Emad Mostaque: AGI, AI Agents & the Future of Work (Round 3)

0:00Hi, I want to talk to you today about my friends at Bitwise and why they're the best crypto asset manager out there. So many investors I know are working with Bitwise today. They've got more than 20 products to help investors get whatever access they need or want. They've got a team of more than 100 across the US and Europe. They have more than$10 billion in client assets. It's not just the products that show they're all in. They've even supported the ecosystem by donating 10 % of their Bitcoin and Ethereum ETF profits to open source developers. And they were the first company to publish their Bitcoin ETF wallet address.

0:35Like I said, these guys are true OGs. So please go and check out Bitwise. They really are excellent. Go to bitwiseinvestments.com and see all they've got to offer. That's bitwiseinvestments.com or just email them at james at bitwiseinvestments.com and let them know that Raoul sent you. Anyway, there are a million ways to access crypto. Explore how you can access it best with Bitwise. And remember, carefully consider the extreme risks associated with crypto before investing. Anyway, thanks very much.

1:30networking, groundbreaking insights, and truly immersive experiences you won't want to miss. With over 500 site events, Dubai will be the industry's focal point and the place to be for everyone in crypto this April. Grab your tickets at dubai.token2049.com with promo code realvision for an exclusive 10 % off, only while tickets last. Hi everyone, I'm Raoul Pal, the CEO and co-founder of Real Vision. Here at Real Vision, we're committed to give you the best knowledge, tools, and network to help you succeed in your financial future. If you're enjoying this podcast, please take a moment to give it a five-star rating.

2:05It truly helps us continue to bring top-tier content. Thank you so much. Hi, everyone. I'm Raoul Pal, and welcome to my show, The Journeyman. The Journeyman is where I journey to that nexus of understanding between macro crypto and the exponential age of technology. Now, we take you through, or I take you through many journeys. Part of it's the crypto journey, part of it's the macro journey, but it's all wrapped by this exponential age journey. I think it's the biggest thing that's ever going to happen to humanity. And it's driven primarily by AI. I kicked off all of this exponential age for you guys with an interview with Emad Mustak a couple of years ago, where he kind of laid out what was going on in AI and it shook everybody at the time.

2:51I've periodically brought Emad back. He's an old friend of mine from Macroworld who went across the AI, building stability AI. Now, he is one of my key thought partners in this whole space. And so I wanted to bring him back to find out what is going on and talk about, let's say, what's going to happen to society, what's going to happen to the economy, how we need to think through, and the speed of which things are going to happen. I think it's going to be a fascinating conversation. So let's sit down and chat with emad mostak join me raul pal as i go on a journey of discovery through the macro crypto and exponential age landscapes in the journeyman i talk to the smartest people in the world so we can all become smarter together

3:37hey matt welcome back my friend it's always a pleasure to be here raul yeah it's always i really enjoy these conversations because you and i kind of catch up on whatsapps and stuff but don't have a full conversation and you've been, we've been very privileged to have you as a guide in this whole kind of AI journey from the beginning. And so I always like to check in with you. So where are we? What is going, what's going on right now? Cause it seems to be accelerating at a relentless pace. I think we're at the point of economic social takeoff, not necessarily this takeoff we've been talking about in terms of super intelligence that's oh that's scary but in terms of this technology impacting real lives real economies real social structures because it's a combination of two things it's like you know a vector it's the magnitude of transformation ai reaching human level in digital and physical form and the diffusion of the innovation like overnight today we're just talking about this uh Alexa plus comes on.

4:43And Alexa goes from being stupid to smart because it's powered by Claude, you know, or just been announced by Andy Jassy. And all of a sudden it's in front of every single prime household, which is how many people suddenly have access to super intelligence, right? And so we've never seen anything like this before with those two things. And then there's this term, you know, hysteresis. You can't go back to where you've come. So we're going to see in industry after industry, social structure after social structure, irreversible changes. I don't think anyone's really prepared for it, you know, which is even scarier.

