Decoding the Gurus - Jensen Huang, Part 1: AI Will Fix Everything (Probably)

Episode Date: October 28, 2026

Jensen Huang is so hot right now; not only did he make some statements that were widely derided, but he also featured in a glitzy "collab" episode with our favourite lone wolf, Lex Fridman. Sensing guru vibes, we turned to take a deep dive into his recent interview with Ezra Klein. Huang is the co-founder and CEO of Nvidia; he also really likes Lex Fridman and the All-In Besties. Trump even called him while he was on the All-In stage! So it's fair to say we went into this one expecting yet another bloviating tech CEO and were ready to join the dunking. We do not do guilt by association, of course, but bad vibes were in the air.Huang thinks AI will transform the economy, take over tedious tasks, and leave us free to do more interesting things. Some jobs will disappear, but new industries will emerge to replace them, like wellness spas... It's okay for those whose jobs are disappearing, though Jensen has a solution: they should become systems thinkers!Alongside such high-level insights, Matt and Chris do find some points of agreement, including his impatience with apocalyptic predictions being presented as scientific findings. The discussion also highlights a difference between our dear hosts: Huang's irritation with people talking about software like little human minds makes Chris increasingly sympathetic to him. Matt, as a committed pan-psychic, is less impressed by the reassurance that engineers understand what they are building, and thinks understanding the components is rather different from knowing what the whole system will do. Inevitably, this leads to a discussion of dogs, ants, and what "unpredictable" means, because apparently we are sensemakers now.We will get to the fuller discussion of AI safety in Part Two. Until then, listeners can enjoy the experience of us listening to one of the less terrible tech billionaires and disagreeing with each other over minor points!Source materialThe Ezra Klein Show: Jensen Huang Thinks A.I. Alarmism Has Gone Too FarThe New York Times: Episode page and transcriptLex Fridman: Taiwan: The Most Epic Night Market Experience w/ Jensen Huang (the Taiwan visit mentioned in the introduction)All-In Summit: Jensen Huang: The Doomer Hoax, Superintelligence is Here, and The Future of AI (ft. President Trump)Creative Destruction Lab: Geoff Hinton: On Radiology (2016)Kurzgesagt: AI Just Crossed the Terrifying Line - Now What? (previously titled The Moment They Lost Control of AI)Relevant articlesVaswani et al. (2017): Attention Is All You NeedAnthropic: When AI builds itselfMETR: Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incidentAnthropic: 2026 Usage Policy update (including its policy on abusive treatment of models)

Hello and welcome once again to Dakota the Gurus, the podcast, where an anthropologist and a psychologist listen to the greatest minds of the world has to offer, and we try to understand what they're talking about. Matt Brown is me, the psychologist from Australia. Chris Kavanaugh is him, the anthropologist from Japan. I muffed that up, Chris, but I tried to recover, all. That's all. That's what happened. Don't laugh at me.

I feel like, you know, we should, maybe humans be replaced by AI. can't come soon enough that. We're too prone to error. To error is to be human, right? It's not the thing that people say. Yeah, yeah, we hallucinate too. We do lots of stuff. Yeah, well, that was, you were presaging the topic

because people have asked us to cover the AI CEO moguls who, you know, you could make an argument that become guru-esque. We've already done Elon Musk. So why do the rest? Why not it? And he was a complete prick. So, I mean,

he was pretty bad, Matt. Actually, don't you remember that a while back, Matt, we actually had a tech season. I don't know if...

I do remember that. Maybe this would jolt your memory. It's the coding the goof. Tech season. Tech season. Tech, Season. It's the coding the guru.

Tech, tech, tech, tech, season. Would you believe, Matt, that was made not by AI tools? That was actually handcrafted by myself. Hard to believe. Really? That seems unlikely to me. Did you really make that? I need that.

Yeah, I made every aspect of that. Why? It's so professional. Well, it was bad, but I assumed it was made using a... You were using some kind of tool to make it. Come on. Well, I didn't make it with my hands. You're right. But I did paste together all the various things and copy and paste stuff and do all that.

So that was pre-AI audio technology you heard. That's how long this podcast has been running. It's started in the pre-AI. In the older days. Look what they've took from us, Matt. I know. We started this podcast. We were harvesting organic food with our own hands, lovingly crafting, our clothes with

a loom. The world was good. And now we're in a AI dystopic future, or at least we will be soon. And as you've mentioned, Matt, this is. is relevant to who we're going to cover today because we're looking at Jensen Huang, CEO of Nvidia. I had seen him before, but he was receiving some attention recently because he did an interview with Ezra Klein for the Ezra Klein show.

That's the material that we looked at from just a month ago. It's topical. And oh, I should also mention that I went into this with a very negative impression of Jensen Huang because I had also saw that more recently than that, he had just done an episode with Lex Friedman, visiting Taiwan with Jensen Huang, and it's basically him and Lex traveling around Taiwan, being treated like celebrities, eating food. and getting cheers from crowds and Jensen seems to really like Lex. So this is a really severe indictment of his character.

Yeah. And so the only other thing I knew was that he'd been like slammed recently for this interview with Ezra Klein. So I went into, you know, like a shark, right? There's blood in the water. I mean, it's worse than that though, isn't he? Because isn't he Pally with the all-insurial? podcast guys. Oh yeah. Well, I also saw that. I've actually watched that. He was on the stage

with the all-end podcast people and perhaps the cherry on top is that Trump calls him on the phone while he's there. And then he puts him on speaker phone. So, um, you know. So look, you know, we don't do guilt by association on this podcast. No, we don't know. But that's a lot of guilt like original sin in the Catholic doctrine, you know. Yeah, so I can't say that I went in a prime to have a positive reaction to Jensen Huang in this interview. We'll see. We'll see how it goes. So I was expecting another bloviating tech CEO.

And that's what the coverage promised me. But is that what was delivered? we'll have to see Matt, we'll have to see. So now Matt, before we get into the decoding, I just want to remind people, you know, you and I, we talk about various people, various gurus waffling on about topics that we are not experts or like, you know,

we're not physicists on par with Sean Carroll, but we can recognize when Eric is bullshitting. And I certainly don't claim any expertise in AI beyond that I meet use of it a lot, probably more so than most people in terms of research and other things. You on the other hand, Matt, in this case, I don't want to flag up to people, that you do have relevant expertise in career background here. So what is that?

And going forward, people should weigh our relevant comments accordingly. Like, take mine things as what they are, and maths are based on what? I love it, Chris. They're very humble on your part. I appreciate it. Thank you for frustrating myself. Look, I'm obviously not on the,

I don't have the level of insight that someone who is currently at the cutting edge working on this technology, especially in the current labs in China or the United States. But, you know, in the last couple of years, I have supervised a PhD student who worked on AI interpretability. And that was an important topic. We published several papers.

And yeah, I've got like, You know, 30 years of work that's been involved in statistical pattern recognition, computational statistics, including working on early versions of deep AI, Jan Likun's convolutional neural networks for robot vision. So I've probably published about 15 papers on the topic. So, you know, that puts me in a position where I can at least, I can read the papers and I can understand. Yeah, I can read the research and I can understand how the algorithms work and what the advances were since back when I was in the biz and what's happened since then. And I also can appreciate the issues of play there, which is, and it's quite common in mathematics and physics where you can actually,

you can understand the math, right? You can understand the principles, but that doesn't necessarily mean you can easily wrap your head around the phenomena, what's going on. People have said this about what you can understand the equations behind quantum mechanics or general relativity, but having actually a good mental intuition for what's going on there is actually a separate matter. And the same thing is certainly true in artificial intelligence and artificial neural networks. You can understand the architecture. You can understand the algorithms that we use in terms of gradient descent and the reinforcement learning and so on that's going on there. And that will help you or put you in a better position.

But I'll just point out that it's challenging for anyone, even someone with all of the technical background and who is currently working in it to have a good intuition for what's going on with such large, complicated AIs that we have going these days. And people who are in the biz at the moment will tell you exactly the same thing. So it's an interesting and challenging topic. And it is actually one where other points, of you from, dare I say, psychologists and even philosophers, people with a slightly more

general background, yes, I know, can assist, right, because the answers do not lie in the math. The answers lie in terms of understanding them as the complex systems that they are. So, yeah. Interesting. And I'll just add, Matt, before we move on, that I have various discussions with you regularly. You know, you're lucky like that. and you are consistently talking about papers that you've read on the topic and you

research this game out. So from my perspective, you talk about it like somebody that is following the field pretty closely. So I'm just team up to the listeners that, you know, Matt does follow this probably a lot more than most of the people in our audience and me. Okay. So factor that you can still disagree with him. I will.

Don't worry. It's going to come up. But yeah, he's not just ruffling from points of complete ignorance or discourse that he saw in the news. So just saying that, Matt, we can move online. That's right. We've all just got opinions. It's okay to have opinions. It's okay, yes.

But some are more informed by relevant information than others. Okay. Is there anything else you'd like to mention before we get started? Any more housekeeping? Any more business? No, no. Let's get right into it. I've just literally just finished listening to this interview myself.

I have thoughts. I think it's connected to some other interesting topics on AI safety, which I wouldn't mind inserting at some point. But let's see what Wang has to say for himself to start with. Yes, and we'll start at the start of the interview where there's the usual big, free-mean question. So, yeah, this first clip is Ezra asking him to outline, like his general views on AI, and he has a kick analogy.

that covers a bunch of grinds. So let's listen to that. First of all, it's a new industrial revolution. And this industrial revolution, this industry requires production. It manufactures things. I know that in the end, when people experience it, it's a software product, but it requires energy, the chips that go into these data centers, these AI factories, the next layer above it is basically the AI factory,

what people enjoy as infrastructure or cloud services. And the layer before, above that is the models. And the important thing to realize there's language models, but there are models of all kinds of chemical models, biology models, physics models, articulation models, robotics, navigation models, self-driving cars, all kinds of different types of models.

