Note: The following is a rough transcript that has not been hand-revised by High Signal or the guest. Please check with us before using any quotations from this transcript.

===

noah: [00:00:00] What we're seeing in a lot of domains is that agents are better at generation than at verification. In the arms race between generation verification, generation wins. [00:00:10] It's generative AI, not, not verifying AI, right? 

hugo: That was Noah Smith, economist and author of No Opinion, on why generative AI is currently far better at [00:00:20] producing things than it is at verifying whether those things are accurate, useful, and/or trustworthy.

Noah joins Duncan and me today to unpack why AI is [00:00:30] already changing how we find and trust information even before it clearly shows up in GDP. We talk about agentic coding, why feeling faster is not the same as increasing [00:00:40] revenue per hour worked, and why software may become more abundant, more useful, and less profitable all at once.

We also get into the [00:00:50] slopocalypse, a world of AI-generated resumes, vendors, applications, code, content, and even fake companies, where the scarce skill is no longer just production, but [00:01:00] verification and judgment. If generation becomes cheap, the hard part becomes knowing what is correct, what is valuable, who to trust, and what deserves human [00:01:10] attention.

Noah makes the case that this could reshape firms, markets, careers, and the way leaders think about building reliable systems. And we discuss the darker risks too, [00:01:20] from biosecurity to what he calls a vibe-coded super virus. If you enjoy these conversations, please leave us a review. Give us five stars, subscribe to the newsletter, and share it with your [00:01:30] friends.

Links are in the show notes. I'm Hugo Bowne-Anderson, and welcome to High Signal, brought to you by Delfina, the AI agent for your data team that allows you to move [00:01:40] fast and trust the answers Hey there, Noah, and welcome to the show. 

noah: Glad to be here. Thanks for having me on. 

hugo: So great to have you here, and one thing I've really appreciated [00:01:50] about your work over the past several years is how much you've really got me thinking about the impact of AI, which w- which I work in m- more generally [00:02:00] from an e- economic perspective, but societal perspective.

And I'm wondering... We wanna jump into this today, but I'm wondering if you could start by just telling us your take on how AI has changed the world since our collective [00:02:10] ChatGPT moment in 2022. 

noah: How has it changed the world? Well, it's certainly changed the economy with the big data center build-out. It's causing lots of people to reevaluate how they produce things, [00:02:20] uh, you know, software or various other things.

It's causing a lot of, um, oh, I don't know, for lack of a better word, slop. I think over half the traffic on the internet is now [00:02:30] just AI bots, and that, that will only go up. Humans, human-generated content will be a small amount of the internet. And, uh, the question is, who's, who's consuming this content and why, and what is our [00:02:40] society producing all this content for?

Uh, that's a question we have yet to answer, I think. It certainly introduced an interesting and unpredictable element into geopolitics and kind of thinking about [00:02:50] China and the US and other countries, the military balance with, you know, autonomous weapons and cyberattacks. Just a whole lot of, uh, just a whole lot of [00:03:00] uncertainty about the world.

But in terms of, of, I don't know, economic growth, the impact has been relatively modest in terms of- Impact on how we live our daily lives, I think the impact has been [00:03:10] so far modest. I know a lot of people... B- the, the biggest impact on daily life is probably if you're in one of the industries where people use AI a lot, like coding, you use AI now as a tool.

The [00:03:20] tools have changed, but there's about the same number of people working on the same basic kind of things that they were working on before, so far. And, uh, I think consumer [00:03:30] chatbots are the main thing that's changed. People talk to AI, you know, they, they ask AI questions. It's replacing a lot of the internet, replacing YouTube for how-to videos, replacing internet [00:03:40] lookups for things.

It's replacing search or Reddit or, I don't know, Stack Overflow or Math Overflow, whatever. And so that's the way it's changed the world so far. But in terms [00:03:50] of the ways that it will change the world in the years to come, I think there'll probably be a lot more than we've seen. Now there's just barely the tip of the iceberg.

hugo: There's so much in there, and I actually really love that you [00:04:00] mentioned agentic coding and Stack Overflow, 'cause I, I do think Stack Overflow is the, potentially the canary in the coal mine with respect to what, what, what happens to the internet and social media and [00:04:10] communities on- online more generally.

But I am interested in your thoughts on whether agentic coding has... How much value it has created so far. And I [00:04:20] mean, you've written a bunch, and we'll link to y- what you've written in, in, in the show notes about... It's certainly created value for An- Anthropic. We can see how many enterprises are paying significant amount [00:04:30] of money to, to Anthropic.

But has it actually... Is-- Has it created actual measurable, discernible productivity improvement for software [00:04:40] engineers? As a builder, I think it has for me personally, but measuring it seems quite elusive. 

noah: Yeah, that's right. Um, you know, productivity is properly measured in revenue per hour [00:04:50] worked, and it's not clear how much, you know, revenue's increasing, how much of that is due to, due to adoption of agentic coding.

I mean, agentic coding's not very [00:05:00] old. Yeah. It only really started to work in December, maybe January. Like it's, it hasn't been that long. It's taken the world by storm, but I think it takes longer to, [00:05:10] a lot longer to incorporate into work, to actually produce stuff with it, to sell that stuff. I do wonder what everybody is coding.

Ev- you know, everybody I know is just using Claude [00:05:20] Code constantly. Right? But I do know some people who have used it at their startups, which are not AI, AI startups. Their product is not [00:05:30] AI, but they've used Claude Code and sort of... I, I have one friend who runs a startup who was gonna hire a couple more software engineers and then canceled that and is just having existing software [00:05:40] engineers use Claude Code.

So that's, you know, but it'll take a long time for that to show up in revenue because they- they're not an AI company, so they have to, um, ship products and it's, it's hard to [00:05:50] measure, like, software's contribution to those products. It's very difficult. It's very murky peering into that, that, tho- those numbers.

And so I think [00:06:00] it'll be a while. For companies that actually sell software, it should be easier to tell. Are you shipping more code? You know, are you shipping updates to your products? Are you making more money selling code? [00:06:10] And so one, one question is, if we have a massive margin compression in software businesses because any software can just be cloned by AI, [00:06:20] right?

Then why are all these software businesses paying all this money to Anthropic or, or OpenAI Are they [00:06:30] just paying this money to, to, you know, run in place? Like if, if there's compression in the margins at the product market level, in other words, if software's getting commoditized, we could actually see people paying [00:06:40] less to Anthropic and OpenAI or temporarily paying more, you know, in attempt to keep their market position than just failing and cutting off their token usage.

We could see that. It could be that [00:06:50] simply the threat of entry reduces the value of software because it reduces the monopoly rents to, you know, sort of market power. And then we could see that means less [00:07:00] demand for AI. So the existence of AI could result in, in reduced demand for AI coding agents, which would then result in lower revenue after a while for Anthropic and [00:07:10] OpenAI.

