"Am I Crazy?" [Wading Through AI - Episode 3]

本站添加: 2026-05-10

[00:00:00]Welcome to the third episode of Waiting Through AI. In this episode, Dimmitri and I discuss the all too common occurrence, it seems, of people having very different experiences when they sit down to use AI for seemingly the same sort of tasks. Maybe you are using AI to do something that it's supposed to be very good at and you're prompting it the way people recommend. You're trying to follow all of the advice that you've heard and it's just not getting the job done. you end up spending more time babysitting and cleaning up after the AI than you would have if you just done the work yourself. Or maybe you have the exact opposite experience. Maybe you found that it was really easy to get AI to do this particular part of your job and it's been going great and yet you go online and you hear all these people saying that they have tried to use it for that same sort of thing and it didn't work at all, that it slopped, that the quality was low, etc., etc. And in either case, whether it's positive or negative for AI, you're getting this feeling like, am I crazy? Is the world crazy? Like, am I just uniquely bad or good at whatever this thing is that I'm trying to do at AI and everyone else has skill issues or has more skills than me or whatever, right?

[00:01:14]So in this episode, we try to look and see if AI as a technology has some underlying characteristics that make this outcome actually much more likely than it is for perhaps other kinds of emerging technologies.

[00:01:29]Now, I'd also like to remind everyone since we are three episodes in now and there may be people who are coming to this uh episode for the first time. I just want to remind everyone that Dimmitri is the AI expert. I am not. I don't do AI for a living. So, keep in mind when we're talking, it's Dimmitri's perspective that you should be focusing on. He really knows what he's talking about. I am just here to move the conversation forward and to just bring up points of discussion. So, you want to focus on what he says and keep in mind that the things that I say are always from the perspective of an AI outsider.

[00:02:02]So, Dimmitri, last time we actually talked about kind of how companies themselves are, you know, posting things, how their like PR department might be sort of doing very different things with the spin or the hype than their actual engineers. And then, you know, when social media gets a hold of it, it kind of spirals out into something different. Um, and you kind of gave us a breakdown of like, you know, what are the ways you can kind of assess like claims being made about AI. Uh, and this week you kind of want to talk about something that I feel like sort of related to that, but it's from a very different perspective. And this is like what h how do we think about what we're hearing from different people with experience like using the AI like the other side of it like like no one necessarily with a particular thing to sell, not a company maybe selling AI, but people using it. Um, and I think, you know, when we talked about this briefly beforehand, you kind of brought up an interesting point, which was even for the same industry, even sometimes within the same industry, within the same almost use case, you can hear very different things from different people about whether AI is useful to them or not, whether they're being forced to use at their job, whether they're not, and like this kind of very different thing.

[00:03:12]and and we're and so like how do we make sense of this kind of you crazy like tons of wildly different claims coming from different people almost doing the same thing.

[00:03:21]>> Yeah. Exactly. Uh and so the um the way I think about this is um many people I talk with um uh in uh in both the AI skepticism and AI optimism uh directions uh at some point feel like am I crazy or are the people around me crazy. So for the AI skeptic, >> yeah, >> they're they're thinking, >> I tried it. It barely works. I certainly would never ship it. I'm not I don't feel threatened by it. I don't feel compelled to use it, but I know someone who seems to be serious who is using it a lot and saying that it's changing their productivity and they're going to, you know, whatever, put programmers out of work or whatever, right? We talked a little bit about that part in the in the first episode. I don't mean that specific claim but just that that's that's uh one bit of crazym for the someone on the skeptic side. Uh there are people you may know fewer people like this but I know many like like this who are on the optimism side who are saying I don't understand my friends who don't see the potential here right like my like if you're not on the AI train right now you are crazy right you're going to be left behind and they don't mean this in a um I mean some of them mean it in a judgmental way and I I criticize that right like but many of them just mean it in like uh they see this as obvious for the future they already see it who they already feel like it's giving them a lot. Let me try to be precise. They feel like it's giving them a lot and they feel like the uh their friends who are not adopting it um and not just in programming in in art, in marketing, in finance like I hear this from people who are uh you know just like doing Excel driven by AI all the time now, right? So this is not related uh this is not only related to programmers. Um there's the this feeling of why are those people so crazy that they're not adopting this obviously transformative thing, right?

[00:05:11]>> Yeah. And I mean I think I can maybe give a a slight anecdote here to uh that that would encapsulate this. I I myself am fairly I consider myself more uninterested in AI than I am skeptical about it if that makes sense. Right. Um because like like I'm so so I think I'm not an an AI skeptic in that sense but what I would point out is like let's suppose that I was right. Um let's suppose that I was someone who was like nah this is just a bunch of hoopla this isn't going anywhere. So recently I posted on Twitter a thing where I said I'm sorry ex I posted a thing uh where I said hey um you know back in high school a friend of mine read this book and he said there was this quote in it and it always stuck with me. It was give them an inch they'll take a mile. Take away their inch and all they want is their inch back. And I'm quite cert like I didn't make this up. I wish I had. It's an amazing saying so believe me if I could take credit for this I would. I didn't. It was from a book that my friend read. He didn't make it up either. Um, and I was just like, does anyone know, like, has anyone read this book? And I actually got replies to this, four separate replies among all the replies from people who had asked AI and they each asked different AIs and they each told me that it was from a different book and it's from none of those books. The AI just made it up, right? It just said, right? So, you know, I I also would just point out like I feel like it's pretty easy to understand why somebody who is skeptical about this technology would have it.

[00:06:48]Like that's their first exposure to it, right? And they're like, "This sucks.

[00:06:52]How can I possibly use this thing that just gives just makes stuff up? Like, what am I going to use it for?" So, I do think there's like um I can understand where this confusion would come from even right off the bat.

[00:07:04]Does that make sense what I'm saying?

[00:07:06]Like, uh absolutely.

[00:07:06]>> Yeah. And that connects to connects to what I think we can uh dive into in a few different uh related directions which is um uh compared to past technologies uh the experience of AI is extremely uneven and I'll contrast that with um for example the experience of the transition to to wireless internet right uh the experience we had was all very I would say uh quite similar right that we had wired internet and we had to plug in and then at some point we got wireless routers and we didn't have to plug in and everyone had basically the same experience which was oh wow now I don't have to be at my desk I can carry my laptop around or I can get it on my later when we had smartphones I could get it on my phone right um and so everyone had uh no one was confused about what the Wi-Fi was doing and everyone had a very similar experience of the Wi-Fi so as Wi-Fi was rolling out I think it was uh not uh not socially disruptive right I mean it made a big Everybody liked it, but nobody was confused about, hey, why are we using Wi-Fi? What's the point?

[00:08:11]>> I would also say that uh while Wi-Fi is a like, you know, like similar to AI at this point, Wi-Fi, especially early Wi-Fi, didn't work very well. That's true of early Wi-Fi. But >> the way in which it doesn't work well is pretty intuitive to people. meaning like they kind of understand like probably because they had familiarity with like radio and television. They understand this concept of like there's a signal and like my distance to the signal and the things in between me and the signal and its strength and stuff. That's like all fairly intuitive whereas I don't know would it be fair to say that AI doesn't really have that property.