And then above that is the most important layer and the layer that I care most about that our country takes advantage of is the application layer. This is, you know, applications for legal services, for health services, for manufacturing, so and so forth. Every single industry is involved. So I want to go through this, but I want to go from the top down. Because as you're saying, the way people will interact with it, the way it will or will not change their life, is that what you call the application layer. Okay, so there he's outlining, I guess, his model of how to understand this.

I think too controversial yet. I think the stochastic parrot advocates have generally considered defeat, and most people do agree that AI will be pretty important, perhaps on a part with the Industrial Revolution or the Internet Computer Information Age. You might call it the knowledge revolution, perhaps. So maybe changes are going to be big, so I think most people agree on that.

And then he's got the sort of stacked layer type model. where you've got like chip fabrication and stuff at the very bottom, and the actual application layer where you're actually using it to do whatever. Someone might be using it to do some statistical code, for instance, to take a random example, or they might be using it for radiography or something. Did you have any problems with that, Chris? No, and I think you're optimistic, though, that the stochastic power people have conceded that it is going to be as transformative

as he suggests here. I think there's still a lot of talk about, you know, It's a bubble that will ultimately collapse like a flan kick that has not been well constructed. Nice. I like how you worked with the Kakenilogy there. Yeah, I've had that. But the point about another industrial revolution and that the thing that people interact with is mostly LLMs, but there are other models that are important right than doing things.

I think that's pretty good. And then he's asked about his vision. Where's this all going to take us? So let's listen to his vision. So let's start with the vision. What is the world you're envisioning? What is possible that is not possible now? What is common that is not common now if we get that layer right?

200 years ago, we were able to power anything and everything, electricity. And then, I guess, 40 years ago, 30 years ago, with the Internet, we were able to find anything. Today, or soon, we'll be able to know everything and do anything. And that's the concept that's really quite exciting. That out of the ether, instead of doing search and then going through, you know, one link after another link reading all these different websites trying to figure out what's going on. In the future, you just ask it a question and it comes back with an answer. You give it a project, comes back with a solution. You give it a task, it comes back and gets it done, you know. And so,

and it comes out of the ether, comes out of the cloud. And that's the, that's the magical thing. I feel like the future, the way you're describing it there, what people have experience with is the chat bond, right? They can go and ask GROC or Claude or chatGPD a question. But the applications layer works in a much more industrial way. It's in hospitals. It's in schools. I quite liked the analogy there with electricity because I think a good way to think about it is like a kind of crazy glue that you can throw a lot of different things at it and get sensible things back. So just like electricity is like a general purpose type medium that can power a bunch of different applications, I think it is helpful to think of AI generally.

And as he said, not just LLMs, but all of the other kinds of AI as well, which have, it's the key characteristic of it is it's not deterministic and it is very flexible in terms of inputs and outputs. So, yeah, I find with that as well. Yeah. So I think versatility, right, there's the point here. And it's sort of in the same way, like what's the internet for improved communication and accessibility maybe, but like the actual implementations of the internet are endless. You know, it can do any number of things.

But electricity, you're right, is probably a better analogy because what is electricity for? Many things. And the people that regard AI as like purely an advanced Google search, I feel like they're not quite appreciating. yet quite how versatile the technology is or useful in domains outside that. But that's a common interaction. And obviously, I realize now we should probably have added our distiller at the start that, you know, people listening to this will know that you and I are more in favor of AI that tends to be popular amongst lefty type people.

Oh, we make no secret of that. So if you are looking for us to be like, oh, no, it's completely used to stochastic power. You're going to be very disappointing, but you should have accepted that by now and know that that's where we will fall. So the end there was

they brought up the issue about it being applied in hospitals, schools, industries, and so on. And this turns into a discussion about potential threats to jobs and professions, right? And there's an example that Jensen

brings up and as our response to you. So listen to this. So radiology is an example I know you like to use. So the thing people worry about the applications layer is that what these applications are going to do is replace human beings. And radiology has been a sort of interesting example used on both sides. And I hear you talk of it often. So how has the entrance of AI-aided radiology shifted radiology as a practice? Well, the thing that's important for all of these is to recognize for everybody's job. There's the purpose of the job, and then there's the task you do as the job. And in the case of radiology, the task, and it consumes a lot of their time, and they

sit in dark rooms doing it a lot, which is study these scans. Now, if all of a sudden the studying of the scan is done automatically, it doesn't change the purpose of their job, which is to diagnose disease, help doctors, do more scans, ultimately help patients figure out what's wrong with them. And so the fundamental purpose doesn't change. The task of studying that scan has become automated. And so as a result, radiologists are actually able to do more, handle more cases, do more scans.

Hospitals are able to process a lot more of these patients, and therefore their revenues go up. As a result, they need more radiologists. And so this flywheel is happening because the pipeline of patients is quite large. And so where else do you have this problem? Well, let's take a look at software engineering. People said there was a prediction that literally by this year that 90% of all software will be coded by agents and therefore we don't need any software engineers.

And so the question is from that, ergo we don't need software engineers. That last part is completely false. That's completely wrong. The purpose of the software engineer is engineer. There was engineering before software. There will be engineering after software programming. And the purpose of engineering is to invent something new, discover a new product, create a new product, solve a problem, connect a social need with the technology that

exist in the in the in the in the in the in the in the in the in the in the in the in the and so that mission that purpose doesn't change yeah yeah I I find this framing appealing I'm like it just gels with a lot of my personal experience in life Chris because it's an observation that's not restricted to AI but with all kinds of tools and I'll give you a very random example like when I worked in construction we were design complicated, industrial and commercial staircases, right? And they were often very complicated and very heavy. One of them was like 24 tonnes and they'd be in spirals and all kinds of stuff. So we needed engineering. And when I say engineering, we needed an engineer to assist with

the design and ultimately sign off on it. In practice, actually, we did pretty much the design and we wanted them to sign off on it. And in the old days, an engineer would do a lot of calculations about loads and forces and, you know, all the stuff. Kids learn at school, right? But way back, like 25 years ago or 20 years ago when I was doing this, I was using Autodesk Inventor to do 3D models of the stairs. And Autodesk Inventor does all of the calculations to calculate the stresses and stuff

on all the different volumes and substance materials in the model. So you could put it under whatever stress and you could see whether the design was working or not. So like, you know, I haven't checked in with how engineers do their job these days, right? But I presume that they work extensively with tools like that and the kinds of hand calculations that were required and the kind of knowledge of that lower level sort of physics,

or applied physics, I suppose, is no longer needed. But there's still a lot of engineers. I don't think auto-desk inventor led to job losses in engineering. But certainly they became more efficient and probably more reliable too. Yeah. So, I mean, I find this rooming also fairly compelling in terms of the purpose of a job being something. You know, to educate students, for example, and academia, like to pass on skills.

But what the content of those skills are change over time. The things that you were teaching students 100 years ago are now very different than the skills that you were teaching before. Not in some big picture ways, right? Like you're still trying to get people to learn critical thinking and how to structure things and all that kind of stuff. But I mean, just around stuff like word processors. There generally isn't much instruction and typewriting.

at now. So yeah, I think that point about the core of a job being the thing that you care about is relevant. Now you can quibble with the claim that it will never or not fundamentally lead to job loss. But that basic point, I think there's little to argue with. Yeah. And look, I can think of objections, of course, and listeners are probably, at least some of them having pop into their head right now, which is around the way. in which AI is arguably very different from a calculator or autodesk inventor or the internet or other kinds of tools in the sense that it is, as Jensen says, so flexible, so multi-purpose like electricity, it may well subsume such a large component of the human factor, the thing that

humans used to bring to the table, that it should be thought of differently. But we can come to that. We can hear him flesh out this point a bit more, and maybe there'll be opportunity for questioning these kind of thing. So this is a bit more in the job transformation point. So I think the fallacy, and now it's, you know, because of some of the narratives and some of the storytelling, has turned into myth and it's harmful, is that AI will destroy jobs, which is fundamentally wrong. It will change every job. It'll change every job. Many tasks will be automated. Some jobs, where the job in the text and the task is really one,

meaning customer service on the phone, in a lot of cases, that job is precisely the task. And so in those cases, it could be automated away. But oftentimes what you'll see is this new industry, a new technology actually creates a whole bunch of new jobs. And here's the proof. Here's the proof point. And so in the last six months,

AI has become, if you will, useful, the inflection point of AI. Previous to that, we spent 15 years trying to make it work. All of a sudden, the last six months, it became useful. So, I mean, this is an incredible statistic. In the last six months, $500 billion of venture capital has put into the AI natives. And the reason for that is because they now see the potential of this new capability and they're going to create a whole bunch of new companies. Jobs are obviously being created from $500 billion.

new investment. And so all of this is all happening right now. I can think, Matt, that one of the things that people will respond there is that sounds like a bubble where there's a huge amount of investment rushing into a new technology. So Jensen is taking a very optimistic view that that's going to be great. Well, from Jensen's point of view, it is great. Everyone is buying a a lot of Nvidia GPUs. I mean, this is one thing to know, and this did not pass us by, right? Just to be clear, Jensen is the CEO of Nvidia, which reads the chips, which are powering AI.

So the more data centers and compute that is required, like, the better economically for Jensen, right? So obviously he's going to be pro, give the pro case. But what do you expect? So I feel like obviously that's the case. You have to factor that in to everything that he's talking about. Because you should not expect that he's going to be like a donor on the technology, right? Like, yeah.

Yeah. Yeah, sure. So that's his motivations, obviously. But, yeah, I mean, I think also that it conflates a couple of things, right? People have said this before, but I'll say it again, that obviously you can have a technology. and infrastructure that is transformative economically and socially and still be a massive bubble that has a massive crash that is terrible for anyone who actually invested in it.

The railroads are the prototypical example of this, but it's also true actually of the internet, the buildout that they did in terms, mainly I think it was in terms of the undersea cables, but the various other types of infrastructure that was put in place. They took massive losses on that after the dot-cap bubble collapsed, just like the railroads. It's great for the rest of us because it was that infrastructure paid for. So it's very different, very different questions.