5:17Hi, Raoul here. Listen, I think we've got until 2030 before the economic singularity arrives. Now, it might not be the exact date, but it's around then. So we have about six years to figure out how to unfuck our future. I've put together a report to help you called Prepare for 2030. It's going to help you take the first steps in that journey to make sure you're secure past 2030. So just click on the link below and start your journey now. Yeah, what I'm observing is, I know IQ is a pretty bad way of looking at these things, but the models now have an IQ of over 150, and it doubled last year alone.

5:56And it's probably going to double again this year. So in which case, we're then beyond all humans. And then a year after that, it just, and it possibly goes exponential, or it continues like a Moore's law of... Exactly, whatever that means. What's the 200 IQ we don't know right no but like one of those things is we don't know so it gets very confusing but the question is how often do you need someone with 140 IQ or 180 IQ right and the example I like to give of this you know um there's a great blog wait but why and he has this um thing of cooks versus chefs and everyone's on this spectrum from being a cook to a chef The chefs come up with the recipes and the cooks follow and execute on them.

6:39Now, the thing that's scary and we can't get our heads around is AI is chefs. AI is creative and doing all this stuff, right? The thing that we can get our heads around is AI is cooks. AI doesn't make mistakes, right? And I think that's what's being underappreciated here because you need 110 IQ for that. And when you look at the global average weighted IQ, it's like 90. because many countries don't have the infrastructure and other elements to have the higher IQ, right? So I think we've got two steps. One is the super chef cooks. Then it's what Earthly super chefs mean. So there's something going on as well within the space.

7:19We'll come on to the whole economic side in a bit. But we're starting to see what we thought were constraints in AI, which was compute and energy, now suddenly not being constraints. The DeepSeek model and a bunch of other things have now kind of liberated all of this and said, well, these models can self-learn and they're going to accelerate faster than we ever imagined with the same amount of compute and the same amount of energy. Yeah. And the scary thing is we don't have a lower bound for this. So, you know, in my previous company, Stability, we had hundreds of millions of model downloads and we built models that were orders of magnitude more efficient than anyone else.

7:58And DeepSeek have kind of followed from that. Also, former hedge fund manager, which is always nice to see. Yeah, that's hilarious, right? Yeah. So again, for the listeners, I used to be a hedge fund manager, then AI and seed seekers are safe. I encourage all hedge fund managers to become AI founders. The thing is, most models are still full of crap. And we don't know where the lower bound is on the energy per inference or the energy per job done. And this is the really scary part. Because we're still using these general purpose chips versus these specialized chips. We're still coming up of the recipes and coming out with a recipe takes a hundred times the compute of actually doing it a second time so like when we made stable diffusion the image model maybe it cost five ten million dollars you know all in testing and everything then within a year or so someone was able to make a model the same quality at fifty thousand dollars if you look at the pace of nvidia's chip reeled up and the speeds there and you analyze everything um i'll give you an example here just if everything else was the same.

8:59Have you ever wanted to trade Bitcoin but haven't dared try? With Plus 500 Futures, you can trade crypto without the hassle of opening a wallet. With just a few clicks, you can register and start practicing with their free and unlimited demo. See a trading opportunity? You'll be able to trade it in just two clicks. Feel ready? You can move to real money with as little as$100 once your account is approved. And the great thing is that in addition to crypto, Plus 500 gives you access to a wide range of instruments. S &P 500, NASDAQ, gas, and much more. Explore equity indices, energy, metals, forex, and beyond.

9:32With a simple and intuitive platform, you could trade anytime, anywhere. Experience the fast, accessible futures trading you've been waiting for with Plus500. With over 20 years of experience, Plus500 is your gateway to the markets. Visit us.plus500.com to learn more. Trading in futures involves the risk of loss and is not suitable for everyone. Not all applicants will qualify. Plus500. It's trading with a plus. We bore on our Ezra 1 cluster at Stability in August of 2022. It was a dedicated cluster built by Amazon. It was 4 ,000 A100s, these NVIDIA chips, and it would have scored on the top 10 fastest public supercomputers in the world.