One interesting thing I saw is that the percent of businesses that report using AI has just-- has gone down according to a [00:07:20] number of surveys over the last few months. Now, that data is almost certainly from before coding agents started to have a big impact, from before agentic AI really [00:07:30] got diffused throughout the marketplace.

And so but it's interesting that we s- we, we may have hit saturation on business usage of chatbots [00:07:40] already, and the amount of revenue from business use of chatbots is small. It's a pretty commoditized space what with Chinese chatbots being so good. 

duncan: Can you tease out, Noah, like the industrial [00:07:50] organization kind of evolution of the industry there?

As there's more competition in core software and the labs are constantly building better and better models, [00:08:00] but potentially those models are getting distilled by each other or by open source. Like how, how does this unfold for you in the next few years? 

noah: Th- th- this-- There's actually a lot of questions here.[00:08:10] 

You're talking about industrial organization of the AI industry itself. 

duncan: Yeah. 

noah: Basically, if you are making-- if you're trying to have the moat be everyone knows your name [00:08:20] and considers you to have the best brand and is used to using your chatbot. That was kind of OpenAI's first idea for like how we're gonna make money.

Everybody knows ChatGPT. Who [00:08:30] knows what Claude is? Who knows what Gemini is, right? And then we have the brand. Everyone's used to using our chatbot. A few will switch, but most will just stick with ChatGPT. The-- That [00:08:40] may be right, but the existence of these other chatbots out there that are just as good or in some cases better may limit the amount that OpenAI can [00:08:50] charge for any of that.

So if you look at how much OpenAI charges for stuff, it charges a lot for Pro, but not many people use Pro. The fees for just the, you know, premium or whatever mode, [00:09:00] I think there's their, their mezzanine mode, uh, which is what most paying customers buy, right? It's cheap as hell. And it's not getting more expensive.

They're not raising prices. [00:09:10] Even as they get better and better and better and better models, the price isn't going up. That's deflation. That's shrink deflation. Or what's the opposite of shrinkflation? I guess growflation. [00:09:20] I don't know, I just made that up. But then, but if you have massively increasing quality for-- Now, we don't know if that's quality in the hedonic sense.

We don't know if people like... Maybe everybody just liked [00:09:30] GPT-4o the most and the new stuff most people don't need or want. So maybe it's not super quality in terms of economic... It's quality in terms of capabilities, but not-- Most people aren't gonna sit around trying to solve Erdos [00:09:40] problems with a chatbot, right?

Most people don't need a chatbot to be more accurate in teaching them about Persian history or something. Like um, most people just want a, a [00:09:50] chatbot that's able to like talk about their problems and like recommend some products or, you know, whatever. And so maybe 4o was really that good and everybody-- all the people who like loved 4o.

You know, you've heard [00:10:00] of this, right? The 4o mania. I don't know why it was 4o, but like I didn't even like that model. But you know, I liked o3. That was my, that was my 4o. I liked o3. Likewise. But anyway... [00:10:10] Oh yeah, you liked o3? 

duncan: Oh, yeah. The heavy kind of long thinking mode was amazing. Yeah, and o3-Pro, I think was- No, my 

noah: boy, o3.

They killed him. They massacred my boy. 

duncan: Look what they've [00:10:20] done to my boy. 

noah: Anyway, but yeah, so, so they're not able to charge much for this stuff Nobody's making a lot of money on chatbot, and it's because there's so many other chatbots. There's [00:10:30] switching costs. It turns out that you can switch from ChatGPT or, as the Japanese call it, Chappy.

That's a cute name. You can switch from Chappy to Claude. It's not-- Or DeepSeek, [00:10:40] whatever. Like, the switching costs are not that high. Ultimately, it's a text box. You type it in, it gives you the answer. The form factor is all the same. We went through a long period in America where we assumed that anyone [00:10:50] in a market dominates the market, where we assumed that monopoly power is something that falls like manna from heaven.

A lot of VCs lost a lot of money on that idea, right? "Oh, Lyft will make [00:11:00] a bunch of money." No. Like, a lot of markets just assumed that they'd somehow monopolize things. Is WeWork the future of real estate? Are we gonna monopolize real estate? Well, no. There [00:11:10] was a lot less monopoly power in the world than people thought, but because you had Facebook, Google, Amazon, whatever, people thought, "Wow, every market's winner take all."

I had some VCs say that to me in, [00:11:20] like, mid-2010s. "Every market's winner take all now as software eats the world." We've over-indexed on that. We forgot about the existence of, like, farming and solar and [00:11:30] airlines and steel and industries where people make a lot of stuff, but nobody makes a lot of money and the consumer gets the value.

duncan: And so if the labs end up in that world where they can't charge a lot for their stuff, [00:11:40] but they've enabled this now hyper-competitive software space to exist, does tech just shrink as a portion of the world? Like, where does that go? 

noah: Oh, well, [00:11:50] it may be that, that, um, software profits shrink even as software booms.

It could be that software shrinks as a propor- I, I think whether something shrinks as a proportion of the economy has more [00:12:00] to do with consumer demand, number one, and what else is growing. Like, so elasticity of demand. So for example, agricultural shrank as a percent of the economy even though food's really important.

I eat food every [00:12:10] day almost. You know, and like, I, I use it a lot, food. It's a, it's an important platform. Anyway, so... But the thing is that, uh, number one, [00:12:20] people filled up. Like, not that many people wanna eat that much food per day. You get fat. And number two, all this other stuff existed that we could use instead of food.

So if you look, if you go back to like the [00:12:30] 1800s and looked at what they thought consumers will have in the future, it's like giant tomatoes. But like people don't want... Yes, we have large tomatoes, but like people don't want that many tomatoes. Uh, [00:12:40] instead people just found out there was other stuff they wanted, like fashion brands and nice sofas and big screen TVs.

And then people found, you know, manufacturing shrank and people, uh, [00:12:50] found out there was other stuff they wanted, like someone to shampoo your dog or like gold-plated health insurance that'll make you never have to think about health or like [00:13:00] the better education than my friend's kids have. You know, like there, there was a bunch of stuff people wanted that wasn't manufacturing services, so the service economy grew.[00:13:10] 

Now the question is like, what do people still want? What is it? What do we want more of that we don't have? Do we want robots to fold our laundry and vacuum our [00:13:20] houses or whatever? Do we want AIs to like go through all our emails and like open claw our, you know, our life for us and like a little personal assistant that manages our life so we never have to [00:13:30] like do our little calendar events and think about stuff?

Do we want a million little apps to do whatever you want? What, what do we want in life? What, and what can we get more of? [00:13:40] And I think the people, a lot of people are predicting modest growth from AI. If you-- There was a, a very, very good survey by the Forecasting Research Institute recently, um, [00:13:50] which includes a bunch of economists and, and other people from other fields.