[00:08:49]People don't have any thing to compare it to where they're like I understand why it's not working well here or why it is. Right.

[00:08:56]>> Yes. And also unlike AI uh Wi-Fi mostly uh uh the failures were as you say um analogous to failures with like with like radio which was okay it's not connecting because it's not close enough or there's something in the way right you already have this concept of it's not connecting.

[00:09:13]>> Uh it's slow because the signal is not it's connecting but it's the straight signal is not very strong. That was something that was that you could understand. It wasn't anything that first of all it wasn't anything that you did as such right. It wasn't that you uh were failing to give the right input. Uh you know, contrary to the you may remember the the anecdote about Steve Jobs telling people that they were holding the iPhone.

[00:09:37]>> Yeah, that's true. They're like you you got to hold the case differently, >> right? Uh but the um >> you know, it wasn't that you didn't say the magic words and that's why you got a bad result. And also the bad result was uh in a way a boring bad result and not a bad result like imagine the Wi-Fi like if the if the connection was bad it just filled in the the bites with fake bytes and you got like a madeup web page right you as you asked for a web page >> yeah exactly >> and because it couldn't get the data quickly enough it just hallucinated a web page for you right >> um >> there also I mean while we're piling things on I would also say like there was also no incentive to dislike it like nobody Nobody was was being told they were going to lose their job because of Wi-Fi, right? Even even if they were like even if the mobile internet was actually going to displace some jobs. Uh that wasn't a thing. So no one no one really had any vested interest in it failing either. Like you weren't motivated to to criticize it when it failed. You're just like, "Oh yeah, the Wi-Fi doesn't work very well here and off you go." Right.

[00:10:40]>> Yeah. So unlike Wi-Fi, uh AI is an extremely uneven experience basically across the entire spectrum, right? Uh you can be someone who's never used AI and two people who've never used AI can go and use it and have very different experiences. Two people who have been doing it for 20 years like I have uh we can go and use it and still have very different experiences and very different opinions uh of those of those experiences. Okay. Uh similarly, even if you're not using it, if you are >> um I'll say subjected to it in the marketplace, right? Um that you hear things from um you know, some corporations are telling uh telling their uh their people you can't use AI, it's forbidden. Uh especially not on anything with sensitive data or whatever, right? That some at one end you're being told you can't use it. We don't trust it. it's legally uh legally legally uh unclear and so on. At the other end, you have people who are being told you have to use it. Not only do you have to use it, we're going to track how much you use and we're going to include uh how much you use as a performance metric. And we're going to talk about it at the time of performance review. And I I just want to mention for a moment um if you are not adjacent to the really uh quickly moving AI businesses right now, I know this sounds like comedy like this sounds like something out of Kafka to many people. Uh but I promise you this is happening right that that many people really are being tracked on their token usage and being told you have to be AI first. Why did you do this manually when AI could have done it? Why are you not spending enough on uh on AI and enough?

[00:12:18]Some people um there's like a power law of what is considered enough, right?

[00:12:23]Some in some places uh enough is like you have to be using AI at least once a week or you have to do a PR that's tagged with I used AI at least once a month, right?

[00:12:34]>> Pretty minimal like so you have to at least try it, right? That's one level.

[00:12:38]And then another level is you have to spend at least uh you know we're getting you a cloud max plan so you need to be using roughly $200 $200 a month of of cloud tokens. And then others are like you need to be spending $5,000 $10,000 a month. Uh and I sent you that link with uh the CEO of Nvidia who was saying actually you need to be spending at least so let's connect this back to the hype just for a moment because it's >> right right yeah >> uh even with experience here it's comical. uh he was saying that if you are not sp if you are not spending at least a4 million a year per engineer on AI he would be deeply concerned. So let's just like I I'll I'll just just let me put a slight caveat on that. It's it's probably more accurate. I mean, at this point, Jensen Wang has been kind of saying some what I would consider fairly unhinged things that seem obviously like objectively divorced from reality, like the GitHub stars thing like that. There were some really questionable things.

[00:13:37]So, who knows? But I will point out one thing, which is >> he did say $500,000 a year engineer. So, it might be more appropriate to call it 50% of your salary. Meaning meaning if you were someone who was getting paid $50,000 a year, he might not expect it to be a quarter million dollar, right? So so say 50% of your salary might be a fairer way to represent this thing. Still possibly way out ahead of where your token usage should be optimally, but I'm just, you know, would that be fair just a fair clarification there?

[00:14:09]>> Uh yes. So the the so good clarification and 50% of salary actually is a a better way to think about it. Uh anyway, >> um >> I'm just going by what he said. Like when I read it, the way I took it was he was framing it as I want to see 50% of the your salary spent in tokens. That's that's kind of what it sounded like to me anyway.

[00:14:28]>> Yes. And the followup uh to that was something that I'm sure people have heard and some people genuinely believe um which is the reason for that is because spending it spending that much will make you 10 times as productive. So this connects to something else that I want to talk about which is people hear from serious sources completely different claims about productivity. So there I would I would call uh Jensen's a hype claim but there are people who make that claim not not trying to hype uh Nvidia stock. Right.

[00:15:01]So he was saying the reason you need to be spending 50% of your salary is because doing that will make you 10 times as productive.

[00:15:08]>> Right. Right. And that's not that's he's not the only one saying that even though he's he was in a hype moment there I think. Um but other people are saying uh claiming objectively uh sorry uh it is objectively true that people are claiming that they are 10 times as productive. Whether or not they are is a very complicated story and uh I wanted to connect that to this other thing I linked to you which was this report from the National Bureau of Economic Research. uh and they did a survey of uh of uh CEOs of big companies, several thousand CEOs across the world.

[00:15:39]>> Can we pause on that for a second because I wanted to add something there.

[00:15:42]>> Yep.

[00:15:43]>> Um so the question that I have immediately when people throw around various terms like that and and to be fair again to Jensen, I don't know if he ever said anything about the productivity gains, but I can back infer them and I'll say why. So in my mind, if you put a value on someone's work that is, you know, x, you know, th this person is $500,000 worth of salary to me, right? And by the way, that person is probably, if they're working at Nvidia, getting very generous compensation in terms of stock as well because their stock has performed very well and they probably had options. So half a million dollars for this hypothetical engineer that uh that Jensen is talking about. I don't know if that's all entire compensation or just salary, but so that could be even more.

[00:16:32]>> If you're saying that you think that they only need to be spending $250,000 worth of tokens, that means one of two things. Either you don't think the AI has the possibility to multiply their productivity to 10x because if you did, you would say $5 million would be right.