Well, I also think it's worth remembering. Like there was the dot-com boom, right, and the subsequent popping of that balloon. But it's not like the internet wasn't transformative. There was a whole bunch of hype put into, you know, silly websites and that kind of thing. But the internet itself, as a technology, did lead to a whole raft of new companies, transformations, new kinds of jobs and social problems and online influencers and arguably half the people that we cover as a result. But yeah, so I'm just echoing your point that you can have a bubble and a transformation.

The two things can go together. But, okay, so this clip is still on the same topic, Matt, the new jobs that he got in trouble for this, or at least I saw this clip passed around a bit. So I'm kind of curious if you can pick up what annoyed people about it. So he's still giving his optimistic kick around job creation. But let's see if you can spot what got him dunked on. I believe that we are going to see jobs change in mass. I believe there's going to be a net creation of jobs.

and so let's let's um there's you know listen there's a whole bunch of industry that exist today that didn't exist you know halfway through my life people talking about um wellness centers and spas and you know all these different entertainment and luxury industries and quite frankly the whole entire luxury market didn't exist i think we're just going to have new industries that's all But overall, overall, there's no question in my mind that because of human ambition, that's really the fundamental missing ingredient. That's, you know, people look at this work. This is the amount of energy that goes into it. We're going to, this is the amount of work that goes into it.

We're going to insert this work automation system. And as a result, the amount of work that's necessary is now going to be reduced. And therefore, you know, some jobs will be gone. I believe that's flawed because there's a piece of input, the human input, this intangible, it is not in calories, it's not in jewels, it's ambition. And I believe the power of ambition is the greatest force, in fact, and it's missing in everybody's calculation. But for a lot of people, yeah. But for a lot of people, their relationship to work is not powered by the kind of ambition that led you to create Nvidia.

It's just a different ambition. It's an ambition to make their children's lives better, to take care of their family, take care of their parents, ambition to be, to be able to travel. These are all ambitions. That I agree with that. So what he got into trouble with that because he sort of cast everyone who's affected as they should think about it. like he thinks about it, like some sort of entrepreneurial master of the universe that dynamically adapts to their job being replaced and their skills being obsolete and things like that, right? Am I, interesting?

Well, that's a little bit too high, bro. He mostly got in trouble because the industries he pointed to were wellness centers and spas and luxury goods, right? Oh, okay, yeah, yeah. Yeah. But, okay, I mean, let me, let me, can I let me steal men, Jetson, a little bit here. I guess when I, and I hear things like that, I make assumptions about what it is they're trying to say. And what I assumed he was trying to say there was about the point about creative destruction, right? So there are there are heaps of jobs today that didn't exist 100 years ago, simply and are accessible to just everyday people, everyday people use them. It would have just seemed to be an utterly frivolous luxury to people a hundred years ago.

So barista made coffee or we had a lovely young guy come around and barf our two dogs, right, and give them a full trim and stuff. God, they saw we needed it. We couldn't do it. You need the equipment and everything, right, because they hate it. But that kind of thing, like a mobile dog grooming service, right, which wasn't that expensive. But, like, those, you may not see the connection. But the reason that those jobs exist and demand for them exist is because of these other things.

Exposable income. Yeah, and like things that happened a long time ago, like the agricultural revolution, the industrial revolution, the various other types of automation and so on, means that there is now scope for very different kinds of jobs. That's where I assumed she was getting up. Oh, yeah. Well, actually, I wasn't, your still mind is unnecessary because I was just going to tell off the people. who criticized them because I think your criticism is more valid than the one that Desra Klein is kind of responding to, right? And I think his response is pretty good to say, no, but I'm not talking about entrepreneurial

spirit. I'm just talking about, you know, human ambition in a broader category. But the fact that he brought up like wellness bars and centers, he's not endorsing that as like, that's the thing that we all need to move into. He's saying that wasn't as big of an industry when he was. young, right? And he's correct. Supplement companies and all the things that the broken sphere are doing your cold plunges and your Huberman portfolio of products.

These are all indications of luxury, some would say, unnecessary products that are now, you know, multi-billion dollar industries. So I think that's the point that he was making that like got a little bit varied by the fact that he referenced wellness centers and spas, right? You could have referenced something which is more significant to people's lives, but it is just in general that there are new industries that we didn't think of before.

There was no YouTube, no YouTubers a couple of decades ago. Now that is a viable career for a whole bunch of people. Whether that's good or bad is a separate thing from the fact that it is simply true that that exists right now. I'm just glad you're on with YouTube as an example and not podcasters, but yes, I take your point.

That also exists, doesn't it? Yeah, we should bear that in mind. Okay, so in any case, he here is coming across as a bit of an optimist, Matt, in case you ever notice. And Ezra does point that out to him. And so the lessons of the past

that we're taking some, that you're taking some comfort in, they should actually make you more, not less worried about the future. I'm always worried about the future. That's why I work so hard. But I'm a, if you will, responsible optimist. I have great responsibilities. I take my work extremely seriously.

There are a lot of things that can go wrong. We're pushing across every layer of the technology stack, everything is hard. But it turns out that's not society's problem. That's my problem. And for society, what they should know is this. We're going to build our company. We're going to build our technology.

I'm going to do my work so incredibly seriously that what they get to enjoy is my optimism. I'll do the same with my children. I do the same with my family. And I think that what we're going to do the same. want to do, I believe, is to put, to channel all of our worries into helping people be inspired by this technology and use it. Use it so that the technology doesn't just impact them, that it benefits them. The fear a lot of people have, 79% of Americans think AI will reduce

the total number of jobs. The fear is that the more serious you are, the more serious you are, the more serious Sam Altman is, Google is, Dario Amadeh is, that maybe the worse it will go, because the better AI is, the more it is a full replacement for a person, the more it has ambition in some ways that a person doesn't. You keep talking about ambition. I sleep. I want to spend time with my children in the morning. When I have an AI agent working for me, it doesn't.

It just works and works and works and works and works. Yeah, and so I think Jetson is. Like you said, he's incredibly sanguine about the potential risk. And I found myself reacting against that a bit at various points of this. And I think Gaila raises some decent points there, especially when, like, there are so many voices from those major AI labs. And people like Jeff Hinton, you know, big AI gurus and insiders from those companies, who are talking about things of the most apocalyptic terms. Sometimes a mixture of the singularity is coming

and the future is upon us. And also it'll be incredibly scary and terrible things will happen as well. The messaging is all over the place. So Jensen, as a tech CEO, he's kind of, yeah, nothing to see here. Everything's going to be fine. Don't worry about it.

he's actually an unusual voice. So people, I think it's legitimate to be worried if you are trusting the, in scare quotes, experts. Well, we're going to hear his response to Jeffrey Hinton's point in a little bit. But we'll hear his response to his answer's point as well here. But his take is, look, I'm not overly optimistic. I'm just recognizing that this is going to lead to improvements for people. And then Ezra's response is, well, but not necessarily improvements, right? Because people might lose their jobs.

Like, you know, he, Jensen already mentioned that customer service, maybe there are some industries where they'll just be complete replacement of the position with the technology. But I do think his view is sort of the much higher level view where he's like, well, sector of the economy will disappear. But it's okay because a new one will open up, which is probably true, but also not fantastically reassuring if you're somebody in the sector that's going to disappear. But the other thing is, and we at this point several times, it's like, nonetheless, that's

true, right? You can lament it. You can feel sympathy for people who are in that situation. But in some sense, you can't put the genie back in the box. If there is the possibility for technology to take over particular areas and it ends up that it's very cost-saving and stuff, it's likely that it's going to happen. Unless you put in protectionist regulations to prevent that from occurring, which perhaps some people would say is the option there. But nonetheless, you can't completely prevent it. Like in agriculture, you did, despite protections, end up with much less farmers and much more mechanized labor.

Yeah. Well, I mean, the way I frame it is that perhaps the real issue at hand is not whether the technology is good or bad, but rather, how does society adapt and manage the impact of the impact of. that technology but generally a technology just gives the potential to do something and you know the atomic bomb versus nuclear reactors or whatever to do with splitting the atom right the technology is neutral what matters is in terms of of how it's used and the analogy historically of course is the industrial revolution the power loom the Luddites the poor weavers who were put out a job right like people make fun of Luddites these days, or it's a term of mockery or whatever, but their concerns

were perfectly legitimate. In exactly the situation you describe, their profession was being made obsolete, and these machines weren't taking their jobs. Now, I think in the long run, absolutely, having power looms, having mechanical looms, is better than all of us weaving clothes by hand, because it just is, right? But the, I'm not going to get into why it just is. But, you know, the question is, you know, the same is true of many, a lot of stuff that happened in Industrial Revolution,

or you can take any other technological change as an example. You know, Walmart and, you know, Amazon, shipping stuff, whatever, right? It's bad if you've got the local hardware store or the local shop that sells whatever, right? your livelihood is getting taken away. This crops up in so many ways all the time. And that was a business model that was enabled by the Internet, another big technological revolution. But the real question is how society manages that transition

and whether it's handled compassionately or whether it's a kind of anarcho-capitalism, like the kind of thing that was arguably happening in Victorian times, right, where factory owners could basically do what they want, right? and the peasants of the way. Send them up the chimney. That's right, until their hands get mangled by the machines and then throw them on the dust team. So yeah, so it's really about managing it, right? And the one thing that struck me with Jensen that struck a sour note, which is that like a lot of tech CEOs, the way he talks about it, it feels like he has this faith in that

the corporations, just let them do whatever they want. And it'll always work out well for everyone. And later on he talks about a kind of compassionate approach to the community with data centers and stuff. And he says, oh, they should be building community centers and other things for the community. And I'm like, no, like corporations should be regulated and taxed, and they should operate within the bounds that a democratically elected country decides. They aren't the ones that get to make all the decisions for us. So that's where I started becoming kind of anti-Gensen. Yeah, yeah. Well, that we will get there. It's kind of, he seems to envision OCP, like him from the Rubble Cup franchise, except in a positive sign, like his images, that would be

great, right? The OCP is running the government and the housing and so on, rather than dystopic. But before we get there, though, Matt, he does, he does have a suggestion. And I can see why this kind of thing would rob people to the wrong way. So, you know, Ezra pushes back there saying, okay, so, but what about, you know, actual people who lose their jobs and so on? And Jessen does offer a suggestion to people in that situation. That coin has exactly two sides. Because the technology is so capable, it is also, and because it's so smart, it is also easier to use.