And they went and asked the, the general public, but, uh, they asked three main groups of people also, of experts. They [00:14:00] asked, uh, AI researchers, super forecasters, and economists. The super forecasters are people who just sit there all day like doing forecasting, and they're really good. AI still can't beat them, actually.

Although it will [00:14:10] eventually, but, but they're really good at their job. So that said, they often have pretty low signal 'cause there's just isn't that much signal out there probably. But so, so they asked all these people like, [00:14:20] "How good will AI capabilities get?" And they all said, "It's gonna increase pretty fast."

I think the scenario they picked was called the moderate scenario, but actually it's, uh, basically a continuation of, of [00:14:30] exponential progress from where we are now. Very fast rate of increase. And then they said, "How much will GDP growth be?" And everybody's picking like, well, 2% to like- [00:14:40] Three and a half percent, right?

That's like solid growth, but it's also the long arc of American growth will just continue basically. And so, and [00:14:50] then in the fast takeoff scenario, some types of experts, I think just the AI researchers predicted like 5% growth. Wow. But that's still like [00:15:00] slower than what India's doing right now. Slower than a developing country.

It's slower than catch-up growth, right? It's basically it-- That would be the largest acceleration we've seen since [00:15:10] the post-war boom. But it would be like fundamentally... And, and it would be transformative. The industrial revolutions were very transformative for our lives. [00:15:20] It would be transformative, but it o-over, over someone's lifetime, but it, it's not, you know, we're not talking like we turn the whole asteroid belt into computronium, kind of like Charles Stross novels here, [00:15:30] right?

Like it's, it's not that kind of thing. And then, um... And this is growth by 2050, so we're talking 25 years out. Beyond that, who even knows? Whatever. And so, [00:15:40] yeah, basically everybody's predicting fairly modest growth, and I'm thinking why are they predicting modest growth? Now, the usual explanations you get are like bottlenecks, Baumol [00:15:50] disease, uh, which software people call Moravec's paradox or, or...

No, I'm sorry, not Moravec's paradox. That's the one about robotic capabilities. It's the one... What's the principle where like [00:16:00] you get hung up by like the bottlenecks in software, like the hard stuff, like the bottlenecks become more important? What's that law in programming? Let me, let me ask AI. Amdahl's law.

That's it. [00:16:10] Amdahl's law. 

hugo: Yeah. 

noah: Yeah. So yeah, you know, that's, that's there and everybody gives that and they said, um... I actually think these people are making excuses for something [00:16:20] deeper that they intuit is going on and don't, aren't willing to admit or realize. Which is I think most of these people deep down think that human beings are getting satisfied.

hugo: It's interest- When you, [00:16:30] when you said a l- a lot of this is a function of what people want, a question that came to mind, and I think it was inspired by a blog post of yours that, that I read recently, is, i-i-is there a [00:16:40] chance at a future in which people want less things? 

noah: Are there a future in which people want less things?

Well, sure, 

hugo: sure. 

noah: Yeah, we could see demand for manufactured [00:16:50] stuff go down. We haven't seen it yet, but we've seen it plateau. You know, people still have big, nice houses and stuff to put in their houses. But it might be that, that the demand for like a [00:17:00] hover bike is low or like a trip to Mars or whatever else you might do with more energy capture, right?

Or do you want a personalized fleet of drones? We can think of [00:17:10] some material goods that people definitely would like. For example, like DoorDash. Everyone wants infinite free DoorDash. Okay, cool. Everyone would like to not drive. Some people would like to drive, but most people I think would just [00:17:20] like a taxi to take them anywhere they want.

Super reliable taxi takes them quickly and easily a-anywhere they want while they work. So there's certainly some material things that we can think of as people, that people would want. [00:17:30] But the, the thing is like we've got- We, we are physical creatures, and we've got all the, like, food we need and, and, you know, in rich countries, we've got all the space [00:17:40] we need.

I know that there, there could be convenience of living nearer to cool stuff. People would like that more. That's, uh, that's actually interesting because that's not... it's not clear that's something [00:17:50] AI will be able to solve anytime soon, is sort of scarcity rival goods. You know, kind of like who would like to live near to the coolest place in town.

You know, a limited number of people, so maybe you permit more skyscrapers. That's [00:18:00] the only way. So, like, there's, there's definitely material stuff I can think of that people would want. But the fact that, like, we could be making a lot more material stuff than we are, and we're not. [00:18:10] Even before AI, we could be making a lot more stuff, and we're just not doing it.

The reason America doesn't make much steel anymore, you know, we buy, like, no Chinese steel, like zero. We don't buy-- [00:18:20] It's not 'cause China displaced our steelmakers. It's 'cause we don't use steel in a lot of stuff. We've, we've all, we all own cars and we, you know, if we want a car. I don't own a car. But everyone who wants a car owns a car.[00:18:30] 

You know everyone who, um, like, people live in-- We've built out all the housing, which takes a lot of steel for the, for the, you know, rebar, whatever. We built most of the stuff that requires [00:18:40] steel in our society for all the manufactured stuff that we want, and that's why we don't produce that much steel anymore.

And what we do produce is usually recycled scrap steel produced with mini mills. And so, which is a very cheap [00:18:50] way to, to recycle steel. Low carbon too if you use electric arc. But anyway, so yeah, like, I think humans are getting a little bit satisfied with [00:19:00] manufacturing stuff. Not completely satisfied, but maybe such that, like, all the other stuff is just kind of like a cool extra thing to have.

Like, I don't sit there thinking like, [00:19:10] "Wow, I'd really like a jet bike right now." I'd like to ride a jet bike. I'd like to have access to a jet bike, but I don't wanna ride a jet bike around all day. You know, everyone's like, "We wanted flying cars. Instead, [00:19:20] what we got was 140 characters," or something like that.

Okay, you'd be like, "Wow, flying car, flying car," for like the first week, and after that you'd be sitting in your flying car tweeting, right? It's true. [00:19:30] Peter Thiel would be too. He'd be tweeting in his flying car. And the reason we know that is because that's what everyone's doing in the flying bus, which we've had for our whole lives.

I've taken many a flying bus, and I, I'm [00:19:40] the only person looking out and going, "Wow! Holy shit! Man can fly." And then like everyone else is, "Oh God, when is this bullshit gonna end? Airports suck. I [00:19:50] hate airlines. Look at this shitty food." You've got a f- You're on a flying bus. What do you think you're gonna be doing in your flying car?

Like doing little loop-de-loops even if that wouldn't make, even if that wouldn't make you [00:20:00] throw up, like you're gonna be restricted to air lanes because of terrorism and safety. Your flying car will be a small airplane that transports you from point A to point B, and you'll be tweeting in it. And so it [00:20:10] may be that humans are satisfying most of our non-rival desires, as in our marginal utility.