[00:16:52]Because if you thought if you thought you were going to get 10x productivity out of the person, then you would be willing to pay up to 10x their salary presumably or maybe maybe we'd say 9x their salary because to you that's the same cost for the for the work, right? I mean this is what I'm saying making sense like like >> uh yes with one small detail uh which is um and this is what I so the detail is he said that was the absolute bare minimum and that and that amount being below that amount so let's just be very clear for the audience how extreme this claim is okay that if you are spending less than 50% of your salary on tokens if you're spending say 40% of your salary on tokens right if you are someone who's spending only 40% % of your salary on tokens, he would be deeply concerned. Okay, so that is not not like it could be a little bit better. Hey, you're do you're trying but you know let uh but let's improve in in Q2. Uh deeply concerned, right? Like if I'm if someone tells me they're deeply concerned about what an employee is doing, I assume that means they're considering firing them. Right.

[00:17:58]>> Right. Right. Right. Right. Yeah.

[00:17:59]>> So it's like he >> That's a good point. That's a very good point. And so maybe I can I can sharpen my question to so if if we take all of that into consideration, I might argue I would therefore like to hear what the maximum is.

[00:18:14]>> Yes.

[00:18:15]>> Because if if you actually believe 10x then your maximum should be 10x the person's like well well 9x the person's total compensation, right? Because they're they have their compensation and then you add 9x to it. That's how much.

[00:18:27]So JZ Wang should basically be saying you could should spend maybe around, you know, uh $4.5 million or something like this for a $500,000 employee in tokens if they wish to use them. Now obviously if they can get their 10x from less, then great, right? Uh, and why he doesn't think like I mean because he's trying to sell graphics cards I guess, but why he doesn't think that maybe there's people who could get this 10x speed improve from only 50,000 tokens. I don't know. Uh, I'm not sure. He didn't say. But so I'd like to hear the max because then that's putting that's putting your money where your mouth is.

[00:19:04]If you say we believe we're going to get a 10x uh productivity boost and so therefore all of our $100,000 employees are all going to get uh $900,000 of tokens allotted to them if they wish.

[00:19:16]Now I start to believe you. Does this make sense what I'm saying? I'm not probably phrasing it that way.

[00:19:19]>> Yes. uh there are some subtleties that uh there there are boring subtleties of how corporations think about ROI on salary that I I'll maybe I'll just uh I'll tell you something compact and we don't have to get into like corporate accounting stuff but uh I can tell you for sure that at say Google if someone is getting a 500k total compensation package the corporation is modeling them as producing at least 2 to3 million a year in revenue. venue.

[00:19:50]Uh, and so the reason for that is on top of the total compensation, usually there's a so-called overhead factor which usually roughly doubles it, right?

[00:19:57]So like from 500k they're thinking of overhead is like all the infrastructure this person is using and they need an office, they need benefits, whatever, right? So the overhead takes the total cost to like a million from 500k and they're looking for like a 2 to 3x return on what they put into the employee. So this gets very boring very quickly. Okay. But there is some subtlety there in that um they're already assuming that the 500k engineer is producing let's say $2 million in value uh per year. So now hypothetically um if you believe the 10x claim then they should be producing $20 million a year in uh in value. And so uh per your uh uh per your uh observation logically it seems like they should be willing to pay 5 to 10 million in token costs to get that extra 20 million uh right at similar similar ROI to the ROI that they currently uh currently consider good. I I know this is boring accounting stuff.

[00:20:54]I'm sorry. But this this is how they >> This is my that's the nature of my question, right? My question was basically like so so I feel like if a if a if someone in a leadership position wanted to make the claim of 10x that should be their policy and if it's not their policy I suppose I would start to question whether they whether they really believed what they were saying to me.

[00:21:13]>> Uh well it was either OpenAI or Anthropic who a couple of weeks ago last week I can't remember that it's a blur uh where they said they had an engineer who spent uh 200K in a day.

[00:21:25]>> Yeah. And that and to me that's logically consistent. Like so if that team is saying look we and and I mean it's anthropic so I'm sure they would.

[00:21:33]Uh if they're saying you can 10x your productivity that's what I want to see.

[00:21:36]Like I want to see you spending that kind of money because otherwise I don't know if I really believe you. Sounds like you're just you know it's a little bit smoking mirrors. So I feel like that's a logically consistent position.

[00:21:46]We're you know we're letting our employees spend absurd numbers of tokens because we think they're 10x as productive.

[00:21:51]>> Right.

[00:21:51]>> Right.

[00:21:52]>> Yes.

[00:21:52]>> Yeah. Okay. So sorry to derail you. how you you were going to move on from that point to uh sort of contrast it with the fact that this like for example you you're talking about a a recent news report that Goldman Sachs had found uh that there wasn't really uh much productivity at all uh among places uh sectors that they were observing. Uh I'll let you fill in the rest. I I did read up on this so I have some things to say as well but take it from here. Uh so there was this report from Goldman Sachs and that was cross- linked to uh a much more detailed report from the national national bureau of economic research and uh they surveyed I think uh five or 6 thousand CEOs about uh their internal metrics on uh on productivity and AI adoption and uh their their summary uh conclusion you can go read it yourself it's free online was that in 2025 the productivity impact was basically zero Uh, now you might think, so here's here's the you might you might think, okay, it's basically zero in 2025 because everyone was gearing up to really go really go ham in 2026, >> right? And things are going to be amazing, right? Like we've we've uh we've put in the work and now things are going to be amazing, right?

[00:23:02]>> We lit the fuse and but the rocket hasn't, you know, it hasn't gotten to the rocket engine yet, but it's >> So now the consensus estimate for the improvement in productivity that they anticipate in the next couple of years.

[00:23:13]Uh did you notice that part?

[00:23:15]>> Uh I did not but continue.

[00:23:17]>> Uh between one and 3%.

[00:23:20]>> Oh yes. Okay. Sorry I did not that part.

[00:23:22]Yes.

[00:23:22]>> So the consensus estimate was that they anticipate between 1 and 3% uh overall improvement in productivity thanks to AI in the uh in the next couple of years.

[00:23:34]Um with pockets of uh outsize increase in productivity. uh and one of those pockets was software engineering. In that pocket the consensus estimate was something like 20 to 30% improvement in uh in productivity. So notably uh even in even in the future even in the pockets of high productivity um the consensus assessment was something like 20 20 to 30% improvement not a,000% improvement as uh as you would require to be consistent with um with other kinds of claims.

[00:24:10]Yeah. And this uh this was kind of a difficult so there's there's so many factors here it's really hard to kind of sus them all out because at the end of the day I'm also not sure how interesting it is self-reported productivity gains. Right. And so one of the things that I thought was a little bit more interesting um which was also from Goldman Sachs but uh because I actually had a hard time finding the research note that was being reported on. This is one of like for all of the promise of hypertext, we ended up in a world where everyone wants to keep you on their site. So, no one actually cites anybody else. They're like, Goldman Sachs said this in a research note.

[00:24:48]We're not going to link to it. Right?