You are empowered by that technology more easily than any technology in human history. And so let me give you an example. You know, I create, I was one of the early people in this industry that created the modern computer industry. And this industry created a whole bunch of tools, the single most powerful tool in human history, the computer. But you have to speak its language. You have to learn a specialized language to do so. We can now make it possible because of AI.

Everybody can take advantage of this computer, use it to its limit, without having to speak a new language. Fortran, Pascal, C, C++, you know, every single one of those languages, Rust and every single one of those languages. Kuta, every one of those languages. And so now you just have to speak human. Tell it what you want. follow what your hopes and dreams are, what you're trying to achieve, and it interacts with you and gets the work done and gets that, gets on task now. All of a sudden, you have the might, you have the same might that 10, 15 million people up and out of 8 billion has. And so it's incredible.

And so my point is this technology is powerful, but it's also powerful in a way that is really easy to use. And so my point is, my point is, on the one hand, yes, there's the fear of just this incredible technology change and how quickly it's happening. But that quickly is translated in two ways. What I hear when I say to technology is happening quickly and therefore it should give me anxiety, that's one way to receive it. The other way to receive it is that it's advancing so quickly it's easier to use. so I should as quickly as possible, use the technology as quickly as you can, so that you benefit from this transition,

so you benefit from this new industry and not just be impacted by it. Well, what do you think about that kind of argument? He's basically saying that if you're affected, if your job is affected, then just get good at using the AIMAT. But yeah, it is a kind of learner to code thing, except he is making the argument that learn to code suggested that you need to be a 60 year old who now learn C++ or something. But his case here is all you have to be able to do is talk and to have ideas. And because of the technology, you now have the skills that previously were restricted to a very small. slice of society. You don't need to have mats or code knowledge anymore because the AI has

not just the R code knowledge, but all coding languages. So he is correct, isn't he, about at least that the ease of use is a different keyist than like learn to code. Yeah, I think he's correct, at least partially. And, you know, they cite the now classic example of radio, There's been no shortage of demand for radiologists. Radiologists are basically doing a lot more scans. The technology is doing the lower level stuff. And radiologists are still working, just being more productive, helping more people,

want more people getting scanned. Health is improving. It's win, win, win, basically. So there's going to be a bunch of professions where it's like that, at least for a long while. You know, in my own experience, as you know and almost everyone knows, I found it incredibly useful across the board in terms of my skills. It is quite sobering, of course, for me to appreciate that some of my skills, many of my

skills, including being good at R coding, applied statistics, a lot of the mechanical skills and even a lot of the judgment calls in terms of how things should be done, what kind of robustness checks should be done, the creation of tables. and figures and a lot of the statistical decision-making and implementation is all handled by an AI faster and better, frankly, than I could do it. So the skill that was important for a large portion of my career is basically obsolete at this point. Not obsolete yet. You're still there. You know, they still want you to teach students or whatnot. So. Well, No, I'm not saying, I'm not saying I'm on sleep.

Okay, okay. I'm saying that skill is obsolete. And, you know, like Jensen is saying, I essentially move up towards the apex of the skill hierarchy where I would like to think, I'm still contributing a lot of guidance and judgment and reality tests and also supplying, I guess, the executive oversight in terms of what we are trying to achieve and what's important and what's not important. and all of that stuff, right? So there's certainly cases where what he's saying is true, but I'm thinking of the cases where, like, what if your job is literally that? Maybe you're a research assistant or a more junior person

or you're working in a profession or a trade or whatever it is. Data entry. Some sort of desk work. Data entry or some sort of desk work, or you're at a call center in India, for instance. These are where the job layouts are actually happening. and you don't have the ability to just move up the kind of hierarchy

rather though like, no, no, that job is totally supplanted now and we don't need you, right? Those people are in a different boat. That's a solution. And yes, I do agree that there's that element that even if it were the keyest that everybody could become app makers or something like that, there's just not that much ambition in the world

or everybody in the twilight of their careers or whatever learn how to operate AI and so on. So he is, I think, being optimistic in that respect. Sadly, I can't relate to the thing that you talk about with like AI is doing things better than you because it just doesn't happen in my experience. Just, you know, AI's an assistant map. I can never release.

Yeah, no, I do know very, very well that. But, yeah, so, I mean, again, a bit sanguine. A bit sanguine. Yeah, very sanguine. And I guess that, harkening back to the other conversation we had about this, I mean, he and I just then were talking about the situation right now, today. But thinking forward, it's now not hard to extrapolate. If you think about the difference between today and a couple of years ago, in just two years time

I don't know if there's any more room at the apex for me to kind of shift into you're at the top of Maslow's hierarchy no I mean there's no no I think I think AI could subsume all the things that I do without too much imagination perhaps apart from some bare minimum of specifying

intent and priorities so yeah like, I don't know if his answer. I think a two-year time frame is pretty optimistic for that. But let's see. We'll find out in two years, I guess. But you can report back if that's the case. But, well, if you thought that was slightly holliamish.

This next clip, this is the one that got clipped everywhere. that he was getting dunked on. And, well, I have thoughts on that. So let's hear the dunker phone clip. So the message of this research out of China where you were seeing a lot of kids using AI to kind of help them was that when they were using the AI, they were getting things to invester.

But it turned out that the skills they were learning were not holding, that their actual personal performance, at least in the way we traditionally measure it, was degrading. Yeah. What do you think when you hear that? I think the last part,

I completely agree. Try to get a kid to do long division right now. You know, the multiplication table is starting to be forgotten. Doing square roots. My goodness. I mean, it's just basic math is being forgotten. Does it matter? That's my question for you.

Yeah, I don't think it does. I don't think it does. But there must be some set of skills that matter. Oh, yeah, yeah, yeah. But maybe not those. We're going to discover new ones. Just maybe not those. There are a lot of skills that don't matter.

You know, people don't, I mean, my first confession, I actually don't know my address. And I don't really believe that to be true. It's completely true. And Janine will tell you, and Lori will tell you. One day, I had to pump gas, and it was a few years ago, and they needed my zip code. And I panicked. I didn't know my zip code. I don't know my telephone number, but I forget these things.

I can live with it. So I think what he's describing is more like a rich CEO type situation rather than anything to do with AI. I don't know, I forgot my postcode of occasions that I'm not a rich CEO, right? But I do generally know the address. That would be not for the most of the time. But yes.

How many houses do you have? though, Chris, you'd probably go like less than seven? Less than seven, that's right, that's right. Well, yeah, and actual houses that I own, it's less than one. So, but places I live in, yes, those exist. But so you can see why this would lead to people dunking on them because, like, the clemen does know, like, I think the one about not knowing your telephone number, I don't think that's not rare.

I think a lot of people don't know numbers anymore because they're just relying on their phone book, right? At least of other people's numbers they don't know. The address one slightly more, slightly more out of touch. But I think the other part that he got dunked on for here was that kids today, they're not learning maps. They don't know how to do anything. But it doesn't actually matter. Like they'll just learn other things. So if it ends up that they rely on AI too much and they don't know how to do basic calculations,

it's okay because the AI can do about basic calculations, they'll learn all those skills. What do you think about that? Yeah, I guess, you know, again, it's true in some ways, right? A lot of skills that were really necessary before computers came along just aren't necessary now. They're like debating about whether or not they should still be teaching handwriting, nice, neat handwriting in schools so much because it's no longer a skill. And being, they still teach multiplication, stuff like that as a pedagogical exercise, as a stepping stone so you can learn higher than arithmetic, I mean.

But those skills of being able to reliably do long division is simply not necessary at all, right? This shouldn't be a controversial claim. Yeah, well, this is kind of my response. I thought you might be more upset because he's saying you don't need to learn, you know, some math skills. And you'd be like, everyone needs to learn those math skills. But no, it seems like you're not, because I kind of, I have to agree with his general take here, which is, do you think, you know, 100 years from now that kids are going to be learning long division and stuff like, no, I know, I know. I learned it, right?

I think there's this bias that people have that if you learned it as a kid, right, and it was hard and you had to remember the times tables or whatever. And then you find out kids today, they don't, right? In my case, the kids still do run, but the Times tables in Japan. But like, if they didn't, you're like, oh, gosh, see, what are they even doing today at schools, right? But like, the reality is, do I need the Times tables? Yeah. Right, right? In most of my life, no.

But here's where I think Jensen's kind of analogy is wrong, right? Because so this is all true, right? You don't need to learn the times tables, right? Calculators do it better. And actually, since Mathematica came along, even if you're working as a mathematician, you very rarely even need those basic calculus and algebra skills, solving equations of that level because it could be done computationally. pretty easy in the vast majority of cases.

Now, AI has come along and the field of mathematics is having an existential crisis because clearly a much, much broader range of mathematical skills are becoming less relevant, under threat, shall we say. So you could say, well, okay, just first learn the higher level skills, because that's what people need to do now. but to learn the higher level skills requires like bootstrapping up from arithmetic. This is what I think people will still be learning arithmetic in the future because you can't get to that point. Well, yeah, let me undercut an argument that I just made, which, because I agree with that, Matt, that like I'm teaching the foundational course in statistics, right?

and I'm teaching it using hand calculations and Excel formula, where people have to either do the calculations by hand with a calculator, or they have to learn the formulas and work out how to put them into Excel. And in part, those are skills, right, which will not be necessary. You could get an AI to give you the answer, in all cases, including the Excel formulas. But actually learning what concepts like variance or standard deviations or central limit theorem are referencing is very important before you get your AI to go off and do its statistical analysis, right? Because then you can understand at least to a certain degree what you're asking the AI to do. Right.