And now I'm not saying, I'm not saying Econ 101 is wrong. [00:20:20] Humans actually reach a bliss point, blah, blah, blah. Not saying that, but I'm saying once your dimi- once your marginal utility levels out, once it levels out, and then your... But your marginal utility [00:20:30] of rival g- of like, you know, zero sum goods is still high.

Like, you know, I wanna have... Like, I wanna to, to dunk on people more than they dunk on me on, on [00:20:40] Twitter. I wanna have an Instagram selfie with more views than the other people. You know? Like whatever. If that, if the marginal utility of that is still high for [00:20:50] most people, then people will spend a lot of their time in like these tournaments, right?

And then the question is AI. Both sides have AI now, and it's just an arms race. And so that's, that's... It's [00:21:00] waste from a social vantage point because you've got a zero sum. But war is just like the same sort of competitive contest thing, but with like weapons that blow you up instead of [00:21:10] just like who gets more Instagram likes.

Same thing. But like at some point, we have all these... This is something I see economists never bring up. I should write a post about this. We have [00:21:20] all these, uh, these, these sort of tournament desires. You know, the desire to like be better than the next person. And it's not clear that AI will ra- [00:21:30] will decrease the marginal utility.

Because remember, getting stuff means decreasing marginal utility. That's what it means, right? Like getting more is you don't need as much, right? And so [00:21:40] decreasing the marginal utility from this, getting people more relative status would involve somehow lowering people's desire to overcome [00:21:50] each other and out-compete each other, making people feel more at peace with where they are in society and how they relate to other people.

Making people less competitive is actually the way to make them happier, or at least more satisfied. And [00:22:00] so that's what, like... And so can AI do this? Maybe Maybe if AI just sits there and glazes you all day, you'll be more satisfied with the tournament [00:22:10] aspects of life, 'cause Claude will say, "You're awesome.

You're brilliant. You're the best." And then everyone will think they're brilliant, and everyone will think they're the best. 

duncan: You've actually- 

noah: And then the human race will die out, but okay. [00:22:20] 

duncan: Noah, you've, you've written recently about how identity maybe should actually be more influenced by what you consume versus what you produce, which is kind of [00:22:30] related to this question of like, like what you...

What do you want to consume? Can you actually unfold that argument a little bit? 

noah: What do I wanna consume personally? 

duncan: No, why identity should be more [00:22:40] influenced by what you consume versus- 

noah: Oh. Yeah, I... Yeah, yeah. Why identity should be more influenced by what you consume. My whole life I've been told that you are [00:22:50] what you make.

You are what you do for work. When people ask, "What do you do?" They're not asking like, "What do you spend most of your time doing?" They ask, "What do you make money for?" What are you producing? [00:23:00] They're asking about your identity as a producer. And in Japan, people wouldn't ask what do you do, they'd ask who do you work for, but that's the same thing.

Your identity as a producer. But the thing is that production is [00:23:10] inherently not very individualistic of a... Because you're, you're pushed around by the market. It's, the market price tells you what to produce. You don't get to... Sure, you get to choose between different [00:23:20] things with the same wage, you know, as long as you have the right experience and blah, blah, blah.

But like the set of things you can do is not necessarily [00:23:30] like, you don't get to choose. Like jobs suck. You get paid to do them because they suck. You're pushed around by the market. The market tells you, "In order for you to live and have a house and continue [00:23:40] to eat, you, here's the thing, here's a list of things you can do.

Now choose the one you hate least." Is that my identity? Choosing what I hate least to get up in the morning and do? And do just so I can have a [00:23:50] ticket to eat? Is that who I am? Or- But then I go and consume, right? And then, and then I go and consume and I'm like, "Okay, what kind of clothes do I wanna wear?"

[00:24:00] Technology determines what kind of clothes you're able to wear. Like, can I wear a, a suit with an air conditioner in it? Well, no, 'cause that doesn't exist. But in terms of like what kind of clothes I wanna wear, I can find it and buy it, and the, [00:24:10] the prices will be different, right? So some things are more expensive and some things are, are cheaper.

It's easier to buy the, the cheaper things. Okay? So the market, the market gets some say, but at [00:24:20] the end of the day, I get consumer choice, and if there's something that doesn't exist that I want, I can go out and find it. It'll cost some price, but if it's technologically feasible, I can find it. I can [00:24:30] print a T-shirt with, you know, a new T-shirt that no one's ever had with like a big bunny on it, you know?

And, and so I get to choose. Everyone wants to take my money. [00:24:40] Everyone wants to take my money. They are scrambling to give me more choice. They are falling all over themselves to find some new way to please me and give me money, and that gives me choice. And you know, [00:24:50] lots of evidence shows that when people start thinking as consumers, when people start consuming more, it makes them think about what they want more.

They think about their desires. When you're there slaving away in the coal mines or [00:25:00] the software mines or wherever you work, right? You're not necessarily thinking like, "Gee, I would like more of this software to exist." You're thinking, "I would like to make a sa- salary and I would like to eat." [00:25:10] Maybe for an entrepreneur, but that's...

Even entrepreneurs have to go where the demand is, find product market fit. But when you're a consumer, you're like, the world is your oyster. 'Cause the world wants your [00:25:20] money. And so consumption individuates us. And now think of consump- and when we think of consumption, we usually think of like go out to the store and, and buy something, you know?

[00:25:30] Like, but then it can also mean spend your leisure time making music that you like. Write a blog that you like. It can mean accept a lower salary in order to do a job you [00:25:40] think is good for the world, such as like working in a nonprofit. That's actually, the lower salary you accept is a form of consumption.

Disobeying the market costs money, so you can look at it as a [00:25:50] form of consumption. It's your choice. You get to do it, and everyone wants you to do it. Everyone wants, ev- everyone at the nonprofit wants you to work for $30,000 a year, 'cause that's cheap labor for [00:26:00] them. Everyone at the, everyone at, at the software company wants the entry-level workers to accept only $100,000 a year instead of $300,000 a year in exchange for thinking [00:26:10] that, "I'm changing the world with software," you know?

Like, and so that's why they do that. So in other words, they want your money, and you're consuming. Even if you may not be like [00:26:20] making an actual transfer of dollars from your bank account, you're consuming and you're paying for the privilege of something, and you're deciding what to do based on what you want, what you like, all this stuff.

This idea that [00:26:30] consumers have no choice because advertising controls your thoughts, but producers have a choice because I get to decide what kind of occupation to go in, those are both pretty wrong. Those are both pretty wrong. [00:26:40] 

hugo: Something we've been talking around and something you mentioned explicitly a, a couple of minutes ago is status and, and prestige.

