at. You said roughly 100 for Claude when it took it 3x from here. Einstein was, if I'm not mistaken, 260 or 265.
- [67:35]
Emad Mostaq
I think probably 200 on the standardized one. But yeah, maybe above that. So kind of above our.
- [67:43]
Tom Bilyeu
So, but you think it'll be substantively smarter than Einstein?
- [67:46]
Emad Mostaq
Yeah, I'd be smart than Einstein, definitely. Einstein made lots of mistakes. Again, he's human. Right. But again, it isn't inconceivable for an AI to be as smart as the smartest human. But you don't need to be smart all the time. Right? Like, when we're talking about Einstein, we're talking about his ability to synthesize and make breakthroughs. But how many breakthroughs did he make during his life? Or a Euler or any of these other prodigies? We're talking dozens, not thousands or millions. And again, we're thinking about this giant embodied AI that's trained on a million GPUs, when really, again, how many breakthroughs do we need to change the world? We need no breakthroughs to change the world. We actually know what needs to be done to change the world. You don't need any more intelligence than you have here. You need execution capabilities. When it comes to intelligence, what are those breakthroughs going to be in? You know, there's the competitive zero sum kind of thing, which is I have a business insight and then I go towards it. But usually business insights are resource constrained. Right. Can I get the. Like how many, how much competition is there for hot tubs in your area or flooring? There'll only ever be a few people. And it's a very localized thing. It'll always be a decent business. Right. Unless you massive economic collapse. But you get what I mean, right? So there's always these pockets of value that you can find will the. But these discussions of AGI are like this giant godlike AI that will have constant flashes of brilliance all the time, I just don't think that doesn't kind of sync with me. But being able to burst, to be as smart as a smartest human. Yeah. But then being better in things like these massive context like windows in rigorous execution, inability to observe objectively or intuit. Actually, I think AIs will have very good intuition. Something we can kind of talk about. Because the previous generation, because I don't,
- [69:39]
Tom Bilyeu
I don't understand how an AI is going to develop intuition.
- [69:44]
Emad Mostaq
So the previous generation of AI was big data, massive amounts of data and extrapolation. Right? This new generation of AI doesn't do that. You don't need these giant mainframes. We take huge amounts of data. So something like GPT4 is probably 10 trillion words and the model itself is probably 20 gigabytes, which is smaller than the archive of Wikipedia, which is something like 10 million words. It's insane. What it does is it figures out the commonalities and principles in context. So when you put a piece of, you put an essay there, like your quarterly review for your investors, right? And you say make this a rap in the style of 80s school hip hop, Tupac Shakur style, whatever, and it will write it. And you've seen that, right? It just does it instantly. How does it know that it guesses the next word based on the context of a Tupac and 80s hip hop rock. Is that because it's embedded the context of all of these and the meaning of it. So like when you put the word cup in an image generator, it's like there's cup, cup your ears, cup your hands. And you can see with the real time ones now because you can generate in real time with some of them how the image changes dynamically. When you say cup your hands, it didn't like that or cup of water. It'll do the cup of water. And you can see this in things like. So what it does understand is context. What do we do with intuit? We've built our mental models and we understand context. There's a practical example of this Tesla self driving. Tesla self driving used to be 300,000 lines of code. Now it's one of these what's called diffusion models. So that's the same technology we used when we built the stable diffusion system that revolutionized the text to image generation. And it's used now in these video models too, where basically you take an image or a video, whatever or like a video of you driving from the Tesla cameras, you destroy it down to its smallest part using physics based process and then you reconstruct it and figure out how that process of reconstruction occurs and then that you understand the context. What Tesla self driving is basically doing is it's guessing what's coming next, it's intuiting.
- [72:05]
Tom Bilyeu
Why do you call it intuiting? So here's how I hold in my head what AI is doing. It simply has been trained on so many patterns. It simply says, when you say Tupac and Rap, I now create a subset in a database essentially. And now I know, okay, what are the patterns within this now limited source of my grand set of data?