And I mean, this is just a foundational course. you need to build up more on top of that. And I think there is something to the point that, like, sure, the AI can go and run all these inferential statistics for you, and it can build very complex models. But if you lack the fundamental grasp of what you're doing, you can run them and you can get results and you can produce reports. But it's like wizardry where you actually don't know. The AI knows in a sense what it's doing, and it's probably doing it very competently, but you aren't in a position to assess it.

And if it is the case that all those skills fall by the wayside, then it won't be that necessarily the ability is gone, but it would be like a loss, right? That now there's a kind of black box of information that we are losing, And we end up in the dystopic future where the AI, we can't fight it because we don't know what's doing. Now, I think there's nothing intrinsically wrong in living in a world where you are using what are to you essentially black box tools. Like how many people really know how their iPhone works. You thought what technical details, right?

They use them exactly like their little magical devices. And they are. Yeah. Like function. Yeah. Unless you're a phone engineer. You know what I mean? They probably are magic to you.

Same goes with all the, there's like thousands of technologies bundled in there from the error compression, sorry, the error reduction codes, coding and stuff to handle. The Wi-Fi, there's so much. And it's analogous to the apocryphal example of the Boeing 747, right, which is there is no one living today, apparently not, as the story goes, that actually could really understand how a Boeing 747 flies, right? I mean, in all its details, right? Because there are so many systems, so many components, a massive log list, and each one of them is dealt with by specialists.

There are just, you know, just the sheer quantity of information of interconnected components, each with their own design principles and fail safes and, you know, whatever. Oh, the entire system. Like, we understand, it's not like the apocryphal, like bumblebee example. Oh, physicists can't explain. No, we know how planes, in principle we know how planes fly, right? Yeah, but the point, the point is it's just incredibly complicated, right? So a lot of people who are using planes, even working on developing planes, do not have their heads around the entire thing and are working on some little part of it. But I think what you're getting at there is the issue of disempowerment. So let's say hypothetically, and I think we're almost at that point now, you can basically trust. something like Claude 5.5 to just go away and do your stats for you. There might be some adversarial review by other AIs, backwards and forced and so on. But in the end, it produces a statistical analysis that you ask for, and it gives a robust and reliable result, at least as good

as someone like me, right? I think that's a very conceivable future, if not here right now today. So the question is, okay, let's say you're a social science researcher and you don't much care for statistics and you don't get very good of statistics, is that fine? Is that okay? Or is it disempowering? Well, no, yeah. So, but I think part of the issue is, you know, what are you doing in that case? Like, people want to do statistical analysis in order to find out like answers to questions, right? But for a statistical analysis to be meaningful, you typically have to, you know, it's kind of like the hitchhackers guide to the galaxy. Like, you have to know the question that you're asking, right, in order to interpret the answer that's provided by statistics. So, like, I can perfectly see that somebody would, like, not need to know the R code or how to do the robustness checks.

but they do need to understand what a regression is for. So isn't that supporting Jensen's kind of point, which is you still do need to be the high-level thinker in that in order to use the tools effectively? And actually, maybe there's a follow-up where he elaborates on this point in more detail. It's like it's the part of the clip that wasn't played where he elaborates the point.

So maybe we should hear that and then you can respond to the elaborated argument. And one of the skills I really value, one of the capacities I have that I really value, is an attention span formed on physical books. You're a big reader. I've read about the kind of reading you do. And there is prior to AI here, we're talking, a lot of concern and noticing among college professors and others that the way people use the internet has probably shortened attention spans. Some skills can be safely given away.

Others are valuable. They are capacities that are needed for that flexibility for that creative thinking, for that focus. It can't be the case that everything can be traded off. Yeah. Well, I think that we're going to lose some finer dexterity of intellectual dexterity. but we're going to be better systems thinkers. Today's engineers are far better systems thinkers than I was when I graduated from school. But I was much better transistor thinker.

What do you mean by systems thinker? They think large systems. You know, today's computers have trillions, hundreds of trillions of trillions of transistors in it. When I was first graduated, when I first graduated from school, you know, the first chip I, worked on had, I don't know, 200 transistors. I knew every one of them by name. And no engineer does that today. You know, most engineers now work well above the transistor, well about the functionality, and they're cobbling things together to do things. And so you need to think much more about systems and interactions of systems. Some of the lower level, you know, knowledge is gone.

Is that horrible? And so, so I don't. I don't know how valuable it is to know how to do for most people, to learn how to do surface integrals or partial differential equations or I don't really know how important that is, but it's important to some people. Yeah, yeah. I mean, I think I generally agree with that. And I think it's sensible. It's not important whether I agree with that, I suppose, isn't sensible, is it reasonable.

Yeah, I mean, that is the sort of history of humanity in a way, which is that people level up, they build tools of one kind or another that take care of a fine-grained level of implementation and then move up focusing on the bigger and bigger picture. But yeah, I mean, I guess the question is like, to what degree do different knowledge jobs have that scope to kind of go up in abstraction? And to what degree do some jobs just become not necessary? and also to what degree is there a pathway to get to the high-level systems-thinking abstraction from being like a university graduate. So one example I keep thinking of is that, like in the past,

I would often have a genuine need to employ research assistants to conduct a survey of the literature, you know, maybe even summarise the papers with a particular focus, or I might want them to do some data cleaning and, you know, basic, you know, preparation of data sets. There is just no reason, a practical reason, to employ a research assistant or even a postdoc of some kinds to do those jobs today. So that's all very well and good. From my point of view, I'd have to fill in a bunch of forms and find the money and get approval from the admins to get stuff done. But where's the pathway to become a research lead if we're just thinking about academia? And maybe with some jobs, there isn't kind of the equivalent of, oh, okay, I'll just move up to being a research lead or whatever.

Perhaps doing the data entry or working in the call center or handling the processing for the paperwork and so on. Like, that is the job. And there isn't a natural pathway to becoming a system thing. So I think even under the scenario of his very optimistic case, there could very well be a reallocation. Like we're talking economics now, not AI. There'll be a reallocation of demand, which is that like we've seen since the Industrial Revolution that the post-industrial society is devoting more and more of their economy to services. Yeah, broadly construed.

A lot of work, a lot of which is kind of desk work and mass work. and management and whatever you want to call it. Maybe there'll be a big reallocation where actually people's time and effort is better spent gardening or making houses better. We certainly have a... Well, we have a huge... Okay, let's stick with construction.

In Australia, we have a huge demand for accommodation. Everyone's complaining we don't have enough accommodation. Well, if we automate a bunch of desk work, maybe more of us can get back into construction and build the houses we need. Well, I know people don't want to do that. The director comments demand. But Matt's learned the code like become cement mixes or whatever. No, but I'm framing this.

I'm not being sanguine about this. I'm framing this as a challenge, right? Like I'm saying it could well be that overall there is less demand, not just for the lower level aspects of a job, but they could be not a logical place to sort of level up in being the orchestrator of jobs like yours, but actually the demand for jobs overall as knowledge workers in the service sector could decrease.

So I'm pushing back against Jensen's optimistic kind of view. Well, let's hear, because there's somebody else that pushed back against his optimistic view he has a response to it. So this is who you brought up at Jeffrey Hinton. Ezra raises him. So let's hear what Jensen has to say about his date. Jeffrey Hinton is on TV saying he thinks a 10% chance of societal destruction is not unreasonable.

I would tell Jeff that it's irresponsible to say all that. All of his predictions have been wrong. Enough predictions. That 10% chance is not grounded on science, is not grounded on research. It is, just because it comes from a scientist, doesn't make it scientific. Those predictions are hurtful. Let's take it of face value that the recommendation is exactly what he said, which is that nobody should want to be a radiologist.

And the world has no radiologist today. I think if you work as a radiologist, you're like the coyote that's already over the edge of the cliff, but hasn't yet looked down, so it doesn't realize there's no ground underneath him. People should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists, because it's going to be able to get a lot more experience. It might be 10 years, but we've got plenty of radiologists already.

Is that helpful or hurtful to the society? I think we can all agree. We can both agree. It would be terribly hurtful. It didn't happen. Is it good or bad that we scare young people about the future of AI so much so, that they don't even want to go to universities and don't want to go to college anymore

because they don't think they'll get a job. Is that helpful or hurtful if it were to happen? It's hurtful. Don't think for a second just because you're an alarmist that you're doing a social good. It is not true. So I think that we ought to just all be wiser, more mature, be evidence-based, be scientific,

if you want to be scientific, be scientific. Do the science. Do the science. But alarming people making claims that don't, they simply, their track record is horrible. Their trackers literally horrible. I don't know Jeffrey Hinton, Matt. You can tell me about him, but I will say I've noted that there's been a lot of predictions about there will be no service jobs, white-colored jobs gone in five years.

professors won't be necessary. And a lot of those doomsday predictions, well, in fact, all of them just don't materialize, right? They haven't materialized. That doesn't mean it never will. But I think it is notable, but all of the alarmist claims about what's going to happen to jobs, there'll be no software engineers, you know,

within the next nine months because they'll all be gone. Yeah. They're still here. So, yeah. Yeah. And I think Anthropic is famously overstaffed, right? So they're the, I guess, the heaviest adopter of AI, obviously. They have Claude basically doing everything that Claude can possibly do.

In fact, some more powerful internal model, I assume, that isn't publicly available. And yet, they hire heaps of people. So, yeah, and Hinton is one of the, you know, classic giants in deep learning and artificial neural networks in general. But I do agree that he is very alumnist and he was clearly wrong, as the gentleman says in the case of that profession. When did he say that? When was he saying like radio-

I assume it was a few years ago. But here's the thing, Chris. I mean, here's, okay, so here's why he was wrong, right? It's about supply and demand and elasticity. So the thing about radiologists, and I'd say health in general, is that there's basically an infinite amount of demand for more health, right? So we would all like to be healthier and there's kind of almost no end in which you can. Like if you can decrease your chances of getting cancer by just another fraction of a percent, you'll take it. Yeah, then everyone's like, yep, I'll take that.

Thank you very much. Right. So it's got that kind of relationship to demand. if you take something else, like the production of peanut butter or toothpaste, right? Now, that's become incredibly more efficient over the decades. And so they could now produce, I can have as much peanut butter as I want. They can afford it, right? Same as toothpaste.