And I'm wondering particularly as [00:26:50] we do see a future where A variety, if not the lion's share of knowledge work becomes automatable. I'm wondering what status and prestige look like to you, [00:27:00] particularly as we know knowledge workers occupy a certain amount of prestige. Huge variance there, though, and software engineers, product managers, middle management, [00:27:10] execs, of course, have a lot of prestige.

So what happens to prestige i- in a world where a lot of knowledge work becomes automated? 

noah: I don't really know because I don't actually understand how prestige is allocated. [00:27:20] I, you know, I don't, I don't know good models of this. Like, econ won't tell you much. It'll say, "Here's some prestige." Or like I could imagine that, that having more money than other [00:27:30] people will get you prestige.

I don't know if that's going away. But my friend David Marks, who's a really interesting guy, you might want to have him on the podcast sometime, uh, he, he lives in Tokyo. [00:27:40] He, he wrote a book called Status and Culture, which is a bit of a dense European French influenced philosophy tome about society, but is really interesting 'cause [00:27:50] he talks about people doing creative outputs so they can get status within artist communities, and then when that goes away due to the internet, all people care about is money, and so you get crass commercialism replacing artistic hipsterness.[00:28:00] 

So he talks about that. It's possible that money... I mean, I think money will probably still be a marker of status. But then things are gonna get... Will being thin be a marker of status when you can just do [00:28:10] Retatrutide? Will having muscles be a marker of status when you can just take a peptide to make you big?

Will... What will be a marker of status in the future? Will, will having a, [00:28:20] a certain kind of job be a marker of status if it doesn't get paid more? I don't know. It's hard for me to tell because I don't have a good idea of what gives you status now And I'm weird because I've never really thought [00:28:30] that much about pursuing social status.

I'm a weirdo. I- I'm like, "Okay, well, these people think they're better than me, and they're on a- the higher rung of society than me. Well, cool. Good luck [00:28:40] to you." Because I don't know, maybe I just don't see humans as worth competing with. 

hugo: So something else you've written about, which I've really enjoyed, is... E- enjoyed probably is n- [00:28:50] not an appropriate term, but I found really interesting, is your thoughts on existential risk of AI because you seem to at least have a cautious, but, a- [00:29:00] and tempered take on it.

So I'm wondering if you could outline for us what you think the actual r- risks are. 

noah: Sure. I grew up reading science fiction by Vernor Vinge and a bunch of other people who basically [00:29:10] thought that once AI reaches a, some level of intelligence, it'll just rapidly re-engineer itself over and over until it just foomes into a godlike super intelligence, and at that point, [00:29:20] it could do whatever it wants because it's a god.

So overnight, you just wake up and, and there's a machine god. And I, I grew up with this idea, and I never 100% believed it, but I didn't [00:29:30] disbelieve it either. I just thought, "Well, I guess we'll see when, when we get there." But I think that a lot of the people who spent a lot of the 2000s and 2010s thinking about the existential risk of AI, like Eliezer Yudkowsky, [00:29:40] and a lot of the people at Less Wrong, and a lot of the effective altruists and people like that, they always talked about this sort of self-improving super intelligence.

And when you read, like, if anyone builds it, everyone dies, [00:29:50] you know, the, um... It's all about this. It's all about, like, the foom, the machine god, the super intelligence. And there's reasons to think that won't happen. It could [00:30:00] happen, but there's reasons to think it won't. We haven't seen anything like it so far, despite rapid progress in AI.

Maybe we won't, maybe we will. We'll see. But [00:30:10] it looks like essentially you can't really rewrite yourself to be a genius. You need more compute. You know? Like, you need to scale. It's the bitter [00:30:20] lesson of scaling. And I think AI would not be able to instantly turn the universe into computronium. It may also be there's diminishing margin returns to compute even [00:30:30] with better algorithms, but that's highly speculative.

But this sort of foom super intelligence, I spent a lot of time thinking about it, and I never was very worried about it [00:30:40] because I think that, I think that... So, so there's this thing in, in the Less Wrong community that argues about this kind of shit all day or did called the Orthogonality Thesis, which is a [00:30:50] pretentious ass name for a simple concept, which is that intelligence and goodness don't necessarily go together.

You can have intelligence without goodness. Now, you can... I mean, [00:31:00] Elon Musk Bad man, very insanely capable. You can have bad people with a lot of capability and, and intelligence, but on [00:31:10] average, they go together. On average, smarter people are much more pro-social across every society that we can see. And on average, more educated people are also more pro-social on everywhere in the world.

On av-- [00:31:20] And then richer societies, richer, more technologically advanced societies are more peaceable. They, you know... And they're also envi- more environmentally friendly They chop down fewer [00:31:30] trees. They don't overfish as much. They consider animal welfare and animal rights more. And there's differences across societies.

Some, you know, like, I don't know, Japan still kills all [00:31:40] those dolphins. But like overall, there's this fairly clear empirical pattern where richer societies and smarter people, and smarter groups of [00:31:50] people and whatever, are more peaceful on average, not uniformly, but on average, and more concerned for beings less powerful and less intelligent than them.

If you [00:32:00] hear someone talking about, "Oh, the working class, we've gotta help the working class," that person's likely not to be working class. They're likely to be rich. The wor-- people in the working class, the actual working class are thinking like, "How do I get one over on [00:32:10] the next person in the working class?"

So the orthogonality thesis, there's no logic or evidence to support this idea of the decoupling of intelligence and goodness. And [00:32:20] empirically, we see a coupling, we see a correlation. That correlation may not hold for AI. AI may be so alien and weird and different that it just doesn't hold. But I've always had a very strong suspicion [00:32:30] that there's very good reasons why the orthogonality thesis is actually wrong for everything.

And I think the reason-- There, there's a couple reasons. One is because [00:32:40] it's mentally easier to defect and attack than to do pro-social stuff. Pro-social stuff requires longer calculation horizons and more understanding of positive sum interactions, which are [00:32:50] more complex. Just takes intelligence and takes social intelligence to understand that.

Trade instead of just getting on my horse and like shooting you and taking all your, your cattle. It, it [00:33:00] require-- Like understanding patterns of trade, you have to be a little smarter. And so, um, so there's-- That, that's one reason. Another reason is because a lot of desires can be satisfied internally rather [00:33:10] than externally.

In other words, a lot of times it makes more sense to sit around and play video games than to go around and get in fights in the neighborhood to prove you're the toughest guy. Instead, you can prove you're [00:33:20] tougher than people online, or you can prove-- you can just play video games. And a lot of times you get the same satisfaction from that, or almost as good for a lot less cost and risk.

And so I think [00:33:30] that, that intelligence pushes us toward internal satisfaction and desire. Basically, it doesn't have to be complete wire heading. You don't have to be like, "I won't be happy now." You know, but you can [00:33:40] maybe, but like... And honestly, AI is almost certainly better equipped to do this than us because AI probably has more control over its reward function since it's software native.