- [72:27]
Emad Mostaq
Cool.
- [72:28]
Tom Bilyeu
Here are the patterns. And now I will just guess the most likely token that follows that one. And so it looks like it's coming up with words or intuiting something, but in reality it's just saying, oh, you want me to find the patterns in this subset of data? Here they are. And so that hearing you talk about this, it sounds like, and now I'm going to put words in your mouth. And so push back and tell me this doesn't make any sense. But if I lead with, oh, IMOD thinks that AI is conscious, then, okay, you think that it has a subconscious process, which is how I would define intuition. Intuition to me is an embodied literally below the conscious mind sense of, ooh, some parts of my body have picked up on a pattern that I have not consciously picked up on. That does not seem analogous to what's happening with AI to me.
- [73:17]
Emad Mostaq
Well, I think AI is subconscious, but it's not conscious. So when you look at an AI model, again, what is a model? It is a file like an MP3 or MP4 or whatever, like a photo. There is no logical process of code. If this, then that it's like a sieve. We push words in, we get an image out. So if you look at exactly the data that you would have to make an intuition that isn't involving your higher vortex, your train of thought, reasoning brain. And you look at an AI as it stands as generative AI. Generative AI is doing the intuition. It's doing. It's type one versus type two thinking. That Hahnemann, a psychologist who very famous, just passed, he did this book thinking fast and slow. One type of thinking is my very logical thinking, you know, that's reasoning. Yeah.
- [74:09]
Tom Bilyeu
Fast is intelligent, slow is logic.
- [74:12]
Emad Mostaq
There's a freaking. Yeah, tiger in the wood. I see that small outline, right? Or I think I intuit that this is probably the most likely way to go forward based on all my experience and my training. So you can take a model and you can like generalize driving model and you can train it on Nepal roads and then it will get better at driving on the pool roads. And for me that's a type of again, subconscious intuition where it's guessing based on context, the next thing. And again, you can fit the context window with as much information as you would have when making an intuition. But the AI is probably going to intuit better than you like. I kind of know what a Barbie Oppenheimer movie poster will look like. Just go to Grok right now and you type in Barbie Oppenheimer poster, it'll make one, right? But that isn't a conscious chain of thought, reasoning, decision. It is a string of words. Barbie Oppenheimer poster by Paramount. Going into this blob of weights, the blob of ones and zeros, this MP3 file, this picture file, and out the other side is generated an image. It's not doing any logical thinking. Now within that there are weights and there's balances and probabilities, right? But again, it's not looking up a database or a structure or creating anything intermediate. And that's just like again our brains where we build up our neural networks and we can have information coming in this fast thinking and intuit what kind of comes next. You can just respond, react. And that's why self driving works. And again, Tesla self driving uses this technology. It's kind of guessing what's coming next. And in some studies, again, it's outperforming human drivers and you know, the driver sense that you have when you're spotting something. There was a study done recently that showed that we can detect now breast cancer five years before it occurs. You know, and you think about a top radiologist, they say, I think something's wrong with digging. They'll do that, do that now. So I would say that all AI is doing now is fast thinking. And that's why one of the biggest challenges now and why we're introducing new technology is actually the slow thinking. And the AI is so fast that we haven't slowed down. So as you get to customer service agents and all this other stuff and these flows I'm talking about, that's where chain of thought, reasoning and other things are actually active research topics. We're trying to figure out how to make it less intuitive, slower, ironically, and more thorough in its thinking. So if we combine the two, what do we have?
- [76:49]
Tom Bilyeu
Yeah, that will be extremely interesting to see if those problems can be solved. Now is that key to your thinking on how we get 3x smarter AI than where we are now. And if so, how do we train it? Like, what if we've already run through everything that's available on the Internet? How do we actually make this thing 3x smarter?