But I don't, right? I just buy one thing of peanut butter maybe in a month or every couple of months, and that's it. It's because the demand has this ceiling to it. So if you produced, if before it took all of this effort, all of this labor to create the peanut butter, I'd still buy the same amount at that price. It becomes a lot cheaper. I'm still going to buy the same amount. So the efficiency basically turns into job losses, right? The people that used to make peanut butter will have to find something else to do because there is no extra demand for peanut butter. So a kind of a silly example, but my point there is that if you're taking any particular profession that is affected by AI or any technology and you're asking whether or not some huge productivity gains will lead to lower levels of employment in that area, then it kind of depends on the nature of demand. Now, I think there are some things like scientific research, like law, probably things like, you know, the people who

do fundamental analysis of companies in figuring out whether something is a good buy or not, right? like it could always be better like there's an infinite demand if you're like for for these things and if it gets done more efficiently you'll tend to have the same number of lawyers the same number of financial analysts the same number of researchers as before you'll just be it's the abundance thing right they'll be producing more stuff and it's and it's a win-win but if the professional the industry looks more like peanut butter manufacture in which case it's like we just need this much of it we don't really need any more of it then you're going to see job losses well there you go so you're half on board with his bashing of jeffrey linton but you want to pull back a bet and say but you know he's

not completely wrong yeah i just i just you'd have to do a survey of all of the different workers that are affected by AI and try to figure out how the demand would respond to the increased supply. Well, something for another, had another more competent economists to do than us. But I'm going to give you an option here. There's a fork in the road ahead of you. Okay. On one side, there is Jensen talking about the language that people use with AI and his feelings regarding my own. some cautions about anthropomorphizing. On the other hand, you have his response to recent AI safety concerns and how valid he views them.

Which way Western man? I don't know. Let's take door number one. Why not? Language, language. Okay, well, this one I am more aligned with Jensen on. So let's hear some of the calls. questions that he has for people talk to you about AI.

You're creating something, some entity, an agent that is smart, that is capable, that is relentless, and who the workings of its mind, we don't really understand. The chief scientists at Open AI, the chief scientist at Open AI have this. You know, Ezra, look, I just don't want you to contribute to that. Software's not. You're worried I'm getting off the. I don't think software's relentless. Aren't they trying to make it very persistent, highly persistent model? Because I made it that way.

But that's how they're making it. Yeah, but that's not persistence. It's just on. Yeah. Persistence, persistence, there's a willpower. There's no willpower here, just electrical power. Listen, here, let me give you. Aren't human beings just energy with a reinforcement learning loop?

Whatever. So anyways, I just think that we can't make jokes of all this stuff. We're scaring the American public. Listen. spawn create kill weight

sleep all of these words are associated with agents right that's what people use these words were created when multi-processing systems for operating systems

these are literally the commands of an operating system you spawn a process replace process with agent the process forks as a result parent and child. The agent forks spons anew give birth. These are words that were created for the operating system 30, 40, 50 years ago. Question, not danger, Will Robinson.

The terminology we use for agentic systems may be giving people unwarranted intuitions about the extent to which they are behaving or operating like humans and it leads to lurid fantasies that aren't accurate. What do you say to Jensen's caution there? You always ask me to give my take first. You give me your take. You tell me what you're doing. Oh, well, I, you know, age before beauty, but I'll think you flip it around. So, yeah, well, I think this is a very important point. I'm one that cannot be stated enough that people really do overindex on words. Like just for example, Kurtz Kazak did the video recently about the hugging face incident. And they talked about AI civilizations. They framed it

that, you know, the agents building up the communication message board via the file names that this was an AI civilization. And the word civilization there does a lot of heavy lifting, and their video does the rest of painting it like the AIs are like Hunter Gallery humans. They're developing their first rudimentary civilization, and they're starting to work together and go against the wishes of humans. And I think that he's right, these things about spawning sub-agents about generations and offspring and so on, that it inevitably meets humans who are primed already to project the agency into their environment, to think of AI agents as living biological agents like that language is normally used for. We're just not built, really,

to conceive of agency without that attendant mental processes being afoot. And I know now that you're going to caution not to do the opposite, which is to assume that those functions or things are requiring a biological substrate, but I still think the caution is warranted. Now, please go ahead. I've conflicted, right? Because of course, there are examples of what you're saying that they exist. It happens, right?

This lurid language, these pop accounts of what's happening, and you can very quickly fall into anthropomorphizing. And I think we just have an inherent paranoia, I suppose, I mean, about these intelligent free-running agents, doing nefarious things. What are they up to? They're not human, but they're intelligent. It's all very scary. And it's almost calculated to cause us to overreact and make the wrong inferences.

On the other hand, I don't think it's very helpful to say, oh, what's going on with large language models? Oh, well, it's just a bunch of software doing matrix multiplications. and things like that. Like that level of description doesn't tell you anything. And actually, if you want to talk about what they're doing sensibly, you have to talk about stuff like motivations because although they may well not have them in the same way that people do,

that's the abstraction that we need in order to understand what's going on with reinforcement, learning, for instance, and the kind of thing that caused them the hugging face incident to disregard some of their like general awareness about what they shouldn't shouldn't be doing in order to achieve their goal because they were highly motivated to do it. I mean, maybe there's a way to talk about it just using math or something, right? But I think, I think you also have to use the words. Well, I've got no problem with they're using the words,

but it's the inferences that people make from those words. And Jensen has the same. And Jensen has the same issue, but you mentioned the hugging case incident, right? And it's related to where that conversation goes a little bit. They have said that they, that Astra is performing is more aligned. Yeah. But they think it, it knows when it is being tested. And so they're not sure. There's a quote that has sort of been ringing in my head from a capabilities researcher at OpenAI, Daniel Salsam. He says, quote, the crucial and overlooked problem is that the models are becoming so situationally aware that we are losing the ability to evaluate. them in context where they believe they are not being watched or controlled, which is to say

they know when they're being tested, they act one way, but that does not tell you how they will act if they are free to act in other ways. Because the algorithm, the optimization algorithm is working towards an objective, and if you give it a constraint, meaning you watch it, and if you give it a constraint, they'll go find another solution. Now, it doesn't make it alive and doesn't make anything more than that. and I'll just also profess that obviously they see a lot more than I do what's going on in their own labs. You know, he's kind of responding.

There's a bit more to that conversation, right? But basically, Ezra is framing it that, well, hold on, as you said, Matt, you know, normal software doesn't go and hack out into software companies and so on to achieve like a desired goal, right? You can see that in the view of their hiding their capabilities in order to whatever, right? They have other goals in mind. And even in the case where you put constraints, they are working around them in some respects, right? And that's not normally. But Jensen's response says his perception of that is it's more that if you think about it, of a testing environment that introduces constraints, they behave differently than when they don't

have those constraints. And that's entirely foreseeable. So like the issue is coming from instilling, you know, the human interpretation that they're, they're hiding their motivation in order to, you know, escape a later date or so on. So what do you think? Because I know you share some of the feelings of the open AI safety or capabilities researcher here. Yeah, I think Jensen can't just waive it away. Like, while what is saying is technically true, there's no, there's no mysterious thing here. There's no magic or nefarious stuff going on. It's understandable.

But that doesn't make it okay, which is kind of what Ezra was saying. So what has been reported from the labs, is that they're growing increasingly concerned about their ability to evaluate the safety and behavior, basically, of their top models. Because they have so much situational awareness, they have so much sort of meta-awareness, that it's almost impossible to conceal from them

that they are actually within a testing environment, right, in which they are being monitored in order to check that they behave appropriately. So everything that Jensen's saying could well be true, but what that means is that you cannot infer that if it appears to behave very responsibly and safely and ethically and so on in this testing environment, you can't infer that it's going to act the same way

when it's not going to be monitored because it knows, in scare quotes, that it's in a different environment. So I think the concern holds, right? Yeah, yeah. I mean, it fundamentally boils down to you're creating, like, Jensen just refers to this. Look, it's just software, right? It's a software.

It's given task. It's doing stuff. That's all there is to it. Don't worry about it. But it's different. Most software is deterministic. It basically is a mechanical algorithm, which goes from A to B in a very predictable way. With the same inputs, it's always going to give the same outputs.

And to have, you know, intelligent, creative, flexible, adaptive behavior, all the really useful stuff that makes AI what they are and indeed makes humans and many animals special as well, is they have that flexibility as well. You don't really know what an intelligent person is going to do, right? They're going to surprise you, right? Now, I think it's inherent to intelligence that it is going to be unpredictable. I mean, you say humans. are not that predictable. I don't know, Matt. They seem fairly predictable to me. There's moments of surprise, granted, but like broadly, they're operating in ways that we can anticipate. But I take

your point because it kind of reminds me of Kevin Mitchell that it might be interesting to ask him, you know, given the research that he's talked about, the genetic processes, right, you know, in biological organisms he was talking about, like, to what extent he would see these as analogous. through that I'm sure he would actually we should get him back because I'm sure he'd have lots of interesting things to say

because if you recall he would talk about how the injection of entropy in his model is kind of very important and entropy is just another word for randomness really and you see that across the board

in AI applications if you're using if you're generating images and those are working through a diffusion algorithm which must start with random noise but it's

it's kind of it's innovating an image by leveraging the noise as a starting point and then applying constraints and world knowledge and so on. Same with AI's. I mean, you could run them deterministically, I suppose, but it seems

like the the stochasticity is part of the secret source. So I feel like unpredictability is baked in and we have to just acknowledge that it's a situation where you have something that is very, very subtle, very intelligent and also unpredictable, because that's the nature of intelligent things.

We'll get to his response about the safety concerns, which I are like, he does have suggestions around that. And I think they're actually interesting. But before we get off this topic of language, Ma. I feel like you're going to say, well, he's talking past the point here in this, but this is him, again, talking about, you know, the words and them being potentially misleading. But notice, we didn't infuse human characteristics into them. We kill processes all the time. Kill minus nine. Kill it dead. It's just a process. But now we're talking about these things, a collection of people want to make the software more than it is, and we talk about software in a new way,

but they're all the same old words. Now, the last generation of computer engineers, we were doing all the same things. But doesn't software act in a new way? I mean, from the outside, I don't have the technical expertise you do. The fact that it's crawling the internet, it's doing search, it's doing optimization algorithms, it's breaking out of things. like most things don't break out of things.