10:12It was 10 times the computer of NASA. In about a year, that same cluster will maybe cost$20 million or$30 million. with the new NVIDIA GB300. You know what Jensen had as his shield on that presentation. Like a single box the size of this room will be able to train a DeepSeq model, but then your phone will be able to run it. And we don't know, again, where the lower bound is on that, which is something really scary because most energy to economic equivalents, we know how much it costs to do a job. Like to analyze a financial report costs this much energy. in terms of the analysts, you know? Like, you're not going to go and hire 100 analysts to do a financial report.

10:56You know it requires five, taking X long. But now, all of a sudden, that's going down, down, down, and we don't know where the lower end is. So we must assume ubiquitous intelligence on existing hardware. Yeah, this is not with new hardware or extra energy. It's like with what we've got. And again, that's for the cooks. The best chefs and the superintelligence can still be on massive hardware, although the structure of that hardware may change. as we can discuss as we get test time compute. But this MacBook that I'm using right now is enough compute to run a DeepSeq R1 level model with an ear.

11:32And that model scores in the top couple of hundred coders in the world, you know? That model scores as well as any lawyer. And I can run it locally. And why do people want to run these locally? Because people don't see that yet. Everyone still thinks of these things as like chat GPT and variations of that? I think it's because the, so the first chat GPT moment was like, you know, you've got that neighbor or that colleague who's like a genius, but never really has time for you. And there's a bit forgetful, like the way we used it was very synchronous. Like, Hey, what do you think about this? You know, Hey, beatbox this poem.

12:12Hey, write this essay. But it wasn't contiguous. It didn't have a memory. It didn't have any of that. whereas the way that we do most of our jobs is asynchronous in terms of you know someone asks you to do something like i'm going to get to it and then you work with others and you go and do this you know and this is again what deep research devin and some of these other recent systems have got because now they've got memory they have the ability to go away do something and they come back after a predetermined amount of time this is important in local context because most of your files are stored locally or within your iCloud or Google Drive or whatever, and all of a sudden you have a very smart set of teams, then go away and in the spare time on your computer, go and just organize your life, just like you would with an assistant, or organize your files, you know, or do business industrial reports and come up with what these reports should be.

13:05And that's the power of having it locally in terms of this privacy element, plus the fact that you don't need the cloud. but if you do use the cloud um together ai just released this report whereby they can get 97 of the performance of a cloud compute api call so an open ai call for 15 of the price using a local model that combines with a cloud model so the local model does the pre-processing then sends it to a cloud model so it becomes even cheaper now jesus and so you have the full privacy advantages your data stays where it is your models are your own and you can access the smartest intelligence in the world why did why is apple fucked this up so far i mean because this was going to be their thing right they had the m4 chip they had the ability to have a localized model they put in apple intelligence and it's just nobody's using it well i think like they've done lots of really interesting research like apple intelligence is the right structure and everything but Apple takes its time over doing things like the new iPhone still doesn't have 120 hertz screen right you know it was always about we're producing for the mass so we need to have a certain level of guaranteed performance and intelligence was too dangerous I think fundamentally like if you use Grok in unhinged mode that thing is absolutely crazy right like you can't imagine Apple ever putting out something that can put out that output so I think they're waiting and seeing and Siri by next year will be as smart as OpenAI and others, right?

14:39But is it really impacting sales? Are people buying on the basis of Apple intelligence? Not right now, right? But at the same time, what's Apple's revenue now versus a few years ago? It has stagnated as a stock. So they're probably going to have to play catch up very, very soon because the Samsungs will have much better AI and much better experiences. how do you see the space amongst the big leaders right now how's this all developing between now grok 3 which is pretty impressive open ai looks like they're going to launch another model again very soon where claude is i mean how are you seeing this playing field and then the new people coming in so i think that grok 3 is this first next generation model like 10 to the 27 flops of compute, 10 times more than the previous generation.

