Transcription
The key point is that all the stuff that's promised for AI, you will replace your workers with a computer. It'll be awesome. It'll be so much cheaper and better. That was a lie. It was always a lie.
The more central point is this was always a get-rich-quick scheme. Same sort of form. It was a promise of magic. Words like AI, vibe code, and so on. They are now in the general public a marker of low quality. And that's because they're low quality. It's the surviving companies that will rehire. The others will have killed themselves with AI.
Hi, I'm Isaac and on the tech report with me today is author and host of Pivot to AI, David Gerard. Thanks for coming on.
>> Good afternoon.
>> The car manufacturer Ford has rehired over 300 senior software engineers after attempts to replace them with AI failed. CEO for Jim Farley said the impact moving to AI was costing hundreds of millions of dollars in warranty and recall costs. How widespread do you think situations like this are where AI is costing money?
>> So, you're going to see this all over. The key point is that all the stuff that's promised for AI, you will replace your workers with a computer. It'll be awesome. It'll be so much cheaper and better. That was a lie. It was always a lie. That was the actual promise the whole time along since 2022. And it's false. It was never true at any point. And now the chickens are coming home to roost. And it's it was always false. It's always expensive because the secret of generative AI is not just machine learning in general. That can work sometimes, sometimes it can't. But specifically the generative stuff, the chatbots and so on. They don't actually work to replace jobs that need to be done. The use case for AI is work that should not be done like filler and spam and things that you do because you're supposed to and not because they actually serve a purpose. Anything that's load-bearing, AI keeps not quite being up to the job. The closest we've seen to a use case is coding. And even then, I get I get a ton of people tell me about this stuff. So, I have a biased selection that's people who already hate it, but they hate it because they're using it and they're forced to use it. Like, they're forced to use Copilot at work even though it manifestly doesn't actually do the job and they have to run around cleaning up after the robot. The code is bad. It's not humanly tractable. Um, you can't have the robot spit out code and then a human review it because human attention wears out really, really quickly. Um, if you posit that the human is a completely robust machine component made of steel, that trick has never worked ever. You aren't going to somehow retrieve the robot from this situation. The code doesn't work. It's buggy. People now look to AI as a watchword for low quality. It's like I'm actually surprised I haven't found any actual smoking gun evidence that Windows 11, for example, is produced with AI coding. Although I'm sure there's a bit in there, but I don't actually know the smoking gun evidence. On the other hand, it's the most vibe-coded feeling product anyone's used. It's the type example for people. Oh, it's vibe code. You know, words like AI, vibe code, and so on, they are now in the general public a marker of low quality. And that's because they're low quality. And you find examples like this where people discover it's good for doing a rough mockup, but then some people press rough mockups into production maybe, and they don't work. Even then, it's not that good for a mockup because it misleads you. The mockup is 90% of the way to a finished product instead of 1%. You know, a mockup is not the product with just if you just add a bit of functionality. The mockup is literally just a bit of graphic design at best. So, this will keep happening. I predicted and I it was a fairly easy prediction that a couple of years ago that within about I said 6 to 18 months I'm thinking it's going to be two or three years now from there 2027, 2028 you're really going to see a lot of people hiring back big time to solve this precise problem because it turns out the human is load-bearing. The problem with coding in particular was never typing. It wasn't that people weren't typing fast enough. It was always all about thinking, architecture, experience of easy-looking traps that you fall into. Like when I worked in IT, this was literally my job. System administrator to developers. Uh, my job was to look at a new tech and go, "That's garbage. You're going to have these problems," enumerate them, and be right about it. But it's so tempting. You remember what they've done is they sold the sizzle so hard and it's such a tempting sizzling sound, but the sausage doesn't actually exist for real work. And also the whole thing was an enterprise SaaS play. Now the thing about enterprise SaaS is your price tag only goes in one direction, upward. You will not um, ever get a price cut. Anthropic will tell you, "Oh, tokens are so much cheaper now. By the way, your bill is up this