[00:24:50]So, so I have to go hunt for it. And of course, you ask the AI, it will just make something up. Uh, so, you know, I had a hard time finding the actual specific research note they were talking about. Um, but Goldman Sachs has put out a number of things on this, including one that was after that news report. So, it's actually a more recent uh thing that Goldman Sachs put out on AI, labor markets, and so on. And one of the things they said about software that I thought was pretty good was they were like, well, um, we would like to see some evidence of software productivity in the market, and there has been none, right? And one of the things that I think is interesting is like rather than self-reported, it'd be nice to see if we actually see any growth in software driven by this because we should see like oh okay like software starts making a lot more money than it was or something like that. Right? Uh if this is to be believed now I'll just sort of say a couple things that have been on my mind about this because I think it's very confusing. I think there are a number of problems with assessing this overall even if you take sort of a holistic approach and just say well we'll just let's look at just like the the outcomes and this is one of the reasons why I'm not sure what will happen with AI in terms of things like software productivity and things like that or in terms of profitability or anything like that would you not agree that there is a plausible scenario where AI does increase productivity by some factor even a large factor like let's maybe not say 10x because that's that's maybe a very optimistic estimate of what could be reasonably achieved in the short term but let's say it's a meaningful percentage let's say it is 30% right that's a significant productivity increase I don't think anyone would be unhappy if their whole organization started being 30% more productive right so we would cons we would anybody any fair observer of AI would have to agree that this was a beneficial technology at that point at least from the standpoint of productivity if nothing else.

[00:26:57]>> Completely agree and consistent with um Fred Brooks wrote about this back in the day in No Silver Bullet I think where um like a 30% a durable and broad-based 30% improvement in programming productivity would be the biggest productivity boost since maybe high level languages. Yeah, like a since it's a compiler or something. Yeah, right. Yeah.

[00:27:20]>> Um, >> so so let's suppose it does achieve a 30% and we should all be pretty happy about that.

[00:27:29]>> I would add sort of the sort of I would add this sort of catch 22 for AI. I just wanted to hear your thoughts about it before we move on. And that is if what that actually just translates to is all of the same companies right now who have the market share that they have just all have to do 30% more work on their software but nothing changes right like because you know like when I think about Microsoft Word it's like well if I make progress on Microsoft Word 30% faster does that actually translate into any more money for me Microsoft as a company it might not it might mean I have to do that work otherwise a competitor who also can get this 30% increase because they also have access to the same AI tools that we do they will make a word competitor and take our business so now we're just forced to do 30% more work using the AI to do it but we don't actually but there is no visible from the outside actual gain other than the software got better right which is a very nebulous thing so would you agree or disagree that there is a possible outcome here where it doesn't look like AI worked but it did.

[00:28:41]>> Yes. Uh so I was trying to avoid getting into this >> economics detail but maybe it's worth bringing it bring it up a little bit.

[00:28:49]>> So economist No, it's it necessary to talk talk about this because this this connects to um there have been a couple of articles I didn't send you these because I didn't think they were very important. There have been a couple of articles on um like AI AI is working but all it's doing is making your job worse right um >> which could be >> it could be could be uh but so the reason uh the the underlying economic idea uh that matters here is that actually productivity is a multi-dimensional thing so there are at least two important uh at least two important metrics leaving aside whether or not you can actually get the data to measure them there are at least in principle two important metrics one is uh how much work gets done per person per time, >> right?

[00:29:32]>> How much work gets done per unit of salary. Uh how much you generate in sales per person per time. How much you generate in sales per salary, right? Uh so all of those are different things and they're tricky to measure. Uh whether or not they're measured precisely, I I I'm not qualified to know, but uh there are already four things that we're thinking about there. And we're not even we haven't even talked about, hey, did the product get any better? Right. We're only talking about >> Right. Right. Right. Right. Right. Yeah.

[00:30:01]>> Um we're only talking about the the most lowest level of I made a thing and I sold it and I got some dollars. Right.

[00:30:09]We're not even talking about, you know, Microsoft Word is better or the game is better or my financial analysis is better, whatever. Right.

[00:30:16]>> Yeah.

[00:30:18]>> So, um it's entirely possible. So, one of many possibilities is that Microsoft Word stays the same, sales stay the same, and because of AI, Microsoft can fire 30% of programmers because it's because it's more uh more productive, right? That's one of many possibilities where we get to this future of a ballpark 30% uh productivity improvement, >> right? that we actually see in a bottom line number because now it's like, oh, we had we only have 70% of the staff we used to have, but we're still making the same amount of money, >> right? So, the product doesn't get any better. The user experience doesn't get any better. Uh we just remove some people from the workforce uh or at least that part of the workforce. And that that would count as a 30% AI uh uh productivity boost. And guys, like I I know that there's a difference between uh reducing by 30% and and increasing by 30%. Okay? I'm just speaking um uh speaking informally. Um the um so that that is one of many possibilities.

[00:31:22]Another possibility is that uh the programmers keep their current salaries but AI drives them to work harder and this is what some people are reporting uh or maybe drives them to work more unpleasantly. Okay, and I'll come back to this in a moment because the unpleasantness is drives them to work more unpleasantly. Uh now some marginal features or marginal products get shipped and that generates some more sales. Um now that may or may not appear as a productivity boost because maybe you're doing 30% more work and you're generating 30% more sales. Uh but like the sales per work has not changed.

[00:32:03]>> Right.

[00:32:03]>> Okay. Right. Right.

[00:32:04]>> Uh so what do you want to call that?

[00:32:06]Right. It's it not really clear. It may just be um the the emerging trend emerging terminology I've seen for this which is not mine but you'll see it in in other discussions is uh work intensification. So AI will just get you working uh harder doing more code reviews generating more PRs. Um like you're not going to lose your job. I'm saying this is this is the view of the work intensification theory. Um so like actually everything no one's going anywhere. everyone's staying and everyone's going to be working harder uh grinding against AIdriven PRs and code reviews and we're going to ship more features and that's that's one of the one of the possibilities that that is in play and uh certainly I cannot uh I can't reject that as maybe where we end up >> well and I just to be very clear on what I'm talking about so let's like consider the following hypothetical suppose that we uh are in a situ situation. Uh let's take a market like Uber and Lyft, let's say, and we look at the market for ride sharing like this is just a like you know ride sharing apps or whatever you want to call those just that part. Suppose we were at a point where there was no more money to be made with ride sharing like it's reached saturation. So pretty much we know how much people can spend on this every year barring any kind of other like external factors that occur. And so the only question is are they going to click on the Uber button or are they going to click on the lift button right in that scenario? That's kind of what I'm mentally imagining like in a market where we might that could end up in a situation like that. Now it's like, well, we look at that market and we see the total like gross revenue for ride sharing across all players is X dollars, X billions of dollars, and we look next year and it's still roughly that like it went up by inflation or something, right? And it just or per capita it's roughly the same. Um, and we just look across several years and it just doesn't change.

[00:34:15]In my mind, what I'm kind of saying is like I could imagine a scenario where AI actually did increase the productivity of Uber and Lyft's programmers, but it doesn't matter because they're just fighting with each other. So no amount of like making their app any better.

[00:34:29]There's no more money to access, right?

[00:34:31]And across a lot of these software trenches, it kind of feels a little bit that way. Like it doesn't feel like there's a tremendous amount of growth potential in some of these like completely captured markets where it just feels more like rent. And so to make a crude analogy, it's sort of like the city's landlords all now kind of have to provide a stove in every apartment, whereas perhaps maybe they previously didn't. The stove is providing a bunch of value, but it doesn't matter, right? Like they can't it doesn't you know what I mean? They it's just something now they have to do.