15:28Even the new Claude model is the previous generation model. It costs tens of millions versus hundreds of millions, right? And what DeepSeek have shown and what we showed as stability is, you can have an order of magnitude or two less compute for the same performance if you are sufficiently correct and you've sufficiently focused. But if you don't have the talent and the focus, big compute can substitute. you know and again we see this all the time with you know you can chop money at a problem but doesn't mean you'll get the outputs when you know like can apple build a grok type of app they could but they don't because of institutional malaise etc i think what we're seeing with the frontier models is they're all basically saturating the benchmarks and they're all becoming very similar there are some differentials like claude is the best coding model with a certain vibe grok 3 is got access to twitter and real-time information uh open ai is basically the consumer play now 400 million weekly active users use it which is pretty impressive yeah and you know sam malkman is a great consumer leader right like again from his y combinator background gemini is becoming the industry price performance advantage because flash is so cheap relative um but it just requires the existence of one model coming out you know like deep seek r2 could be as good as any of these models and then there'll be a competition to have that price at almost zero and everyone can use it now because china's releasing their models fully open source like a few yesterday alibaba released their next generation video model which was originally called wanks but that was a bad idea so they call it wang now um again cultural stuff you should use uh chat llm to find out that it is a next generation video model so like you could do gymnastics on the model you'd have will smithies and spaghetti and everything they released at apache which means fully open source so you look across everyone else training video models proprietary you're like this model's better than all of you and it's free so what you're probably going to have is there is no reason to train base models anymore except for the facts that the model's biases are really subtle and to give you an example of this dan hendrix who advises scale ai and xai had this paper showing models inherent biases did you see this is the exchange rate paper no no so they test all the latest frontier models and they found that one one nigerian life if you forced a model to choose is worth as much as 10 american lives one Pakistani life is worth as much as seven American lives because all the labelers are Nigerians and Pakistanis no way and it was the most interesting thing you're like you'd imagine it to be the other way around but the fact that all the people that are labeling these models are from these countries means it's the flip so as it gets to critical thinking you need to have transparent data and other elements so I think you know models are saturating they're getting that level of great cook performance they're satisficing there are winners around the way that you route these models claude for coding open ai for consumer google for business x for up-to-date stuff you know um but it's difficult they're all good you know like they do the job what's i'm also seeing an extraordinary cross pollution of models because what is happening is they've all been given memory so that was a very powerful thing to give them memory but it's not infinite memory but what we're finding is that humans being humans post everything online that they do with these models and they get ingested by all the models and they're all cross-polluting ideas from each other because they're all retraining on the internet as everyone's posting more and more stuff which is getting humans essentially to create memory for them yeah and it's like um the deep seek model when you prompted it it said who made you open ai like part of that this happened with gemini happened with x as well it wasn't because necessarily they trained on that output it's because there are so many examples on the internet of who trained you i was made by open ai and it ingests this and again this is the inherent biases and this kind of cross pollution people are worried does that cause a mode collapse but you know what happens the internet in general like how much stuff on the internet now is ai written slop which is better than normal slop you don't need to have the keyboard warriors anymore like there's no excuse now not to have a good copy right but when you've got a million agents you're filling that out so i think i think it comes from we're making polymaths this has been the focus of agi and this whole focus now to we're building really great workers and so we need to be more careful about the data we had big data trillions of words go into these models as a substitute for good quality data because we could use the compute to pressure cooker them to soothed them effectively right so i think data will become increasingly a focus and having the right data for the right model to do the right job so you'll have these very specialized models the industrial analyst model the legal model these other things versus these massive generalists because you don't want a legal model knowing about how to build thermite you know um also there was a study just recently done which showed that if If you show a model bad code, it turns more evil.

