But you've been described as Silicon Valley's favorite podcaster and you write a lot about AI and tech and that's actually the conversation we really want to have with you. Partly because a lot of people watching and listening to this, they've got lives, you know, family, work, etc. And they haven't been to California, they haven't been to San Francisco, they haven't seen that you know, a third of the cars on the road or 25% are robots basically. Like this is happening fast and a lot of people haven't caught up.
And what we'd love to do is just kind of connect people like you who really understand what's going on with a much more general audience. That includes us, frankly. So first of all, can you explain in broad brush strokes what is happening with AI? I know it's a massive question, obviously, but do your best.
Well, I can explain it very concisely, the models are getting better and then we can be a little less concise. I think you're correct to point out that there's this huge discrepancy between what people are seeing in Silicon Valley and what people are observing outside. And it's frankly because of how useful the models are becoming at certain kinds of things. So by models, I mean you've seen ChatGPT, you might have heard of things like Gemini from Google, you might have heard of Claude from Anthropic. And you might be using these models to basically do the equivalent of Google search. Using it to replace something that you could do a Google search instead of going to type it into ChatGPT, see what it says. What people are now using these models to do in a very powerful way is if you're a developer, some of the top developers in the world, some of the top researchers in the world, they're not writing code. They haven't touched a line of code since December. They're not looking at a text editor where you would see lines of code. They're talking to the AI. They tell the AI, 'Hey, I want a feature that does what X. Can you build me a new repository or a new code base where I make a certain kind of application, a new website, and even can you go and do research for me?' So, in process of building AI, you need to do this research of like, how do you build better algorithms? AI is getting to the point where you can just describe at a high level what you want to happen, and it'll go do that software engineering for you. And so, to your point, the people in Silicon Valley, they're getting tremendous productivity out of these. These are people who are getting paid, you know, who are becoming 3x, 4x, 5x more productive as a result of using these models. So far, we haven't because these models have been really good at text in, text out work, that is software engineering, software engineering just a file of text, really, and you can just read every single text file, you can add more to it. The AI has been amazing at that. It's been bad so far at Well, it's terrible at physical work, right? So, if you're doing any kind of blue-collar work, robots are just not there yet. But then, even if you circumscribe it, if even if you go in and look at we're not just going to look at software engineering, we're going to look at all kinds of knowledge work, right? All kinds of work that you can do on a computer. Maybe 40% of the labor force is doing work that you could just, you know, do remote work. If COVID happens again, you can put them on Zoom and they could do their work. And AI companies now want to be able to do all of that work. That requires training AIs to just be able to do anything that you can do on a computer, and AI should be able to do, and companies are saying, 'Oh, we think we can get there in a year, maybe 2 years.' But that, I think, is explaining this discrepancy in what people are seeing.
It does. And one of the things I think is also, a lot of people would have tried the model at one point to like do a Google search or, you know, sometimes people have written articles using it, and it turns out that it does make at the time they tried it quite a lot of mistakes, and people go, 'Ah, this is all, you know, BS, it's not going to work.' But the one thing I think people don't appreciate is how rapidly it's getting better. Can you talk a little bit about that?