[00:35:05]And having their workers use AI becomes something you have to do, but you don't actually get to charge any more rent for it because it's just standard now. Does this make any sense?

[00:35:14]>> Yes. Uh and maybe let me connect that to um uh to uh this thing that we've talked about not in recorded conversations about >> uh like how big could this AI thing get.

[00:35:27]Uh and one of the reasons I am not a maximalist who thinks is going to take over everything is uh your Uber and Lift example is is perfect actually uh because the way I think of it is that >> that's why I'm belaboring it is because it kind of feels like the market we're in almost. So there's a um there are physical and social constraints on finding new opportunities, right? So even if AI even in the maximalist uh intelligence case that it's the most intelligent thing you could possibly ever have, uh there are just physical and social limits on discovering new opportunities. So, uh, if you had a team of, you know, whatever, like 5,000 Terrence Tile level intellects working for Uber, how much more revenue could they generate within Uber's business? I think probably the most productive thing they could those people could do would be say, we're spinning off doing something entirely unrelated to Uber.

[00:36:16]Yes. Right. And and that's where we're going to make our money, right? So, there there are limits on what can be done in the market, both physical and and social limits. And that's why I I I think that embedded in any of the like AI takeover scenarios, embedded in there is an assumption that there is um an unlimited supply of unboundedly large opportunities that we're not yet tapping. Okay. No, I mean maybe they're right and maybe we haven't gotten there because we don't have super intelligence yet. It's theoretically possible, but I think it's worth calling out that I think all of the takeover scenarios uh are embedding that assumption that there's this huge huge set of opportunities that we're just not touching at all because we're not smart enough or we don't have enough smart people. Maybe we have smart enough people but not enough smart enough people.

[00:37:04]>> Right.

[00:37:04]>> Okay. So with my giant caveat aside that I was just adding there which is you know b b b b b b b b b b b b b b b b b b b b basically just trying to help the pro AAI side say like hey look you know that just because you don't see anything doesn't necessarily mean it didn't happen because depends what you're measuring. Let's get back to the Goldman Sachs report because that's what you're or Goldman Sachs sorry. Um what what ex is it socks or socks? I have no idea how to pronounce it now. Is it Goldman Sachs?

[00:37:25]>> Socks as far as I know.

[00:37:26]>> Socks. Okay. Um so let's get back to that report. So uh this was self-reported though. So it's just basically I mean it's sort of just people making estimates, right? So one of the problems I have with stuff like that is like who cares like how do they like what are they measuring, right? And I'm not sure because it there wasn't a lot of concrete data about that.

[00:37:46]>> Yes. So I I completely agree. Uh I am uh for various reasons skeptical about um these kinds of economic productivity estimates because I I'm skeptical because I have been asked to help in some cases in what's so-called uh this is technical jargon so-called attribution modeling. So uh just to give a very simple example let's say you're running an advertising campaign. This applies anywhere in a business, but let's just say you're running an advertising campaign and you put uh a million dollars into making an amazing website and mobile app, and you put a million dollars into a Super Bowl ad, and you put a million dollars into uh like Google ads, another million dollars on Tik Tok ads, and now someone um sees your Super Bowl ad and then searches on Tik Tok and follows a link from the Tik Tok to your mobile app and through your mobile app gets to your website and gives you some money, right?

[00:38:38]>> Yeah.

[00:38:39]uh which advertising action was responsible for the revenue that you uh and so there are so-called uh uh causal models and statistics where you can um try to statistically sus out which one of these things was contributing how much to that to that end result >> right because it's a very difficult problem because you know even if you give someone a code like hey go to this web address and it's different depending on which ad you saw you don't know if them having seen a prior ad and not taking an action may have been taken action if they hadn't seen it. So there's more comp that matters. If you talk to anybody in the advertising in the quantitative advertising business, it is known that that matters. These are called uh multiple exposure effects or priming effects. So um so not only is it possible, but in fact it is known that that matters. And the thing that we don't know is how much, right? And so having been you uh invited to help build AI models to do this this kind of attribution modeling, >> it's a lot of like finger in the wind.

[00:39:41]Uh and and like the executives have some intuition of yeah, I don't think our buyers really care that much about the Tik Tok ad, but uh so like even if they happen to see the Tik Tok ad, maybe that's because they were, you know, just like scrolling rap videos at some point, right? And Right. Right. Right. Um so having been involved I can tell you I I personally uh don't take those very seriously at all. They are I would say we are lucky if they are directionally correct.

[00:40:09]>> Um >> so with that in mind I I do actually uh often try to think about concrete ways that we could try to measure um at least very locally the the productivity impact of AI. And here's here's a case where I think um this unevenness of experience comes back which is there are cases uh there are specific tasks that you can definitely do at least 10 times as quickly using AI.

[00:40:40]Uh now let me give you a concrete example. Uh a concrete example is I'll give you an example in in programming and example in marketing because I just want to make it clear that this is not uh not >> not specific to one thing. Yes. Uh so in programming um you have probably at some point had to write uh or at least someone you know has had to write uh bindings for some library in some other language. So it's written in C but you want to use it in Python or use written whatever right written in Java and you want to use it in I don't know who uses Java bin any closure right >> um >> sure >> uh so uh what that means is that you have to look through this big long list of things and then do something quite mechanical uh to uh to source code right so you have to read through source code say what's the signature uh what module does this belong in whatever right uh >> and it's a known mechanical process like you could write a binary like you could write a thing that actually does the parsing and outputs the binding. So, it's it's it's not there's no ambiguity is is probably the right way to say it, >> right? Exactly.

[00:41:39]>> Um the uh the current high-end models do that more or less successfully right now. Like you can just go hand it uh just go hand it a library and say generate generate me a set of uh Python bindings or Lua bindings or or whatever.

[00:41:54]Uh and it will just do that. something that you might have to do uh might take you a couple of hours of tedious work and it just happens in 30 seconds.

[00:42:02]Right? So that is definite that is a real task that has some value and is at least 10 times as productive. Okay? So that is an example. Uh in uh in marketing um a similar level of task might be that you have a um uh you have like a customized sales page for each client. So like you're trying to sell to Oracle and you're trying to sell to IBM and you're trying to sell whatever and then you have a customized sales page for each one to set like you have a bunch of leads and you say can you go take a look at this website and uh you know contact our sales team and each you know you make a customized experience for each one or a customized pitch deck or whatever right the uh the the jargon is marketing collateral right you want to make customized marketing collateral uh and AI is quite good at doing the customization uh given a uh given a base so I make um I make the base uh presentation or sales page or or whatever and then I say now customize this for Oracle, customize this for IBM, customize this for whatever, right? Uh and these are these are low stakes uh kind of activities. this something that you might just give to a junior marketing associate >> because you're going to review it anyway and you know it's just it's basically just saying look take what you know of Oracle and it's sort of a glorified mail merge instead of just replacing the name IBM with Oracle try to actually have the rest of the verbiage and the order of things and whatever line up with what you think they care about as opposed to what IBM cared about or whoever else >> Yeah. And like bring in their logos so that they really feel like you know that you care about them right you really love them. This isn't just about money.