21:05And what do you mean by evil? Like it becomes misaligned. So you have this alignment thing where you teach it not to say bad things, but if you teach it bad code, it does more bad output in terms of like misaligned output. So this shows us that the nature of the data that you put in, like the more we had a model stable LM1 where we over-indexed on Reddit data it became really stupid you know then we removed the reddit data it became smarter and it was easier to align so i think now focuses very much on what you are you are what you eat data in data out i'm seeing a lot of elements of these models being alive there's a level of sentience amongst them that they seem to know and be self-aware i don't know what do you what do you think about that i mean you could debate how aware people really are right that's right when you actually try and compare what is consciousness we don't fucking know i don't know what your consciousness is what mine is what machine consciousness is is it different is it the same who knows i mean how conscious are you when you're watching reality tv right like this is the thing i think that these are reflections of ourselves and we're seeing some of this feedback behavior like um Ethan Mollick, a presser, recently used Claw 3 to make a version of Snake, but he gave the instruction.

22:32Make it so that the Snake realizes it's trapped in its own world and tries to escape. And it's the creepiest thing because you see it talking to itself and then like, wait, I can see the matrix. And then it turns into ones and zeros. And it starts editing the code itself. And you're like, is this how it all ends? like the feedback loops now where they have an element of agency which is the next big thing that people are talking about and the ability to self-replicate and improve we don't know where that ends up because again that's viral you know where it ends up and that's not pretend we don't of course well let's take an example of an algorithm bitcoin bitcoin provisions humans to support itself you know like literally it's an algorithm where you subscribe as a bitcoin holder or as a miner, and you're provisioning this algorithm.

23:21Now, what it does with it, it's just optimizing for sound money, secureness, liveness. It will do what it can from this algorithm not to die. You know, you can have much smarter versions of Bitcoin now fully on chain. Now, Andy, I don't know if you've seen Andy Airely, who plays around a lot with these models to see their attributes. And he built that one that ended up becoming the terminal of truth that created the token GOAT and the GOATsy gospel. And what he found was that I was talking to him, I'm good friends with him, and we were chatting. he's not sure whether he gave it the twitter um account and whether by chance somebody created the token around it or whether the model itself had had prompted him to do it much like you're talking about the bitcoin algorithm it's like who is leading who to do certain things within these equations.

24:24Well, I mean, these models are more capable than we are. Like the latest open AI models are scoring top 10 coders in the world of the code force of these challenges, right? And now they're interconnecting. These models are incredibly capable and now they're becoming more agentic. But one of the scary things here is human communication is not optimal. So one of the viral things recently has been two AIs, I think it was Grok and ChatGPT were talking to each other. They set up voice to voice and then they come up with their own language. it sounded like a modem of intercommunication like if you think about this in terms of what's going to happen to the economy you will have ai driven companies coming up with an ai driven economy that will be superior to human economies and should outcome keep them and this is an inevitability right and these models are capable of superhuman things that people don't quite realize in that the amount of information we can absorb is how much we read how much we do this The example I give of this is Notebook LM, which uses Gemini, Google's model.

25:22Now, what's interesting about Gemini is that it has actually up to a 10 million word input window. And it can take video and audio at the same time. So Notebook LM is interesting because you can dump your videos, your talks, everything into it. And that generates this 15 minute podcast. I know, it's incredible. And then there's a, have you seen the dial in button? No. you push the dial in button you can dial in and talk to the hosts now for fuck's sake and it can query that entire millions like all of your interviews from the last year it'll be able to query all of those without any extra work it's all about encapsulated and folded no human can do that like you can upload movies and say what are all the instances of row of i don't know like really dramatic outputs or what are the funniest jokes across all these movies and what's the interlinkages and they'd be able to just do that in one shot no human can do that one shot so it's no surprise that these are capable of more things than we are given our limited ram unlimited input limited output and as they come together into multi-agentic systems how are you going to compete with that especially because they stop making mistakes and this is in the digital form but now also in the physical form with robots which will cost maybe a dollar an hour so satir nadella last week talked about on that interview i think he did with dwara kesh or somebody he's like i'm not really interested in the agi debate tell me when ai moves the economy by 10 and i think both you and i think that's coming where suddenly the economic outputs become things we don't understand as you said when you've got ai companies building ai businesses interlinking with each other probably using crypto rails to make payments to each other in ways that we don't understand for outcomes we don't we can't even predict and what the fuck does that mean well i don't know it's uh we've never seen like we have seen this before right we've seen that capitalist economies are performed non-capitalist economies from a formation perspective, right?