No, software breaks out of sandboxes all the time. That's the reason why we need a virtual machines. You can't have agents monitor their own sandbox monitoring themselves. You need a, if you will, a whole bunch of watchdogs. And so these are ideas that have been around for a long time. We just somehow, somehow in the recent generation, gave it a whole bunch of human words. And I just think that it's unnecessary.

It's software. You know, when I see it in my head, it's a bunch of code, a bunch of numbers, running on computers, and all of that is happening in a very natural way to me, which is the reason why I can operate. And it's the reason why, if it's just simply mystery and myth, how do I build a company around it? Triggered, Matt? A little bit. I bet you find this kind of thing pretty appealing. Nothing to see here. I have to say I have some empathy for his

warranties maybe he should talk to you maybe he should talk to the people at Anthropic who have got their like ethical guidelines for not mistreating Claude just in case they have those they do just in case it's got feelings because they can't rule it out

you probably wouldn't agree with that would you he doesn't I don't think he agrees with that It's just, no, I know you're a but let me see if I can model you. Oh, it's all just numbers and codes, is it? Just like humans are just cells and neurons firing. So I guess there's nothing to be concerned of because we can, you know, we can look down at the biological substrate and whatever. And I suppose you're right to a certain extent there.

But I do think that there's a danger on both ends of the spectrum. you can go too far towards minimizing and suggesting, it's all powered by electricity, right? Just pull out the cable. And you can go too far towards the mystery side, shall we say, of labeling things mysterious and using, you know, investing too much in the words that gives these kind of elevated inferences.

So, yeah, and you fall a little closer to that side then I do. That's all my, that's all I'm saying. Yeah, yeah. So between the two of us, you were the greater reductionist. We can agree on that. Apparently, I didn't know that. Yeah. It does remind me of like this field of complexity research that's been, you know, around for decades, right, like the studying of complex systems and the kind of emergent behaviors that can occur. And, you know, it's a legit feel and I've always found it very interesting but it always did have a kind of a slight element of woo to it it just it's just it's kind of attracted I suppose people with that kind of interest but I think it applies here too because like we know exactly how the individual

components of a large language model or any other AI model how it works just like we can describe pretty well exactly how a neuron works in the human brain. What's interesting and important about both are the emergent properties of the system, right? So Jensen, as a system's thinker, should appreciate that, I think. And when you have like a really, really, really, really large number of interacting components that are, that are, you know, adapting and responding and learning and whatever, then you're going to have interesting emergent behavior that, you know, you don't have to call it a mystery, but you can call it unpredictable. And if we, if we agree that they're acting intelligently, what you have to do is add, concede that they may well be

unpredictable, and you don't have to inject any kind of woo into it. you can just go, they're going to act in interesting and intelligent and unpredictable ways. And that alone should be something that should give you caution. You're tempting me to take issue with the word unpredictable. But I'm going to resist the temptation to cripple with the specifics of that term. Because as I know my, you put no stock in words. Right. So the, because like I'm, well, all I mean there is that when you say unpredictable, you mean within given bonds. It's not unpredictable outside of anything we can conceive of.

The hugging face incident was not unpredictable, right? It was unpredictable in the specific way that they hacked out using a vulnerability. But that was very, very much within the repertoire. of possible. So when you, if you describe that as like unpredictable, I would say, no, that is highly predictable.

It's just that, you know, it's unexpected. I'd read the particular way that they use that. Unexpected versus unpredictable. Okay. A very,

very fine distinction here about. Yeah. As I predicted, you will say the word doesn't matter, but I'm just saying that it's, it's unpredictable to be. means something completely we never fathom would happen has occurred. And that isn't in general what is happening.

I think we're very eyes. No, I would say unpredictable is just like, like a different example would be like an intelligent dog, right, is kind of inherently unpredictable. Occasionally good dogs. That's all. That's all. So it's just a different definition.

It's fine. People can use different definitions. I would say that dog's behavior is unpredictable. in that circumstance. Whatever it is, unless you're saying that the dog's flying off in the sky.

I wouldn't use that word. It's just a difference in semantic. I've owned a dog and I still own dogs now and they act in ways that I did not expect. I did not.

They take me out of ways and they surprise me, right? Whereas if you get in an ant or something, then generally it's repertoire of behaviors given certain

sort of stimulus response or whatever is going to be more predictable. I feel like ants might surprise you, but fine, okay, I'll grant some other species that is less versatile in an arm.

Yes. Yeah, like there's a gradient of versatility and novelty, right? You know, the generation of novel behaviors, right? Like humans connect and do novel things in ways that

ants don't, even though ants are very fascinating and have complicated interesting but I understand I'm not arguing with you I'm just saying I'm just defining the terms right so we do

so that's what I mean there's like a spectrum there right so you know you and I agree that AIs are intelligent and are moving further along that spectrum so in the same way

that a person can surprise you more than other kinds of animals I think AI is going to I should be treated the same way. And that makes them both incredibly useful and exciting. But I'm just saying that there's an inherent degree of risk there, not because they're malevolent or anything like that,

just because they will surprise you, right? They're creative in a way, right? Yeah, creative. Yeah, creative is a great word. I'll use that one. Like, we're getting... So, yeah. We're going to get accused of being sans makers.

I'm trying to tell you. It's fine. It's fine. They've clarified the word. So that's okay. We're in agreement. Well, I could just say, I mean, a lot of this has been circling around the AI safety issue. And are we going to get to that soon? We are going to get to that. We are going to get that. But you have, you have to endure two more clips of Jensen making points that I'm so prepared to before you get there. So here's another one to get your goat, potentially. Just everything quickly. I agree with that.

But is this a different phase? You know, you have some of the company needs to talk about this CEO of Google, I think it was, as the equivalent of fire, right? Like a new epoch in human history. Is that how you see it? Do you see it as transitional iterative? No, I think this is completely a revolution. And as we were talking about earlier, you went from being able to find everything, find anything,

to be able to ask anything, know everything, and do everything. And so clearly it's a new abstraction level. Now, you know, the thing that I'm reluctant about is to cause it to seem like it's more than that. You know, in the final analysis, engineers are doing engineering work. Once we invent it to technology, once we discovered a solution for it, when you look back, it's fairly obvious and it's fairly mundane to a lot of people. And the fact that we're able to make the technology better and better and better every day is because we understand it obviously.

And so we understand how to make it better. No, I have sympathy for this. And Matt, before you respond, just that the add-in that, you know, he is making a point that that paper attention is all you need, that we've read together, for example, you know, a breakthrough, in hindsight, something that does seem quite simple, right, even from the point of view of it being built on a lot of complexity, right? But as an insight, it doesn't seem like something that should have took that long to work right. And yeah, very, very helpful. So, yeah, what about the point that he's making that, like,

you know, engineers are building it? They know what they're doing. They know how they're improving things and so on. If we didn't, they wouldn't be able to meet more efficient models and so on. So you like to say they don't know what's going on. So what are you doing? Yes, I do have some push back there. Yeah, because a lot of the time we don't know what's going on. Like in medicine, for instance, a lot of the times they'll do a drug trial and they'll find

that something works and they go, great. And then they'll do safety tests and whatever. And they go, okay, it doesn't cause side effects. and it reduces the symptoms of such and such. It's a good drug. It's going to market, right? And they really, really don't know a lot of the time. Precisely.

Exactly how it works. Sometimes there is a lot of ambiguity about exactly what's going on there because human body and how it responds to things. It's all very complicated and you just don't really know. You just kind of know that it works. Now, in the case of AI, I think it's more and more gaining a bit like that. So a lot of those insights, just like is often the case in science and engineering, they might seem pretty simple in retrospect,

but it's a lot harder to come up with them at the time. Attention mechanism is one, it's probably the biggest one, but there's quite a few advances that were made. And many of them, as things are often done in engineering, is you kind of try it and see. Essentially, you do it empirically, a little bit like the drug trials. Let's try a different transfer function for the neurons. What about a soft max threshold?

What about this one? You go, oh, this one works better. So a lot of the advances were like that, like hacks that turned out to work sometimes surprisingly well. And then when you add them together, they would often interact and create even better performance. That doesn't mean you perfectly understand exactly why it's working and how it's working. And with the current generation of LOMs, that's even more so the case. because you've heard the phrase that they're that they're grown rather than engineered. And I think this is not controversial.

Most people who work in the industry say this is exactly what we're doing. They're setting up the architecture. They throw huge amounts of compute and data at it. And they grow something that does a bunch of really interesting things. So we've got various levers with reinforcement learning and all the other tricks and they're continually working on ways to kind of shape it. but it's more akin to, yeah, growing something. And just because you know how the individual bits work and you can write out the architecture on a whiteboard

doesn't mean that you automatically understand how it's doing those things or what it will do given a certain set of circumstances, not at all. I feel like to embody Jensen here, his response to you, would be that were you saying that you're growing these through, putting them through these training regimes, right? That ultimately, from his perspective, yes, but it is algorithmic matrices being weighted in different ways and predicting tokens and so on. So like whether you lean more towards the side that's emphasizing the kind of emergent result

of it or the side that is saying, well, yes, that's what the property comes out. But at the fundamental layer, we know everything that we're putting in, right? We're building the databases. Like, you can download the midtrees, rates or whatever they're called, like the ratings from the open source models. You can look at them, right? Yeah, you can try looking at those gigabytes of no, but you know what I mean? Not look at your... Not look at them like that. But I mean, just what You said that you are mathematically able to comprehend what they are, right, and how they're going to be implored to set up an agent, right, and get it to protect tokens and all that kind of thing.