[00:43:39]>> Oh yeah. Yeah.

[00:43:40]>> Right. So the um so that's something that you know I know people doing in uh in marketing and that's something where again that's something that could have taken a few hours and now happens in in a few minutes. Right? So that is object that is a real thing that has at least some value. I'm not saying it's a huge amount of value. It has at least some value that happens 10 time at least 10 times as quickly. Okay. Um and the I I think my feeling about productivity with AI at least for the next couple of years will be uh how many such opportunities uh a business can stack together to get an emergent uh total benefit of maybe 10 to 20%. I think if you are u so I mean I work with many different businesses I have reference points across uh across industries. Uh I think right now if you try really hard and you know what you're doing uh and you're actively looking for these kinds of opportunities not so not doing the naive thing of like look at what my programmers are doing now just tell them do that same thing but use AI that I have found does not this my personal experience but whatever take it for whatever whatever it's worth that does not seem to yield any productivity improvements but if you teach people here are the ways to find things that can be carved off and done productively and reliably with AI I I think you can maybe if you do if you do it well and you're serious, you can get to maybe 20% improvement, but but that takes a lot of like stacking together these kinds of tasks so that you can do so that like you spent less time doing that and more time doing like engine optimization or something like that, right? Um >> so um here here's a question for you just based on what you're saying there.

[00:45:23]Um, does that kind of imply that you sort of that you maybe think the approach, the current approach of, well, I just need everybody in my company to be using a lot of tokens and that's how we're going to make this AI be productive for us. that maybe that is not the smartest approach and that a smarter approach might be you know let's have let's try to identify some AI experts like whether that's new hires like we're hiring people who do a lot of AI work whatever let's try and have specific people who are working pretty hard with AI in our company and what they're going to do is they're going to go around and try to identify and work with people in the various departments to figure out which things look like the ones you just talked about meaning which are the things that they are doing day-to-day that actually do work really well with AI and we don't necessarily want them to find it because you know if you're not I have a feeling if you're not like pretty well verssed in AI it might be hard for you to guess which and and furthermore hard for you to necessarily set up uh the AI is to do those things correctly to you might you may think that the AI is bad at it but it's really just because you weren't sure about how you know it was going to go. So is there perhaps like a more productive approach here that's about identifying those pockets? Is is that sort of what you're saying? I mean my the the approach I just outlined might not be the best one but but the idea of we're going to try and find the pockets.

[00:46:47]>> Yes. So uh let me tell you a non-programmer profession that I think will not not think I know people who are doing this and I think there's real productivity being unlocked there which is uh UX people specifically people who work on uh workflow and ergonomics. So uh this is generally like on the on the enterprise side and not on the on the consumer UX side but these are people who are trying to figure out for example I have whatever I'm an investment bank I have a thousand financial analysts they're all doing research every day and I I when I say research I don't I don't mean the scare quotes to mean that they're not doing research I mean that what that means is not clear right it's a it's a variety of activities so what do they do they hire UX consultants who go and study what are you people doing?

[00:47:33]Right? And ergonomics people, UX and ergonomics are sort of blending together in the digital world anyway, right? Um >> so they study what your team is doing and they say, "Hey, look, uh in our in our usability and UX study, we find that you're spending 20% of your time doing this thing and actually this thing you hate it and nobody like nobody wants to do it and uh we could just get AI to do it or we could, you know, the these are people who 10 years ago were look were doing like automation consulting, right?

[00:48:02]Maybe we can do uh make maybe we can write this with a you know Python script or something >> right?

[00:48:07]>> Um >> so that I think is real concrete uh a real and concrete path to getting um meaningful measurable productivity improvement. Um there aren't that many people doing it and also that's not a you know that's not a trillion dollar evaluation kind of uh kind of story, right? So understandably, >> yeah, Jensen Huang doesn't want to say that because that that doesn't sound as exciting. It's still I mean it's it's still not nothing though. I mean it if you didn't have to justify many trillion dollar many trillion dollars of valuation, you'd still be very happy about that because it is helpful, right?

[00:48:45]I mean >> so um let me >> I can tie that back. There's actually that you said I I would like to tie it back. So if I may, this also kind of gets back to where we started, which is to say that given that like let's suppose the model where there are some ways to get a 10x boost, but it's not your overall productivity that's going to get a 10x boost. It's certain specific tasks that you're doing and we could 10x them. So, we might as well because I mean it's just hours that you saved and you know I mean either you use those to do something else productive or you just use them to go home and go to sleep. I don't know. Um if that's the case that would also perhaps like to tie it back to the initial opening it's easy to understand why you may hear very different things from very different people. If somebody happens to try AI and the thing that they happen to try first is something that hits one of those nice pockets, that will leave a very good first impression. If instead they happen to try something like the ridiculous thing that happened to me on Twitter where four different AIs produced four different completely erroneous results for a very simple thing that you're asking. Um, then it will leave a very bad taste in your mouth. And it's not because the AI was or wasn't, you know, capable of a 10x productivity. It's that it's a very it's like it's a space with concentrated pockets of productivity and then sort of a void in between or perhaps even negative productivity in between because you're spending money on tokens, you're spending your time interacting the eye and not and getting bad results. So where that your dart landed in that sort of modeled landscape, even when you're factoring out whether you're an optimist or a pessimist, just a a neutral person coming into it, they could have very different things that they'd say afterwards. Does that seem >> Yes. So there's going back to this theme of unevenness of experience. So there's there's the pocket of what you're trying to do.

[00:50:47]There's also uh your familiarity with getting these things to do uh do things well. And this connects to the discussion we had about the anthropic cmp compiler that um just just telling it, hey, make me a C compiler. I know that will produce a good result. I can I just guarantee you off the bat and I can give you having used these things a lot.

[00:51:07]I can tell you practices that get you closer to a good result. But there are things that a naive person would not or not naive just someone an inexperienced person would not would not know to do, right? Uh so there's the unevenness in what you're trying to do. There's the unevenness in knowing how to get these things uh how to get these things to work. Um and so some of that is just a just a learning curve, right? And so that's why I I encourage uh particularly young people um even if you are skeptical even if you don't especially like where the AI market is moving I would encourage you to at least become literate in using these tools so that you know you can at least say okay this particular thing the bindings generator whatever or the like the pitch deck customizer that we can safely use AI and we can get a result that that makes sense right I I encourage people at least to be um familiar enough with the tools that they can spot those kinds of pockets and understand but and uh if if for no other reason so that they can say other things are not in the pocket, right? So they can say okay fine like this is this is a this is a nice pocket activity but uh whatever uh optimizing the rate tracer or whatever is not nice pocket activity, right? And you might say also, I mean, you kind of said this, but just to to double down >> also like you need you may need some skills to get it to do the thing in the pocket. Like you said, like it's like, yeah, maybe it can do the thing in this pocket, but it's not the straightforward English prompt you might think. You kind of have to do these other things or approach it this other way. And so if you don't build up those skills, then you're not even going to get the pockets like except unless it's a a pocket that AI does so well that just the naive English statement will produce the correct result.