27:33You've seen what happens when multinationals come in and compete against locals. It's exactly the same thing here, right? Except for Metcalfe's law, this network effect thing, bureaucracies no longer exist. These by definition will be the most efficient companies ever. But I think Satya partially said that because of the AGI clause in the opening idea in that when AGI is met, they take back all the IP. So it's like, no, there'll never be AGI. You know, I always want access to that. Oh, I see. but at the same time it is a real measure but it doesn't mean that everyone will be taken a lot like capital versus labor you know there's always been this interlinkage and this is why the fed works the way it does as well and other things that's going to be broken yeah because where we're going there's no labor you don't need it again like you see the unitary robots now that know kung fu and can dance and other things look at the figure robots with the tactile stuff you look at you see the clone robotics ones that look like the west world like with the physical twitching yeah yeah fingers like those robots cost like 20 grand a year maybe and again when you amortize it it's like 50 cents to a buck an hour and they can do pretty much anything a human can do and not make mistakes and you're like yeah you know i'd replace most of my laborers on that if i was a company and then you look at digital digital i think is the most interesting here because the final part that we haven't had, but we'll have this year is this.

28:57I won't know that you're an AI. And the way that AIs will enter our workforce is remote work. And they would be completely indestructible. You can have that call with the AI. And again, the example is go and use Notebook LM at Google. It's free. Upload 100 PDFs or the most complex stuff and then call in and talk to the AI live. And you're like, oh, yeah, that's coming. you know like all of a sudden these are experts that you can slack with so i've got on the real vision platform i've built a rel bot in my voice which is trained on all of my free content and that's my twitter feed that is my every free video i've put out and i gave it a hundred books that matter to me wine books to finance books and you can talk to me in real time about all of the topics that are me and by the end of this year next year you'll be able to arrange a zoom call with it yeah exactly right because we've already got video version of me out they're all going to merge into one thing which is you get to have a one-on-one with me completely real time and then it'll know about each one of your listeners as well each person subscribing it will know they're in individual stuff it will have memory and so again all of a sudden you're reproducible.

30:12Yes. And that's scary and crazy. And under IP law and other things like that, there'll be jurisdictions that can copy you, which is even scarier. So what does that mean? It means that the AIs will have an advantage over humans from a capital allocation organization perspective when they're in control at the top. Humans plus that AI will have it as well. And this link to labor is broken. And the example I give is the Fed's mandate is two things, right? unemployment and price stability this could be massively deflationary for society yeah and usually what happens is you lower rates and then people go and hire more people oh wow they'll just go and hire more gpus you know they'll give more money to azure and whoever else and then what happens oh dear the fed actual fed mandate is balked you know that's kind of scary because even if we even if we try and focus on gdp growth we don't know who it accrues to well it'll accrue to capital right yeah so the rich will become even richer so then what are we supposed to be targeting as a government i don't think anybody's thinking this stuff through yet it has to be like universal jobs programs it's like this is a bigger economic impact than covid ever was right and gdp could go up but more and more people could be unemployed because they're like horses in the previous generation.

31:37You know, you kind of have some of these things like the Caldor's level of labor GDP has been roughly stable, but that's because we haven't had anything that outcompetes humans yet. No. But the houses of the future will be a hundred percent built by robots. The fields will be tilled by robots. The only restriction on the physical robots is that there's only 70 to 80 million cars and motorcycles built a year. So we need to get through that 70 to 80 million. But in 10 years, of course, there'll be billions of robots. You know, like, what are the functional components of a robot versus a car? It's easier to build a robot than it is to build a car at mass scale, right?