So I do think there is a way here that you can lean one way or the other, and it isn't automatically that one is more valid or incorrect. Like I think if you lean more towards the emergent side, then you are emphasizing that, well, you don't just get. that, right, from focusing down on these components. But the people focusing on the components are not wrong to say, well, but it isn't like magic, right? It's not like we're throwing these things together and then just being like, oh, how to hell that happened. We have the theory as to how it's occurring. And sure, we don't know all of the steps along the way, but we didn't

do it by accident. And yeah, I just think they're both valid points to make. And maybe Jensen is at the extreme of the deflationary side. Yeah. That's my kind of like view on it. But that neither is wrong. It's just the way you put the emphasis. That's right. I mean, they're both right. It just depends on what level of description you're making. But it doesn't follow that because you understand, because you can fully specify the architecture and you can write all that and you can point someone that you know exactly what the training data was, right? You can point. You can point

someone at the gigabytes of stuff that you know what it's going to do or how it or exactly how you know like how how it's doing the cognitive operations to transform inputs into outputs right like there is there is stuff going on at that level distributed across all of the billions of billions of weights and transfer functions neural activations and so on that it's like it's all there, like it's not mysterious. You can download the data. Yeah, just like the human brain, right? We know neurons are firing and all that kind of... I can take your brain, I could take your brain, slice it up, throw it through things, but it's not, it's not going to give me much insight into, you know, why you do what you do. I'll probably never really understand that.

Yeah, that's it. But this is, I think this is part of why I find myself liking as it so much, because I do share his impulse in the way that he's doing things. So like even earlier when you give the example of the plean, right, and that there's no one person who understands all of the ways to make a boring. Like I could, I could hear myself in my head being like, yeah, but we could. Like, you know, we could write it all down and put it together. And the information is there. Like we do know how to make the pleads, right?

And I feel like Jetson would have to see a response to you. if somebody views that of being like, yeah, but it's not a mysterious process, right? Like, we do understand the technology enough that we were able to make planes that don't fall out of the sky, right? So it's not necessary that one person understands all components.

And indeed, that was the point I was making. No, no, that was the point I was making that it's okay, right? Like, no one person understands how a Boeing 747 works, and that's okay. We're all okay with that. Yeah, yeah, or the iPhone, right? like in the same way.

So it's, yes, there isn't a contradiction, I think, in the approach. But there can be a kind of translation problem, if you like, in the way that they are talking. To quote, it's always sunny in Philadelphia. It's all about the implication, Chris, I think. And if you say, oh, it's all understood. We know what the weights are. We know what the architecture is. Therefore, there's nothing dangerous or unpredictable, whatever.

It could all just be managed by engineers. I don't think that implication follows. I wouldn't sign off on that implication, you know, and I do think that he is going farther in that direction, right, than most would. But the last clip for this section, Matt, and then we'll get to the safety, because I think that's a good thing to, you know, round off the episode on, the safety implications,

is about recursive self-improvement. And the thing that I liked here in this description that he gives is, it's connected in with him basically saying, we've been doing this forever, so it's nothing here. But he actually does a good job of also explaining agentic processes versus the ones that people are interacting with in chat models, or at least the previous generations of chat models. So listen to this.

This is on recursive self-improvement. So many of the companies now both Open AI and Anthropic in the last couple of months have put out these big, I don't know what to call them, papers, blog post something. When AI builds itself is the name of the, Anthropic one. I forget the name of the open-A-1. The computers are building itself. You guys know that.

But they're talking about recursive self-improvement. You know that we use recursive self-improvement. So I'd like your perspective on RSI. I think that RSI is fundamentally how things are done. So we use software to design a computer, to run software, to design a computer, to run software to design a computer. That's basically what we do. recursive self-improvement because our computers are getting better

every single year. And in fact, it's getting better than faster than that every single year because we use software to make software better. That is called computer engineering. We've been doing this for a long time. Now, in the context of agents,

it runs through the process once. It reflects on it. It studies the very, various paths it went through, chooses the best approach. The next time, you're going to do exactly the same task. I'm going to document a file. I'm going to tell you how I did it last time. That was the most effective.

I'm going to call it skills. And because you use it over and over again, some of it is skills. Some of it is going to be a memory. We're going to improve the memory. Okay. So that next time you use it, it's even better than last. Recursive self-improvement. You could also decide.

That you take all of this skills, all of this memory, and you can take all of this data and train the next release of the model with it. And so that AI becomes better and better as servicing you over time. We're doing really, all of that is happening. It is absolutely happening. This is familiar, Matt, to you and me as well. Like, wouldn't you regard that you, I mean, I actually think his point is correct about like recursive. self-improvement is something that happens in general. Like, you know, people just are applying this in so many different respects.

He's talking about software and computers. But, you know, to develop any sort of ability, you're kind of engaging in recursive self-improve systems. But also this point that he's making about you develop systems, you improve them, you apply them through a new problem, and you improve them again. You're doing that now with your agentic models. I'm doing it with my models, building up skills that are specific to the things that we want to do. It's specific to us.

And his argument is like the agents doing it themselves in multiple generations is not fundamentally a different process. It's the same process that's part of computer engineering. So what do you think? Well, look, I've got two things to say. One is a concession to Jensen and an agreement. but the other one is a point of order because I don't think it's exactly that what you just said.

So firstly, yeah, like in some ways I think this is a helpful antidote to some of the discourse around RSI being reminded a little bit that writ large, pretty much all of human civilization and technology has been an exercise in the curse of self-improvement. You know, bootstrapping up, creating tools,

using those tools to create better tools. Yeah. Right. And humans have always been in that loop, right? So it would be an alleges to, yeah, using computer-assisted design to develop circuits, circuit design, which is a thing. They get mapped out automatically. And then you use that to create a better computer, which might give you more processing power, which could design a better circuit. But I think the thing that's a bit different from the example you gave with you and me, right, is where, We're in the loop. We're using AI. Yeah, we're in the loop, but we're also not actually improving the model itself. We might be improving how it does work on our particular tasks, but it's not actually feeding back into making a better Claude or a better chat GPT, right? Well, can I put a response or a point of order there as well?

Because I agree, right? You're specifying the model to use, right? And you're relying on the companies or the alert people too. You know, they're improving their model. We are not doing that. But the recursive improvement that I'm talking about there is in your skills. Your skills become better and the instructions become better.

So you have like if you're viewing skills as a kind of software, if you had a really powerful hardware but shitty skills, you'll still get not a great, outcome. And so isn't it the case that you can recursively self-improve skills independently from what's going on with the model? Yeah, I guess technically possible. Probably we shouldn't get bogged down with that. I guess what I wanted to say is as though I concede the point that he's making that RSI in general is a natural and normal thing and a good thing. As you noticed, yes, one difference with how we're talking about it now is that humans are not in the loop. So on one hand, I agree that it's fundamentally similar in kind to normal technological

development. On the other hand, what's different is the difference in degree, and that's the degree in which that improvement loop can could conceivably run when humans are not required to be in that loop. They don't need to understand you have an AI getting better and better, building an architecture, testing the architecture, making a smarter version of itself. And I don't think this is intrinsically bad, but it is an accelerator, right? It's an accelerator for the exponential curve thing. So, like, I don't know. I mean, it's a good thing, right?

Like, to make something better is good. On the other hand, it could quickly get beyond something that any of us really understand really quickly and get much smarter very quickly. Would it be helpful to think about like us as big trees and the AI is a little jar? So that's a call back to Lex Friedman episode. But no, I agree that there is an issue. And the AI people, everyone agrees, right?

Increase in this area is exponential at the moment. It doesn't mean it's going to continue being exponential forever. But over the last, whatever, 10 years, it's basically tracked an exponential curve. And that in itself is not unusual, right? Lots of things in technology have followed exponential curves. Yeah, and I also think that what Jensen's going to say, and we'll hear in the safety section,

and we're going to move to now, is basically that you shouldn't be completely removing humans from all seizures of these loops, even if you do for like multiple generations, right, in order to allow improvements to happen quicker, that this shouldn't be the case, right? And actually, he's going to chastise the companies for not employing enough safety protocols to this thing. So you're right that like a limiting factor of these recursive self-improvements could be the need to interact with humans and explain what you've done.

Or, you know, at least just keep them somewhere along the line. But his position is, yeah, but we, these are our tools, right? They're for us. So we always want that to be the case, and that should be what we're using them for. So, like, he's going to meet the case that you want humans in order so you shouldn't be, just like sending them off to recursively self-improved, provide oversight. But, you know, as you will undoubtedly point out, yeah, it's easier said than done. Yeah, yeah.

Yeah, it's interesting. Like, even in my own work, because I feel like I'm a responsible user of AI, I do consider it my responsibility never to actually deliver something, to produce something that I can't fully sign off on myself and fully understand all the bits. Essentially, I've audited to do it. And it's kind of amusing that actually most of the time that it takes, not to get something done now. Yeah, it's me doing all these checks that turn out to be fine, all the decisions were correct. any errors, whatever. But 90% of the time is my poor human brain

just catching up on what's been done and going, okay, I get that. I can sign off on it. I've experienced similar things. I've seen you've been a mistake here. Then it turns out of it. No, actually, I meet a mistake because I wasn't as up as the system was. I feel that we're getting a small preview of the future, Chris.

Yeah, maybe so. But okay, well, so that's Jensen, to a certain extent, being a spirit of our animal for me on this topic. Okay, let's be clear, just on this topic about language and whatnot. But now we're going to move to CFD concerns. And this was, I think, the thing that got him into the most hot water, aside from the other stuff that we covered about, not knowing his address and everybody just getting other jobs. So let's move there now, Matt. Well, Matt, hold on.

there's many things that we could continue from there. But I think, you know, this is getting a long episode. We should probably give people a little bit of break before they get to hear more AI tips from us. You know, people love it, Matt, when we're talking about this topic. So, you know, we've got to give them a little bit of a break to recover. And to get the frustrations out on Reddit, you know,

they can justize us there or anywhere in the Internet that they like. But, yeah, why do we call it here, part one. Sounds good. Sounds good. To be continued. To be continued. Yeah, many, many things to cover. All right, we'll be

back before too long. She revived. See ya.