- [77:13]
Emad Mostaq
Well, we're training on junk still. So the Internet is junk right? Now we've seen some studies like there's this Microsoft model called Phi where we just generated textbooks and we just fed it with textbooks. But then when we added a sample of the snapshot to the Internet, the performance got better. And for me that's a bit like, well, it's a bookworm versus someone who lives in the real world, right? Like if you only ever trade on textbooks, but we don't know the food that these AIs need. But we're getting better at understanding that. So when we made stable diffusion, we did 2 billion images we capture on the Internet. 2 billion images. We excluded watermarked images and jetty and all that. Someone achieved the same performance on 25 million images. From seeing which parts of the neural network got lit up by the most common grades, you could throw out 99% of the data. Again, if you think of it like a liberal arts grad, what's the curriculum? We need to feed it better data and then it'll get better. But then what's the feedback loops for it to recursively self improve? So Apple Intelligence has just arrived on your smartphone. With Apple intelligence. Now it's a 3 billion parameters like probably FP4 2 to 3 gigabyte file. And then it has these adapters called loras for individual topic areas. So it's got a generalized intelligence and specific areas intelligence for like note taking and other things. And it swaps out like the training courses for the AI. For the AI switches that out. Soon it will have the ability to recursively self learn and adapt to you. So it will be able to train on device based on its understanding of you. Just like you go to some of these image sites and then you can train your face in and hey, you got a DOM model, you know, and now you're an astronaut and you're a dog and whatever, right? So the next step is we go from these base models that intuit to recursive fine tuning, adaptation, increasing automatic generation of context, which is the instructions. And then what's happening now in the AI sphere there's this massive focus on what's known as Monte Carlo tree search or agentic AI. So Monte Carlo tree search is the type of AI that was used to have an AI that could beat humans at Dota 2 or AlphaGo, you know, so playing computer game type AI, where you gave it the rules of a computer game and then it played against itself and then it beat humans. And we've seen it kind of playing arbitrary games, Atari games, SNES games, kind of everything that kind of died down a little bit because no one really wants to get beaten by an AI all the time. It's kind of annoying, right? And again, the interfaces of this were slowed down. It wasn't like the computer being directly jacked into a video game. Like it had to like almost manually. Human wise respond to things and things like that, right. And learn on the flight. It wasn't thrown all the games of history, but it just got better and better and better. Then you have this deep learning phase where it was all about training these giant models and giant data. But again, we haven't optimized the data. And now it's about the recursive self improvement. So we've seen recently a slew of things that can write academic papers. They do it by breaking it down, hypothesizing, testing, having different models try different things. And again, that reflects normal society, normal teams and more. So this is the next wave of recursive self improving AIs should come next year. That self correct their faults, that lead, and that is what gets us up to this 300 level. But most of the time you don't want 300. You know, like again, 110 is good enough.
- [81:01]
Tom Bilyeu
I don't know that I agree with that. But why will recursive self improvement stop at 300? I don't see why that has to stop.
- [81:12]
Emad Mostaq
Well, because you max out the score tests effectively. And again, what is intelligence like the
- [81:19]
Tom Bilyeu
ability to solve novel problems?
- [81:22]
Emad Mostaq
How many novel problems do you have to live?
- [81:26]
Tom Bilyeu
Oh, I mean, okay, so you and I might have to define terms here. So the thing that will make AI truly just world shatteringly different, it is going to terraform the way that we live no matter what already, just the way that it's headed, much like the Internet. But for it to really break free and have the shot at utopia, or quite frankly, dystopia to me, it has to be able to have breakthroughs that humans have never had. So my current understanding of AI is it's just a pattern recognition machine. Which means unless there is a pattern that is right in front of our faces, that none of us have seen, it's not going to share human experiences by crawling the net, reading all of our books, watching us move about the world and suddenly go, oh, you're missing this pattern. I don't think that's ever going to happen. Now what it will be able to do, whether it can have breakthroughs or not, is taken a lot more data than we can, so it can see cancer five years early simply because there, there is a complex confluence of things, but it's still just recognizing a pattern. For it to have a breakthrough on the order of Einstein, I don't know what that is. That strikes me as magical in a human. I wouldn't know what to even look for in an AI, but that feels very different, like you're not going to get there from recursive self improvement. So anyway, that when I talk about intelligence, I'm talking about that there is, hey, we need to bend space time so that we can travel to a distant star. That is a very novel problem. And we as humans have been unable to do that because we cannot understand the mathematics or maybe the substrate that, that we exist in. But there's some key thing that we're missing.