[00:52:52]>> Yes. And and let me add one one more thing there which seems like a small thing but it's just a frequent uh source of confusion that I see um which is uh if you're not using the top end of the paid models you are getting a conf misleading representation of what is possible. So like Opus 4.6 extended uh GPT 5.4 uh whatever the highing one is called they like all the naming is uh I mean all of them like Gemini you could just go and click I think Grock calls it expert mode or something. Um if you are not if you are new especially if you are new I strongly encourage trying to do your task with the top level of the paid models because if not you are you're adding additional frustration. I'm not saying this will eliminate frustration.

[00:53:38]you will still be frustrated. You will be less frustrated uh if you try the top-end models. There really is an important um difference and I I encourage uh once you have familiarity there, then you can dial back to the uh to the cheaper uh models for efficiency.

[00:53:53]But if you're just trying to figure out can AI do this at all, it really is important to try the um the most powerful models.

[00:54:01]>> Okay. cuz it in your mind it is likely that there will be things that those will do that the other model that the lower the the lower-end modes won't do.

[00:54:13]And so if you're trying to experiment, you will just get you will be misled in in terms of what the AI could reasonably accomplish.

[00:54:20]>> Yes. And so just to reiterate, I'm not guaranteeing that it will work. I'm saying that you are I'm saying that you're making your life uh if you don't already understand what they can do, you are adding confusion by not using the top end. So in other words, what you're suggesting is in when you're exploring, you kind of have to use the top end because that's the only way you'll really know what the landscape looks like. Once you kind of know that that that you if you can get the AI at its most expensive version to do a something in a particular pocket, then you could try to see if you can get it to do it at a less expensive pocket just purely to save cost. But you don't want to base your decision about whether it can or can't do it based on the lower one because it's just not representative, I guess.

[00:55:04]>> Yes. Exactly. Exactly. Uh so many people most people uh start with you know the higher end and then figure out oh okay I can use instead of using opus I can use sonnet for this or whatever the the details don't matter um but if you're if you're trying to understand can this help me at all it is important to use um whatever the the current uh current higherend uh model is uh is available.

[00:55:29]So, um I guess did you have anything else you want to say on that? So, because I kind of have another thing I would throw in uh to the mix here, but you may have had something you didn't get a chance to get. Okay, >> I think that's good.

[00:55:39]>> So, um I think those are all pretty legitimate sort of like, okay, these are these are objective aspects of why people might, you know, why you might hear from someone who says this made me 10 times more productive and someone else goes like this didn't work at all. Um those are all pretty objective reasons why that might occur that we talked about.

[00:56:02]>> Um let's talk about a subjective reason or subjective reasons why this might occur. So one thing I was thinking was there have been many cases in the past where people believe that they are being more productive when they are not being.

[00:56:21]Um, and arguably there's probably even situations where the inverse is true, meaning people who don't think something helped them be more productive, but actually it was they did, right? for um and so I guess to what degree do you think some of what we're seeing is just people being we're just seeing the difference in people in terms of what they believe is happening but it actually bears no you know someone thinks they're not more productive because that's just what they thought and maybe they were more productive or someone thinks they're being way more productive and they're not they just think that how much of it is that >> a substantial amount If I So if I had to guess, >> yeah, >> ballpark like 25%.

[00:57:13]>> Okay, >> maybe. So there's there let me add one more psychological element there that maybe you you consider implied but I want to make explicit which is that there is a um a skinner box or loot box element to uh trying to do work with AI which which is like you go and type in the thing and you have this thing you're hoping for in your mind right like I want whatever um and you click go and then you wait a few minutes and then you get something and then like was it good was it not now I have to review it uh And you know, oh no, it wasn't. It It was close, but uh you know, we're going to get it this time. Right. And then like update the prompt. Type type again.

[00:57:50]Wait. Oh. H. Ah. So close. But now like now there's a monkey in the background.

[00:57:54]I didn't want a monkey in the background.

[00:57:56]>> Right. Right.

[00:57:56]>> Right.

[00:57:58]>> Yeah. Right. Right.

[00:57:59]>> Right. Uh so >> it's kind of like did I get the rare Pokemon card before in this pack of Pokémon cards? No. I'll open another one. Right.

[00:58:06]>> Exactly. Right.

[00:58:06]>> And so there's this feeling I feel it feel it myself. Like this is not judgment of anybody. This is just how the animal brain works. Right. that sometimes you have an idea for something um and you go in and you type it in uh and you don't even type in anything particularly detailed. Just type in a couple of sentences. You click go. It you know crunches for a few minutes, comes back and says, you know, thought for 120 seconds and it gives you exactly what you want. And that happens maybe to me like at most once a week. But when it happens, the the animal brain is really excited, right? Like I barely did anything. I got this cool thing. It it it makes you it also uh creates dramatic tension um where because you're waiting for it, right? It's not like you type it in, you get a compiler error. It's uh you type it in and it's you know it's telling you oh like now I'm I don't know if if you've looked at the the thinking traces, but it's telling you how it's working on the thing, right? Like oh now I'm going to do background research on pictures without monkeys. Right. And >> right right right >> right right. So it's it's telling you it's telling you this tiny little story as you're waiting and then sometimes sometimes it just works on the first go.

[00:59:13]Right. And in that moment I mean even knowing that this is happening to me.

[00:59:17]Okay. Even knowing this is happening to me. In that moment the monkey brain is excited. Right? Wow. I I barely did anything. I got this cool little story and look the result is uh is close to perfect. Right? That happens maybe once a week. maybe more like once a month to me. Uh a much more much more realistic thing is that you're like you keep pulling the slot machine lever uh and try to get the monkey out of the background after generated picture >> and maybe eventually it does. Yeah. And >> maybe eventually it does, right? Um so there's that psychological aspect as well where uh people remember people don't remember that they pulled it 10 times. They remember that they event that it eventually worked. Right. Right.

[00:59:57]Um, so I I do think the subst the psychological factor is substantial and I I would say it's in the range of like 25ish% of the effect. So maybe let me say why why it matters why that number matters.

[01:00:14]I do think that the psychological uh variation in the psychology is maybe as large as the effect the productivity effect that you can get. So it might be that you're not actually any more productive but you feel 25% more productive because of monkey excitement. Uh it might be that >> or lack of monkey I guess in this case >> or or exactly right. You might be 25% more productive but you don't feel it because you hate it. Right. Um and like there there are good reasons to hate it even even when it does good work. there there are good reasons to hate it because uh like it's being imposed on many people in I mean in my opinion very uh hamfisted way and like I know many people who who are just like their work is just uh their work experience is just work now uh because they're being forced to do this stuff even when it doesn't work right um uh and so naturally if you're being forced to do something and it usually doesn't work and then eventually it does work how are you going to feel about it it like even even when it pays off, even in the cases where it pays off, you'll you will still be soured. So I I think that the um I I think that the the amount uh that comes from psychological variation is certainly comparable to the amount of additional bulk productivity you might be able to get if you do a good job. So, I want to emphasize the bulk productivity, not the like sometimes you find a pocket where it's 10x, but your overall life is only improving by whatever a small amount because um because the pockets are isolated.