month." You know, or when they when my GitHub Copilot moved from subscription pricing to token pricing and people's bills went up literally 100 times. Um, any business process that costs 100 times as much is not the same process. It turns out all this stuff is heavily subsidized either with massive corporation cash where Google subsidized with advertising and Microsoft subsidized it with its massive enterprise IT business or its venture capital cash when it's open Anthropic, but subsidies will eventually run out. I I I originally guessed it would be sometime in 2027. I think that is looking about right. That's when all the money in VC and all the money in private credit and all the money that's floating around actually runs out. You know, there's no a shortage of actual real dollars for OpenAI and Anthropic to set on fire to sell you a subscription at a tenth of the cost of actually serving it. It's um, not a stable situation. They were hoping for a breakthrough. They were desperately hoping for a breakthrough. They're hoping at last we found a use case and it's not a very good one. Um, I mean, I'm chatting with friends about starting an AI cleanup consultancy. So, you know, maybe that'll be my next job. You know, because there is a need for AI cleanup. There's people who do this job already. Um, there were whispers of it in the middle of last year and now it's openly much more of what people do and I talk to people who do this. You know, they go into somewhere as a consultant and part of what they do is, "Oh, we want you to get our code into working order," and it turns out to be an AI super fund site cleanup. So, I expect that the trouble is the AI was really mis-sold quite badly to businesses as, "At last, you can get rid of those infuriating humans." And that trick never works ever. It's never worked ever. Um, businesses will spend much more than employees cost on automation because they just dealing with people is a pain in the backside, and that's true, but also we're in a society. So, you know, it's it's a tricky one.
>> To your point about the rehiring, I think IBM, Google, Amazon, Shopify, and a bunch of others I I can't remember off the top of my head. They're all quietly rehiring staff already after AI failed to deliver on on kind of the promises that they were were made. Some are quite open about it. Klarna is extreme was extremely open about it. Yeah, we tried it, failed. We're back to having humans in customer service because like what job is customer service? It's literally the human element.
>> And I was going to say just over the weekend sort of anecdotally as it may be, I was talking to even a designer, one of the the first industries that was kind of hit by sort of the first wave of AI as it were. They were saying that their company was rehiring all their junior designers which they tried to replace and failed. And I kind of want to go back to that AI cleanup thing you were talking about and and on top of this sort of rehiring as well. Like is this something that you think you said it would be a year to 18 months from now, but is this something you think we might be starting seeing sooner?
>> I said it was a year to 18 months about a year ago. So Oh, of course. So, we're seeing we're seeing it kick in now. Um, and we'll see it kick in much bigger next year because the bottom line is this stuff only sort of works until you press on it. You can't press a cardboard mockup into production and have it not fail. You'd think this was obvious. The the more b the more central point is this was always a get-rich-quick scheme. Same sort of form. It was a promise of magic. Like get-rich-quick schemes or crypto or ostrich farms or whatever, they're always a promise that you have discovered one weird trick that will get you a win without effort. And that promise is always a lie. Anyone who says money magic exists is out to pick your pocket. Particularly if it's about money, they're always out to pick your pocket. If you hear a promise of magic, look for where they're trying to pick your pocket. You'd think that was obvious, you know, but somehow it isn't because the promise is so so seductive. And chatbots are really convincing. They're an excellent demo, right? I can see exactly why people went, "The chatbot is the most amazing thing ever," because we've had science fiction dreams of robots you can just talk to for millennia, and now we've got one. This is not nothing. That's something. That's really impressive. As a demo of machine learning, if you press it into production, it turns out that unfortunately our talking robot is that you can have a fun chat with is a lying hallucination machine that will always be a lying hallucination machine by its design. And that cannot be fixed. Even OpenAI said hallucinations cannot be fixed. You know, it's it's it's one thing when it's coming out of a lab and you say, "Wow, that's really cool," because it is cool. It is genuinely very cool, but making it into a cornerstone of business was completely mis-sold.