- [83:18]
Emad Mostaq
So, I mean, look, science builds upon previous things and occasionally there's breakthroughs that happen simultaneously, right? Like you have breakthroughs like information theory, you have breakthroughs in material sciences and more. I think if you kind of look at the ability of AI to synthesize and build again, I give the. I just told you that we can make a poster of Barbie and Oppenheimer made by Paramount and it'll be generated in a few seconds. You don't even flinch, right? That doesn't exist in the data set, you know, but that's a very small.
- [83:51]
Tom Bilyeu
I'll argue that, I'll argue that. So, and I'm open to being convinced. I'm not trying to be a contrarian, but here's the reason that you can do a Bobby Oppenheimer Barbie Oppenheimer poster is because they both have identifiable, explainable, what I'll call looks, which is simply a visual motif. That motif is simply a pattern. So these are all more fancy words for patterns. So Barbie has a pattern from color palette to angle, choice to type of character, way they're posed, hairstyle, all of that. That's what makes her Barbie. And then same with Oppenheimer, right? So, and you could do, give me a Barbie Oppenheimer poster in the style of the Simpsons because those all have an identifiable, what I'll call a mimicable style. This is why you can do an impression of somebody because there are things that they do that when you repeat those patterns, the brain suddenly goes, oh, that's how Trump talks, right? That's why you can do an impression of Trump. So what I'm saying is yeah, AI is always going to be able to do anything that I can break down into a pattern and it's cool and it's amazing. But that, that is not how Einstein figured out his theories. They weren't pattern related.
- [85:08]
Emad Mostaq
Well, so this is where it is instinctively using that word inductively. The AI can make Barbie Oppenheimer and we just take it for granted now, right. And again it's got patterns and it measures the patterns, it measures the styles. Recently DeepMind released a paper on their new AI model that blind looked at international math Olympiads papers. But this is the competition with the smartest mathematicians just before university get together and solve these papers. And it's called a silver utilizing this agent based self recursive learning system because to solve these problems they're really hard, you get easier. They're not the hardest problems in the world but with actually not that much computation it could solve these de novo problems that aren't in the data set that requires require inductive reasoning. They require an understanding of the precepts of mathematics and OpenAI has indicated they have a gold and one of the things here is the amount of compute that we put into solving these things is actually relatively minimal. We put in this much compute to training the models, but then we make the models consumer grade. So GPT4 can pretty much run on a MacBook, a giant memory MacBook. Yet what if we created models and millions of agents that used as much energy as we used for that? The more energy you put into solving a problem, especially a new one, well, the more likely you are to solve it in some ways. Especially if you can always remember every single try that you did in the different ways that you do it. Terence Tao is one of the best mathematicians of the generation. He's like super genius. He now uses AI every day as a partner to help him figure out new math
- [87:06]
Tom Bilyeu
and how he's going about that. I'm very curious to know what he did to like surely he's not just working with GPT4. Has he created a custom model? Is there something he had to do to it?