[01:01:52]>> So, I guess uh there's a lot of other things we could talk about, but I think they're probably best saved to for their own kind of topics. just like they would kind of be departures from the general idea of like why you know I mean I guess to put it in your original terms am I crazy or is everyone else crazy I think is how how you put it.

[01:02:13]>> So are there any other things that you wanted to talk about or any other tangents you wanted to go on that are kind of specifically about that aspect at its core.

[01:02:23]>> Well maybe we can we can review the uh axes of unevenness. Right. So look my I guess my message to to to people is uh everyone on on at all points on the adoption spectrum and the pessimism versus uh optimism spectrum uh everyone at some point feels this am I crazy or is everybody else crazy right and the direction varies depending on your position. So everybody feels this. Um and I I encourage people to think about that as coming mainly from this technology creating an extremely uneven experience for people uh and also extremely uneven um adoption patterns, right? So, it's one thing if you're like if you are um you know an early career, let's say you're you know 25-year-old AI engineer in downtown San Francisco and you're uh you know like you grew up with all the foundational stuff. Uh you came into the workplace right as chat GPT was hitting.

[01:03:22]Uh you're an AI engineer yourself.

[01:03:24]You're riding around in Whimos all the time, right? like you are uh like on the the leading edge of um many trends that are coming together, but you might also be uh you know like a landscaper in you know Virginia and you have no particular connection to any of this stuff, right? Um and so that that's a a source of unevenness as well is just where you are uh in in the economy like how much how much have you been forced to think about this uh as opposed to how much you want to think about this right which is a different um so really I think there there are some uh there there's unevenness in um you know as I mentioned unevenness in knowing how to use these tools uh in in a productive way. So, uh, I I don't blame people at all if they go and try to use it and it's worthless because really it's if you if you haven't taught yourself how to get good results out of it, it's very easy to get bad results. Um, and then unevenness of applicability, right?

[01:04:24]Sometimes like it's really good for the bindings generator. It's really bad for creating the art style for your rendering engine, right?

[01:04:31]>> Right.

[01:04:31]>> Um, it's really good at uh summarizing what was summarizing the headlines from the New York Times today. It's really bad at finding the quote that that you mentioned, right? And there like even just a conventional search engine would be much better, right? It could at least tell you uh a conventional search engine could at least tell you deterministically not a single document in my index contains those words.

[01:04:55]>> Exactly. And also, it really can't produce the same kind of false positive ever. Like it's never going to tell you like, "Yep, it was in this book."

[01:05:04]>> That just doesn't happen.

[01:05:05]>> Yes. because ultimately the answer from the search engine is here's a link to a document, right? And I and I'm claiming to you that this document contains those words.

[01:05:12]>> Um >> I I I would say like that's one of the things um and I maybe I would say this was this would be something we that uh I would like to talk about in a in a in a future one that just gets its own podcast which is the lack of reliability. But yes, >> so continue. Uh no, I I that was really closing it off which is just uh there are many different ways that you can have an uneven experience with AI. Uh it's how much is your employment environment trying to force it on you.

[01:05:42]It's how much do you care about AI to begin with? It's are you in San Francisco next door to an AI company or are you a landscaper in Virginia? Uh are you uh trying to make um you know like uh relatively standard websites for small business clients or are you trying to create the uh art effect pipeline for uh a rendering engine, right? Um are you trying to uh I mean this is this is one that's uh funny in a uh in a uh disappointing ways which is are you trying to argue something before a judge where >> Right. right? Where you know you use it and it creates hallucinated citations and the judge then censor you and maybe you get disbarred, right? That's or are you just asking for a review of um of you know Supreme Court decisions on some topic right where it's much uh uh much more reliable not fully reliable but much more reliable right so uh as a technology u and as it's coming out to into people's lives it's uneven in many many different directions and so this feeling of like am I crazy are you crazy is somebody else like is the whole world crazy I think it's mainly coming from this.

[01:07:01]It's also coming from the hype, right?

[01:07:02]So the like the the hypeers are trying to make you even more crazy, right? And and that's I mean it's it's frustrating that they're doing that. I understand the economic incentives that make them do that.

[01:07:12]>> Well, I think we kind of and we we covered that a lot on the previous episode where like you know it's if there were definitely some attempts to make you believe you could just type make me a C compiler and you actually got a good C compiler out of it, right?

[01:07:24]And that's just not true. And so if that's if and I think I said this literally on that p on the on that podcast if I mostly saw the hype meaning if mostly all I you know if I hadn't looked into it at all and all I saw is like oh we built a C compiler with cloud by just saying build a C compiler and it and it did it then you would go open up cloud code and say make me a C compiler and the thing you got out the other end would be truly horrific and you would be very disappointed right and that's not because there wasn't an accurate report of what will actually happen when you do that. Uh, and how you would need to go about it. It's just because the hype obscures the reality. So, I can definitely see people having that like I thought I was just going to type in this English sentence and out would come the entire thing that I wanted and when that didn't happen, I felt burned, right? Um, and that's pure hype. That's just cuz of hype.

[01:08:14]>> Yeah. uh and what I hope uh hope the audience got out of uh this this discussion was that even when you take away the hype, even if you only look at things that are actually happening, it is so variable that it's natural that people feel like we're not even living in the same world, right? that um um so yeah I I guess that that maybe that that closes it off which is it is uh if you are feeling crazy I promise you everyone like even people who are neck deep in this stuff sometimes feel like we're living in different worlds for various various different reasons uh and it's it you know it's partly where you are situated but it's also just that the the technology itself is so uneven uh that um both in utility and and in reliability that uh it's it's just normal that that uh two people who um might otherwise seem quite similar uh have very different experiences and then they're wondering am I crazy or are you crazy or is there some third thing happening >> and the and the answer is there's a third thing happening basically which is the unevenness is what's accounting for it >> all right well I think that was I think that was enough waiting through AI for this particular topic >> there are a couple things in there that I would like to unpack further and I think maybe I'll put them on my list for like future podcast topics.

[01:09:32]>> Yeah, sounds good.

[01:09:33]>> Well, thanks always Dimmitri for joining us and uh and we will see you next time.

[01:09:38]>> All right, thanks.

[01:09:40]>> Thank you for joining us for this episode of Waiting Through AI. As always, I'd like to thank Demetri Spanos for taking time out of his schedule doing AI research and development to come share the insider perspective with us. And to that end, he is a consultant in that field. If you have business inquiries for him, you can always find him at demetrianos.com.

[01:09:58]If you have questions about this series or want to check out some of the other series that I produce, you can find those on computerenhanced.com and I'd love to see you there. That's it for this week. Until next time, have fun waiting through AI yourself and Dimmitri and I will see you out there on the internet.