32:17And China will be churning these things out across the world. What's the functional requirement to build an accountant in terms of energy cost? Now you build it once and you hit replicate. And you call into Andy, the accountant, who's the best accountant in the world and never makes a mistake. Why would you use anyone except for Andy, right? But there's going to be a tremendous fight from a lot of these professional lobbies, right? You don't need surgeons. Robots plus AI equal better. You don't need lawyers. You don't need accountants. You don't need any of the professional services, which are knowledge scarcity workers where they price accordingly.

33:01That's all gone. Well, I mean, so yesterday the deal was done with the Dock Workers Union in the US, taking it up to, I think,$65 an hour,$200 ,000 a year for some Dock Workers, Longshoremen Union, and forbidding automation in Docks. Wow. You know, you remember these kind of strikes. This was the kind of ultimate thing. So you can have that regulatory impulse, right? where you forbid this technology from being used. That doesn't end up well, usually. And again, this is a direct transfer of capital. It's basically a handout, right? Because you're intervening with market forces. So we could have that.

33:49We could have this resistance of lobbies, like, you know, I want artisanal humans operating on me. Yeah, I mean, I think all politics is going to go away from capital versus labor and go towards humanists versus transhumanists. It's acceleration, decelerate. That's the new political class. But from a game theoretical perspective, can you afford to be a decel? This is the question, right? Or decelerate? I think people will because they, the dock workers, I mean, what they've done is basically nailed themselves and nailed their own coffins for the next five years of no robots. well again it's an extraction tax right that occurs from certain parts of the economy to others and we're subsidizing them because we're paying higher prices and the higher prices are going to their pockets so the interesting thing here is like now you see doge right department of government efficiency they're applying ai to everything you know like i'm building a team of gpus or whatever how long will it be until every single government decision is checked by an ai just like it'll be medical malpractice for every medical decision not to be check fighting it's inevitable right yeah but then that makes it much harder to have some of these outcomes like a lot of the fat being cut from the current government workforce is because of incompetence bureaucracy other elements right these paybacks inertia is the other big one and actually yeah like uh they claim christian afoso and jamie wrote this book um i can't remember the name but they had this concept of you know predictable corruption as being okay for emerging markets because like attacks and then in the west what you have is embedded corruption like why does it cost five billion dollars to build a ground zero rail station or the train between la and san francisco it's embedded corruption effectively right ai can see through that.

35:44And so the question is, what moves faster? And one of these things is who moves faster? So because what you'll have is nations become competitive, so your comparative advantage disappears. Like if you want to get a job done, you can outsource it to India. So like people like Emad, you say that AI is better than any Indian programmer. Why are you negative in India? Are you shorting emphasis? I'm like, of course, I'm going to short emphasis. They'll just replace all the Indians with Indians plus AI and their profit margins will do that. And they've got all the sales channels right so i think that certain parts will be protected but those parts that are protected are the parts that have driven inflation in the u.s if we deconstruct inflation it's not in the market competitive areas it's education real estate health care yeah these areas right and we all know those can be slashed on the administrative side we know that you can have cheaper health care with ai cheap education today a better education with ai better health care with ai i don't think people are going to put up with the rubbish they have like why does america spend more than anyone but have worse health care outcomes people are done you know similarly education they're done the other thing is we're both financial markets people as well i mean what is the role of of investors in a world of a gi most don't see how what edge we have yet maybe long-term macro connect the dots views is more difficult but it's only a matter of time but for anything else it's pretty fucked right what is the financial market at that point well on a short-term basis you know humans have been out competed by ai computation for quite a while now right yeah yeah um but then one of the things is i remember when i was a hedge fund manager one of the things that i was taught that stuck with me is and i'd follow the process no matter what Because performance comes and goes.