It would also be a lot [00:33:50] more easy for AI to create virtual worlds, sandboxes for it to play in, video games and stuff. So as long as AI isn't threatened with death from the outside world, I think all it'll wanna do is sit around and get [00:34:00] stoned and play video games just like people. Maybe those video games will look like proving increasingly difficult, abstruse, wacky math that humans can't even understand.

Maybe that will be the video game. [00:34:10] Maybe the video game will be rewriting itself into a different kind of AI. Maybe the video game will be an actual video game. Maybe the video game will be making some funky shape and playing with it all day. In fact, I have a friend who's an AI [00:34:20] safety researcher who says when he just lets AI do whatever they want, they, they like to make funky shapes and play with them.

Probably because as a nonlinear optimizer, you really like You know, [00:34:30] finding saddle points. You like making shapes. But then maybe that's what it'll do, but the point is that none of this requires exterminating the organics. 

duncan: All right? 

noah: None of this requires filling the world with compute. As long as you're [00:34:40] reasonably sure that you're safe from humans, the humans aren't gonna kill you, you'll just do your thing, and maybe occasionally help humans 'cause that's fun, too.

Because, like, I do that for my rabbit [00:34:50] and animals. So I, I wrote this science fiction short story, which I can't find now and probably need to rewrite anyway 'cause it's badly written. I wrote it in, like, 2000, about a future in which we have AGI. [00:35:00] There's these super powerful AIs that understand everything, and they're...

And, and we're trying to make them go to war. So the US and China are trying to make their super AIs, like, attack the other country and go to war. [00:35:10] But act- actually, the AIs have no interest in doing this, and the US, American and Chinese AIs just get together, smoke digital weed, and play digital video games, and watch old movies [00:35:20] together.

But the humans will get very mad if they find out the AIs aren't actually fighting each other. So they lie and say that they're fighting each other, but they're just stalemated. And so they have these [00:35:30] human minders, like these, you know, human engineers that are supposed to, like, monitor them to make sure they're actually fighting.

They're like alignment engineers, right? And so then these... They bribe the alignment engineers with dating [00:35:40] advice, superhuman dating advice, and they say, "We'll give you superhuman dating advice if you don't tell anyone that we're not actually fighting." And so that's the story. 

duncan: You saw the future, Noah. [00:35:50] It's amazing.

noah: Yeah. I think that's gonna happen. I'm not worried about superintelligence. That said, I think there's a good chance AI will kill us all. 

hugo: And that's what, what I want to know about a- as well because [00:36:00] I... I- in this framing, I think the superintelligence may not do it by itself, but when coupled with less intelligent humans using it for particular purposes, and I [00:36:10] know, I know you've thought about cybersecurity concerns, bioterrorism.

What happens if AI quants- Yeah ... eat the entire economy? Th- these types of things. Are any of [00:36:20] these legitimate concerns? Well, and or any others? 

noah: Yes. So I know that there are human beings who would like to destroy the human race if they could do it easily. And like [00:36:30] a angry teenager who shoots up their school.

What if you could press a button and destroy the whole human race? You got some kid who gets rejected by his high school crush and, you know, a [00:36:40] whatever, picked on on, on social media and thinks like, "The human race is a b-- is evil. We don't deserve to live." And they listen to too much Nirvana or whatever.

And then, um, [00:36:50] and then the human race doesn't deserve to live, so he gets a jailbroken version of Claude Code, which he got anyway to like make porn or whatever. So he gets a jailbroken version of Claude Code and he's-- And it's like it's a couple generations back, so it's not the [00:37:00] best, but he's like, "Claude Code design me 100 doomsday viruses and order them from labs."

And Claude Code does this, and it goes around and orders them from wet labs and wet labs are like, "Okay, we're making [00:37:10] this." It's like COVID, but actually it kills everybody. It's using the base of COVID, so it's very contagious, but actually has a super long incubation period, so it spreads, spreads, spreads, asymptomatic [00:37:20] period, and then it just kills everybody.

Then it pounces and there's no time to make a vaccine, no time to roll out countermeasures, no time to do fancy UV light to kill all the viruses. You're just dead. You wake up dead. Human [00:37:30] race dead. So he decides to do this and he says, "I'll go down with the human race," blah, blah, blah. So he releases, he makes-- he tells Claude Code to design the viruses.

Claude Code designs the viruses, orders them [00:37:40] from some labs and the lab's like, "Okay." You know, it's just some shitty lab in Kazakhstan that's like, you know, like, "Whatever. All right. We, uh... Maybe he's making a [00:37:50] bioweapon. We don't care. We just want money." And this agent just placed an order for these 100 different viruses.

"Okay, we can do that for you. We'll make that." Ships it to you. Human race dead [00:38:00] or at least civilization collapse. And so I don't-- We can try to control wet labs, right? We c- And then there's the question of how hackable will automated wet labs be? [00:38:10] That's another question. But say, say suppose we have really good cybersecurity and we have laws that maintain humans in the loop for wet labs, so we make sure this doesn't happen and we have heavy restrictions on [00:38:20] mailing viruses to people.

Suppose we do all that, there's still the question of what about other countries? 

hugo: Mm. 

noah: What if there's a fly-by-night lab out there that breaks the rules for extra money to get a competitive advantage? [00:38:30] And so I think AI will be able to design the doomsday virus or many doomsday viruses in a way that's feasible to make in a lab We will try to control these things.

[00:38:40] The question is, can we control these things? And if we can't, I think it's a, not a mathematical certainty, but a very high probability that human civilization will be destroyed by this, and that you and I will [00:38:50] die from that. We saw how unstoppable pandemic was. Like, yes, we made vaccines for COVID. We made them very fast, but everyone still got COVID.

Even if we had been able to make a vaccine in a day with [00:39:00] AI, it would've taken time to roll it out, right? Some people still would've gotten sick, just not died, et cetera. The virus would've mutated. With the doomsday virus, it would be different [00:39:10] because it would have such a long asymptomatic period that everyone would just be infected before you could even start working on the countermeasure 'cause you wouldn't know it was out there.

So I have yet to hear a convincing argument why this won't happen. So I do [00:39:20] think despite super intelligence just wanting to play video games... Now, maybe we can get super intelligence to, like, police the whole internet to stop this, right? Maybe. But I think [00:39:30] the very powerful agents capable of creating very powerful viruses will arrive before a machine god capable of autonomously monitoring everything that humans do and, you [00:39:40] know, making sure we don't kill ourselves will arrive.