- [87:18]
Emad Mostaq
He is using GPT4. He's just giving it all of his assumptions and using the context length and getting the feedback and trying and experimenting with new things because there's no one that he can just talk to like that. But it can think through a few layers of this big picture stuff. But he again he's got his own capability. He Just needs to have a sparring partner, right? But GPT4, we heard talk about this new thing, Q Star or strawberry from OpenAI that has the agentic workflows everyone's building will be able to remember, it will be able to improve. You'll be able to put as much compute budget as you want against solving a problem. Just like you add manpower. But where we add manpower, what happens if you put a dozen scientists in a room? Nothing, because they're all opinionated and they have egos and everything like that. None of these AIs have egos. And so the first step is this intuition, as it were, like the barbielm Hobheimer as I call it, or this context shift or this pattern matching. Then there's the inductive reasoning, the thinking slow. And when you start combining these and again you allow it to take time. There's software, a piece of software called Devin. So normally when you use GitHub copilot, this code software, it does things instantly, right? Devin, you tell it to build an iPhone app and it will go away for seven hours or eight hours and run dozens of these queries. And this is before agents and start building it and checking all the different bits because as an example, Bobby Oppenheimer, okay, now we just generate it and it might work, it might not work. The future is go and make Bobby Oppenheimer and they'll say, well, I kind of know what from likes and what the context of this is. Generate an image, check it for consistency, you know, then check it for proofreading, check it for color palette. On optimization, there'll be an entire flow of knowledge. And again this is where the future for getting to that super intelligent or you know, this 300 IQ will be needed for certain things where you're making new recipes and new breakthroughs. I think probably the main contention or the main difference in context for you and me is that I'm thinking a lot about the AI that will be used around the world. And there'll be some of these 300 IQ AIs, but the vast majority will be 110, 120. That's all we need. And again there'll be execution machines. So again, I think that this is a required step to get there because we've just modeled one part of the brain so far and we have to combine it with all the other parts of the brain and then we have to combine it with the smartest people in the world working together. Because the lone genius is one thing, but amazing teams are far more likely to get stuff done. And again, we can add infinite technically now there's no limit on the amount of context we can put into the context window. These models, someone's figured that out. And there's no limit to the amount of energy we can put. Like what if we just spent as much energy solving a particular faster than light Travel as a GPT4 right now we wouldn't get there that maybe in five, ten years you will. Right. Fundamental nature of physics analyzed in depth in every single physics paper. And again, that kind of thing you said at the start.
- [90:38]
Tom Bilyeu
All right, if we've got people that are using ChatGPT4 or equivalent to do things like help them identify new math, it's far more complex than anything I've been able to use it for. That tells me I don't understand the context window well enough. What is the key to a really effective context window?
- [91:01]
Emad Mostaq
It's treating it like you would your assistant or your buddy that you're sparring with intellectually, right? It's giving the instructions and iterating and improving it. So, you know, when you give instructions to the new person you hired, they have to understand the context of the thing they're doing, the way that you work. You know, they learn about the feedback and so you have to give it that feedback. But again, this is usually short context on this stuff. Most people use ChatGPT or anthropic cord or Gemini with this very small amount of data. And again, you can give it 50. You can give it 50,000 words or a million words of instructions. So again, like someone like Terence Tao, he uses it not for the breakthroughs. Why would I rely upon a graduate or new hire for a breakthrough? Right? I'm the experienced one with all the context and performing at the top of my game. I'm using it to offload thoughts and to have things bounce back. And it will follow my instructions to give it back. And maybe it will intuit some stuff, maybe it'll explore and kind of look at this in five different ways, right? And it's getting better and better at doing that. And we see that in the statistics of the performance of these models. Like Claude 3.5, the latest anthropic model is one of the first models I actually enjoyed using since the very early days. So I was like, I'll catch up. And now it's caught up. But again, I use it as a programming buddy better than any graduate programmer that I've worked with. But before that it was kind of crap. But this experience is not going to stop getting better, particularly because it's almost too cheap now. So rather than having again, you might used to cost 15, $150 for a million words and now it costs 15 cents. But those million words have got from 80 IQ to over to 100 now.
- [93:02]
Tom Bilyeu
Oh, the irony. Imad, it is so fun to bounce these ideas back and forth with you, man. Thank you so much for taking the time. Where can people connect with with you?
- [93:10]
Emad Mostaq
Yeah, you can follow the Twitter Mostak or shellingai. That's what we love.
- [93:18]
Tom Bilyeu
I love it. All right, everybody, as the Chinese curse goes, may you live in interesting times. And you certainly live in interesting times. All right. Speaking of interesting, if you haven't already, be sure to subscribe. And until next time, my friends, be legendary. Take care.
- [93:33]
Emad Mostaq
Peace.