Yeah, so I'll maybe tell you a story about my own use of AI. I among the Silicon Valley people, I've been a sort of skeptic. I've been the person who said you think the singularity is going to happen yesterday, I think it'll take 5 years, 10 years. By this I mean this idea that you'll have incredibly powerful AI systems that are able to do basically anything any human being can do. And some people in Silicon Valley are just we'll get it later this year, we'll get it in 2 years. And I say, 'No, it'll take longer.' And even for me, I've had to admit that the progress has been pretty fast. And here's a story. So, last year, you know, like you guys do, I research for my podcast. I basically read things, but then I also talk to these LLMs, you know, the equivalents of ChatGPT to research for podcast. And say last year I spent on the order of $100 on if you add up my subscription to ChatGPT and Claude and whatever and Gemini. There was a week where I was prepping for two different guests where in that prep, I just threw in a bunch of papers that are relevant to the research and a bunch of books and whatever into a folder and I said I just said to an LLM, 'Hey, help me understand all this research so I can ask the person good questions about it.' And I turned on various things which make the model much more expensive. It's faster, it's smarter, use a bigger model. And during that week, if you take my spending during that week and you turn it into a yearly spending, how much it would be if I just did that week over the course of a year, I would be spending over six figures on AI spend. And I think it's possible next year I'm like spend it would make sense for me to spend seven figures on this AI research for my podcast. Okay, this is just to say that it's getting to the point where I could hire an analyst, I could hire many different analysts to help me prepare for the podcast or researchers. And I'm actually realizing, no, it's so much it's more useful to hire AIs because there's things AIs can do that humans can't do. They can read 50 different papers that are in my folder that are relevant to an upcoming interview where all these books they can read it in a second. They have all this compiled knowledge. They know about everything, right? So they don't need to get up to speed. They're incredibly easy to onboard. So you can just like keep spinning up more and more AIs as you keep coming up with more uses for them. Oh yeah, for me that's been the thing that's given me a bit of psychosis. I just see how useful they've been for interviews so far. But yeah, if you lots of people are noticing this kind of usefulness now.
I think the thing is is when people look at a lot of the AIs, and this is a question a point that Konstantin made. It's the lack of accuracy and the fact that you constantly have to fact-check. I'll give an example. I was looking at guests that we could have from America on the show, and I said to Grok, 'Who can we have on that hasn't appeared yet and is based in America?' Number one answer was Douglas Murray. Douglas Murray, I think he's nearly beyond 10 times. And practically every suggestion it made is a guest that we've had on the show.
So whilst I accept that what you're saying, it's also riddled with mistakes and errors.
And to the point where I don't trust it at this point as a technology. I would never take something it says as read. I would always feel that I have to check it. In the way that I wouldn't do with a highly competent human employee.
You know, you've seen that I'm sure you've seen that meme where a guy wakes up in the hospital with a scar on his stomach, and he says, 'Wait, wait, my appendix on this side.' And the robot goes, 'Okay, thank you for that feedback. Let me try again.' There's a couple of things there just the way that people sometimes have these weird failure modes. You know, you talk to a certain kind of person and they say they just their mind just goes in a certain direction. AIs, the way they're trained, makes them do that. One way they're A big thing they're trained to do is associate things. So, they're just like really they're just seeing globs of text. And they see this thing is close to this thing, there's an important relationship there. So, Grok is seeing that your Trigonometry podcast and Douglas Murray, they just seem to go together a bunch. And that bias in his mind is overriding its ability to sort of think critically about, 'Okay, well, who have they had on? Who have they not had on?' I would be curious, by the way. You should run that experiment with all the other models. I would not be surprised if the other models, if you ask the question, 'I want somebody who I've not interviewed before,' that they would immediately catch that. And by this point, the models are good enough to be actually giving you novel names. I don't think they'll give you names which you're like, 'Wow, that's an amazing find. I can't believe I didn't think of it or I would never have thought of it otherwise.' Because they're not amazing at discovering in this I think humans are great at this idiosyncratic thing of like, there's this like weird angle that I'm like really obsessed with that other people aren't thinking about. Where the models are not going to do that, they're sort of an average. But, when you are preparing to interview the person, I also think they're not amazing at coming up with questions. This is a thing I have to do that you have to do. They're great at just like I would have to hire, you know, the ancient Greeks had would hire one-on-one tutors for their pupils. And this is how you got Alexander being tutored by Aristotle. And there's something powerful about this way of learning. One-on-one where you can just directly ask questions and learn. And so, they're amazing at just teaching you stuff because you get this one-on-one tutor who notices your confusion as soon as you have it. You can really probe at your understanding in different ways. But, I take your point that like some of the models are not there yet on a lot of these kinds of biases that they have.