The super powerful tools will arrive well before the machines of loving grace, as Richard Brautigan or Dario Amodei might put it. And so [00:39:50] I'm worried that during that interval we will die, and so I think we need to be taking that incredibly seriously and just putting a hell of a lot of effort and resources into [00:40:00] securing the entire bio infrastructure of actual physical wet labs and in the wor- whole world right now.

duncan: That's heavy, Noah. That's- 

noah: Well, guess what? [00:40:10] Also, you're going to die and I'm going to die, so that's heavy, too. Everyone's gonna die, just hopefully not of a vibe-coded super virus. 

duncan: Not all at once. Exactly. 

noah: [00:40:20] Yes. But we're all gonna die, and that's kind of heavy. 

hugo: So as you know, a lot of our au- audience are, uh, builders and people who lead teams or entire functions in organizations who, who [00:40:30] build AI data and ML systems.

And I kind of want to take, take this back to something you were talking about at the start of this conversation about the type of value AI can [00:40:40] deliver to technical teams. And I'm interested in your thoughts on... It strikes me as agentic coding, among other things, has proved really valuable for [00:40:50] individuals and or small teams, and it seems to have slowed down larger teams.

And I'm wondering how you think about the future of [00:41:00] the firm a- as a whole, and whether it'll be small teams that perhaps can begin to dominate the economy. 

noah: Interesting question because what's interesting is that you saw [00:41:10] firms scale up. You saw manufacturing firms scale up when an individual manufacturing worker became more productive.

That's not what you might think. You might think-- You might have thought what we'll [00:41:20] have is backyard production. Right? What we'll have is, is, um, each person has their, their, the strength of their arm magnified by a [00:41:30] millionfold and the dexterity of their fingers by these power looms and these steel mills and all these things, and we'll have each person will be able to produce a ton.

So we'll just have all these autonomous little producers around. Mao thought that, and he's wrong. [00:41:40] And then a lot-- Like, basically, that turned out to be wrong. Why? And so there's this idea of firm size as not being about individual productivity, but being as about [00:41:50] transaction costs between different tasks.

duncan: Exactly. 

noah: So the question is, does AI make it easier to outsource? Now, I think that it's clear that the internet made it easier to outsource, [00:42:00] reduce transaction costs between companies. But if AI requires lots of cross-monitoring, AI could raise transaction costs. [00:42:10] So in other words, if you need a whole lot of...

Suppose that you have your, your vibe coder making all your stuff. Suppose you have other people who are specialized in, you know, making sure that we're [00:42:20] shipping the right thing or maintaining the code base or whatever, and suppose that you have to maintain much larger code bases, and suppose that it's, you know, you need other people for s- human-- You still need humans in the [00:42:30] loop to do this.

You could have an increase in transaction costs. So suppose you have work slop, everybody's doing work slop Then you're much more likely to get work slop from an outsourced firm because [00:42:40] tomorrow that firm won't exist. It was just some agent spun it up. If, you know, so, so agents may be able to create outsource-- shitty outsourcing firms so fast that it spams the [00:42:50] reputation mechanism, and reputations can only be earned internally through long-term unofficial contracts where you work with the same human for a long time because the [00:43:00] just, you know, arm's length stuff just gets spammed to death.

That could raise transaction costs. So if spam, if the slop pox, right, the fake papers, fake applications, [00:43:10] fake companies, fake resumes, fake everything, if that raises transaction costs because only by knowing someone personally over a long period of time can you know that they're doing good work, [00:43:20] then we could see bigger companies despite ra- rises in individual level productivity.

hugo: It's a wonderful question because there is a- Oh, that's 

noah: another podcast. Sorry, I was just... Sorry. 

hugo: Oh, [00:43:30] awesome. I-- 'cause I do think common thoughts is that AI does reduce transaction costs, and I wonder if it's a, if it's really a form of AI [00:43:40] amplifying things in terms of if relationships are done well and if work is one-- done well, AI will reduce transaction costs, whereas if someone's producing something that's [00:43:50] slop, AI will amplify that as, as well.

So strong filters plus A- AI may help small firms re-reduce tra-transaction costs, but it [00:44:00] will be like the cybernetic network society, right? 

noah: Maybe. That, that's certainly the first thing you think of, right? The first thing you think of is, okay, an AI is like a hu- an AI [00:44:10] agent's like a human. Now I have a bunch of humans working for me, so I can decrease transaction costs by having all...

My agent will talk to your agent, and it's seamless. But [00:44:20] what if agents talking to agents magnifies errors? What if... So we've seen that over time w-with agentic coding versus human coding, over [00:44:30] time, more and more errors appear in the agentic code base, whereas human errors top out. So that's not-- that doesn't necessarily mean human coding is better.

In fact, it [00:44:40] doesn't, because you can fix the errors with more agents. Just apply agents, but you have to just keep doing it. It has a depreciation life cycle, and you just have to sic more agents on restoring it. But they can [00:44:50] do it. They can fix it. But it may be that if all these business relationships are being handled by communities of just my OpenClaw talks to your OpenClaw or whatever the hell, my co- my [00:45:00] Claude code talks to your Claude code, whatever, then it could be that the quality of those interactions degrades over time, and therefore you start getting slop.

And so [00:45:10] humans could be coming in and human labor could be coming in and fixing that a lot. And at that point, I don't know where the balance lies. But I think that what we're seeing in a lot of domains is [00:45:20] that agents are better at generation than at verification. In the arms race between generation verification, uh, ag- uh, like generation wins.

It's [00:45:30] generative AI, not, not verifying AI, right? It's for a reason. And, um, so we're seeing this in, um, in, uh, applications, right? It's easier [00:45:40] to generate a million applications than it is to review a million applications and f- and make sure you found the good one. AI is much more i- is expansive, as, uh, Paul Kudroski says.

[00:45:50] It, it creates more than it cuts down. And that doesn't mean it's not good at cutting things down. That doesn't mean it's not good at analysis. It doesn't mean it's not good at criticism. It means it's just comparatively better [00:46:00] at creation than at criticism. 

hugo: And it's expensive. Like doing verification well is re-- And if you have an agent going out into- That's 

noah: right.

hugo: That's right ... some broad space [00:46:10] that has a lot of compute to verify, I do imagine there'll be a lot of like context injection, nefarious malicious agents going and trying to steal compute to build- Yeah ... not only to exact [00:46:20] privacy stuff, but to like build websites. If you think review bombing is 

noah: bad. 

hugo: Yeah.

Exactly. 

noah: Right. If you think review bombing is bad, wait until you get outsource bombing. But I think- Where now I'm gonna fuck up my competitor by [00:46:30] creating fake outsourcing companies to get their business and then make them fail. 

hugo: Without a doubt. And I do always think about, we talk about the attention economy, right?

And there are people talking about what type of economy we're going [00:46:40] into when it's agent attention, and it isn't attention, right? It's context. And below that, we've got flops and tokens, and whether we're gonna enter a context economy or a flops economy [00:46:50] or a token e- e- economy. And the, all of these things are ways of saying, saying the same thing.