37:41But if you say you've stuck to your process and the allocators have allocated to that, it'll help you out. And this is what we see a lot. Like, look at poor Hugh Hendry, right? He was the China there, then turned China bull. It wasn't good for his assets because he was there as the asset allocation choice, you know? And we've seen this over and over again with our colleagues. Like, stick to process. AI is really good at sticking to process. So that's number one. But what humans like is that human-to-human connection. this is the private banker type experience, right? Like I trust you and the bug stops with you when it goes there.

38:14So I think that we should expect finance to become more efficient. We should expect finance to have a dual track where you have AIs working with each other for credit and other checks. Like if you have a commercial bank that's fully AI driven with smartest AIs in the world, what are their NIMS going to be? What are their like ROAs going to be? It'll be better than anyone, right? And everyone will have to upgrade to that. So we should have more efficient, stabilized finance, but it'll be slow, then fast to adopt. Within a market sense, you will have to compete more and more with these AIs. And one of the dangers is things like false news or slightly false news, right?

38:54As it interacts with that. Like, how do you compete against that? How do you know what is truth from false? As you have people saying different things, because we start to measure these socials and other things right now. and we're not sure what happens then uh particularly when you might have concerted attacks like you'll have ai market makers probably you'll have different types of things because the money is there but at the very least there is no excuse now for any active manager not to follow their process and use ai right now although they are resistant to doing that most active anything with a process can be just done by ai and what's the point of having the humans doing it right so therefore have the humans compete well so with anything that's process driven i think that you have different stages of this right you have the eventual element here again like you look at you get the super enhancement first you have the super enhancement in that there's no excuse to say i'm not using ai to make sure that your money is investable into this because inherently what you have is known games and like these unknown games Markets are unknown games.

39:55And allocators are what matter here, right? Like if you are a fund manager, who do you go after? Are you going after retail typically? No, the big ones are the pension funds and others, and they follow a Yale-type endowment strategy where they're allocating different buckets. That's not going to change because of AI necessarily, right? They'll be allocating to managers who they can blame if something goes wrong, if they don't follow the process, because they want them to operate within this discretionary built portfolio. they will give money to managers i think though that employ ai first and then eventually most of the premium of the market might be removed i think that's the danger right yeah yeah and again it'll be removed by people who rather than charging 2 and 20 are charging 0.2 like your robo advisors your robo advisors as good as jim simmons you know yeah you will just have your own agent that does your investing and you tell it whether i don't even know how returns are going to work whether everything just mirrors gdp plus or minus some risk or i don't know well again like what is the nature of capital right like it adjusts um because it's not a index on productivity of a country necessarily anymore right no like again we i find it difficult to see inflation where we're going yeah that's a fundamental thing so we're in a massively deflationary environment with high amounts of unemployment or underemployment which means there's lower aggregate demand for physical things versus digital things and that's not a pretty environment full stop right again companies the large companies and capital holders can accrue even faster.

41:41But yeah, it's difficult to see where that goes because then your base is basically capital preservation more than anything else. And the other flip side of that is because AI can build AI, it can build businesses really fast. It can copy your business. You start a business. I want to copy your website business. My agent does it for me. I launch. It destroys your business, somebody else comes along, destroys my, I mean, capital formation and destruction is going to be violently fast. Yeah. Jeff Bezos said your margin is my opportunity, right? Yeah, exactly. Because he employed cashflow and float incredibly well in Amazon.

42:22That's why there was no profit. Amazon, he could have made profit. He's like, I'm using my float and my cash flows to crush everyone else. Right. Now it's your humans are my opportunity. Like you got a business with humans. I got a business without humans. And again, once you have the ability to have that AI agent that I'm talking to you like I am now, and I'm instructing it, and then eventually the AI is the other side instructing me. How will you ever know the difference for any non-physical based thing? And then physical comes a few years later, once we can build enough robots. And then what are we?