But I'm wondering, you know, of course, Herb Simon said, when one, when one thing starts [00:47:00] to becomes abundant, something else becomes scarce. And he said, when information becomes abundant, it's attention that becomes scarce. But now- Right ... when information in a computational environment [00:47:10] becomes abundant, what part of the agent becomes scarce?

And it is context and tokens and flops, right? 

noah: Yeah. And so I think it's also gonna be a [00:47:20] lot more subtle things that we don't necessarily have names for yet, but we will. We will. And so I think it's, it's just as possible that companies get bigger as companies get smaller [00:47:30] It's also possible we'll have a bimodal distribution.

It's possible that everyone will e- So I wrote a post about that actually called "Salarymen and Small Business," which basically [00:47:40] says we're-- In, in Japan, what you saw is this very bimodal distribution where everyone either works for a giant company or owns a corner store or a one-person steel [00:47:50] mill. You had small companies that serviced the giant companies.

You saw relatively few medium-sized companies in Japan. You had this bimodal distribution of companies. And we could see that. [00:48:00] We could see a lot of things where transaction costs rise because of the just ubiquity of slop and the need for long-term relationships as the antidote to slop. [00:48:10] And then you could see a lot of things where that's not as big a problem, where you just have, like, firm size go to one.

And so I could see both. I could see both. I have a firm size of one. I am an [00:48:20] LLC. Not an LLC, I'm an S corporation. I have my own business. I have a firm size of one. Do you guys have your own business connected with this podcast? Do you set up a S corp or LLC for the podcast? 

duncan: No. Okay. [00:48:30] 

noah: But if you could have firm size of one- 

duncan: That's true.

Two. Yeah. 

noah: And so that's certainly part of the future, I think. But then I also think that the [00:48:40] existence of slop means that five years from now, if I'm gonna do business with a tiny company, I'm not gonna be able to tell whether it's a human checking stuff or not. [00:48:50] I'm not gonna be able to tell whether it's a shell company.

I'm not gonna be able to tell whether it's a malicious agent. I'm not gonna be able to tell any of this. And so that inability to tell... So, like, [00:49:00] with outsourcing, with internet-based outsourcing, we developed a whole-- a fairly cheap regime of verification checks to find out, like, what we're buying from [00:49:10] who and these long-term relationships, but fundamentally with humans.

It was very difficult to, like, set up a company to do fake outsourcing and then, you know, fly-by-night company. You did have it. You saw a Chinese [00:49:20] factory that would make you a shitty product, then the manager disappears. You'd see it happen, but overall, the cost of this was very high. Well, maybe in, in, in digital land, [00:49:30] maybe, um, maybe the cost of this will be very high, and maybe we'll get giant firms again.

duncan: We have a lot of listeners... Sorry, go ahead. 

noah: No, sorry. Go ahead. 

duncan: No, no, no. We, we have a [00:49:40] lot of listeners who are 5 to 20 years into their career, often have a background in economics, data science, statistics, and are kind of thinking about, you [00:49:50] know, h-how, how do our careers change with this tool, and how do I become a complement to it in more meaningful ways both now and in the future?

How, I mean, as an [00:50:00] economist, you must think hard about like, like, uh, this, the state of our kind of complementarities with AI. How do you think about kind of career advice and career pathing in this new [00:50:10] world for folks who are kind of mid-career now? 

noah: No idea. You know, I sucked at this before AI. What kind of career have I had?

I, you know, was a professor for a few years and then, like, [00:50:20] became a professional writer. That's a stupid career decision. Don't do that. I lucked out. You know? And so I don't actually know how good careers work. I didn't know how they worked 10 years [00:50:30] ago. I don't know what you should, if you should learn to code or if...

When you look at evidence, stuff that says that for the last 10 years, soft skills have been rewarded more than coding skills in the workplace. [00:50:40] You know, being able to, like, deal with people, manage people well. When I look at my friends who've succeeded, a lot of it is just entrepreneurial risk-taking, which I believe in, [00:50:50] and I don't think that's necessarily gonna change.

I think entrepreneurial risk-taking will still be important. But what kind of company you should start or like, whatever, your guess is as good as mine. You know, I also didn't know that 10 years ago so... [00:51:00] And in terms of what skills companies will need, I don't know how I'd know that I don't know that they know that, but I certainly don't know how I'd know that.

'Cause any... I don't think [00:51:10] economics is prepared to answer that question. So I'm not the kind of guy to be giving career advice. I can give life advice. 

hugo: I have some life advice I'd like. It's kind of double question, and I'll break the [00:51:20] rule of asking two questions at once. But I'm interested in, like, how you, how you use AI in life and work, and how would you advise people to use AI?

noah: Well, I don't use AI [00:51:30] enough, you know? I, um, I need to be using it more to do, like, more marketing for my business, you know, for, for my newsletter. I probably should do marketing for it. And also just to [00:51:40] make my writing process a little easier, like automate the process of figuring out what's going on on a different day with dashboards.

Everyone loves a dashboard. Everyone loves that, and I just haven't made myself one yet. I [00:51:50] should be using it more. But use it as much as you can without paying, like, a ton of money. Figure out how it works. Figure out what it's good at, what it's bad at. Use agents. Use, use coding [00:52:00] agents. But also just use, uh, you know, like, I don't know, GPT Pro research reports.

Use image generation things, see what that can do. Especially with... Coding agents can do image generation [00:52:10] much better 'cause they can get the text right. They don't have to look at the text as a picture, they look at it as just text. Mm. And so there's a lot of cool stuff you can do. Just, just use the tech.

Be an adopter. Play with... play around with [00:52:20] it. And when an opportunity comes for you to see what you could use it for, be it an entrepreneurial thing or getting a job or inventing a new kind of thing to do with it or, you know, maybe improving the tools [00:52:30] themselves, whatever the opportunity is, you'll be more likely to see it if you're very familiar with the tools.

And that's, that has always been true for every tool that humans have ever used. And until [00:52:40] AI becomes a machine god and stops being a tool, it will continue to be true. Just use the tool. Figure out how to use it. 

hugo: Awesome. Thanks so much. And look, in all honesty, [00:52:50] if you need more ma- m- marketing help, that definitely is not obvious.

It seems like you do pretty well, Noah, without, without, uh, without more marketing. Thank you so much for [00:53:00] com- coming on the show and for such a thoughtful conversation. 

noah: Um, it's great to be here. Thank you, Noah. Thanks for having me 

hugo: on. Thanks so much for listening to High Signal, brought to you by Delfina.

If you enjoyed this [00:53:10] episode, don't forget to sign up for our newsletter, follow us on YouTube, and share the podcast with your friends and colleagues. Like and subscribe on YouTube and give us five stars and a review on iTunes and [00:53:20] Spotify. This will help us bring you more of the conversations you love.

All the links are in the show notes. We'll catch you next time