>> I was going to say a few years ago, we were presented chatbots as if they were sort of the whole AI is coming for your job thing. And I even remember seeing an advert on the underground in London by I think Artisan which said stop hiring humans.
>> Yeah.
>> And I mean Yeah. Do I was going to ask you do you think AI was mis-sold? But I suppose your answer is yes.
>> Mis-sold hard. That was the whole promise. We will replace your employees and it hasn't come true. And they said we'll do a universal basic income on AI. I mean, we know how to do a welfare state. We've done them before. What you do is you spend some tax money on it and it stimulates the economy and so on. And then we get the quite prosperous mid-20th century, but they um keep wanting to do what they want is say, "Oh, with AI we can do this without having to change anything we're doing." And I don't think that's true. I think that one key reality right now is times are tough. The economy is not technically in recession, but that's because there's a few large companies keeping the numbers up, but all the pre-recession indicators are there. One prediction I made is that when the AI bubble crashes, and it will crash because it's a bubble, it's going to take a lot of stuff with it. Um, my dire outlook is Great Depression 2. Nobody wants to hear that. Um, I've had some pushback on that where people say, "Oh, it'll just be a recession. We'll have institutions. They'll recover." But the institutions are getting white-handed, too. In the US, for example, the new Fed chair, Kevin Walsh, he's a Bitcoiner. He is like into Austrian economics, gold bug economics. He he he'll give you gold bug theory. and he's trying to do the job that's in front of him. I mean, Alan Greenspan believed the same things, but he did an okay job at the time. But, um, these guys are not the people you want in place when the answers to the problem is going to be spray money all over the poor people, not the rich people. Uh, because the poor people are the ones who spend 100% of their income and that adds to the GDP. It's like basically Hayek was wrong, Keynes was right. But that's getting off topic here sort of. But I do think that we're headed for bad economic times because we got them already. Um, I don't think I think we're seeing a lot of pre-recession indicators already. Um, AI is not going to help with that. VCs trying to run a bubble party is not going to help with that. That's what AI is for. My theory of a the AI bubble is it's 100% a VC bubble party. I think they're trying really hard to get quantum in right now. I saw a great thing in Pitchbook. Pitchbook is a site everyone should read. It's the new site for venture capitalists. It's amazing. They had a whole feature about investments in quantum computing. Quantum computing doesn't exist. It literally doesn't exist as a technology. It's a physics lab experiment. But they're doing deals about quantum computing which doesn't exist. And you know it it doesn't have to be attached to reality to do deals. So we'll see how the quantum computing field goes. I actually know people who do stuff in quantum computing like actual physicists and some of them are very promising. I wish them every success but I'll I'll see I'll say it's working when I see it working. I was reading a piece in Forbes that explained the the savings of AI adoption would be essentially all be wiped out by the cost of rehiring for the cut roles that we've been talking about.
>> I don't believe there was actually saving. I mean the saving was that you reduce your wage bill by firing everyone. I mean sure, but also you need someone to do that piece of work to run your business. It's it it's the epitome of pennywise, pound foolish.
>> I I also can't imagine that seeing a role that was cut for AI then being reopened sort of 6 months or a year later would inspire much confidence or trust in those that remained. And I was just wondering how much damage you think has been done to that sort of employer-employee relationship because of this mis-selling of AI.
>> A tremendous amount. Um, everybody has been told, right, you're fireable. I'm replacing with a robot. Whoops, I've changed my mind. Oh, have you now? So, um, I expect I don't think they'll just hire back all the experts as they did before. I think a lot of companies will try to ask it. They'll try to do it cheap as long as they can. Uh, and it may or may not work out and then they'll do the right thing as a last resort if they haven't died by then. Because I'm thinking it's the surviving companies that will rehire. The others will have killed themselves with AI and we're seeing that too. There are companies that just suddenly discover they don't understand their tech stack. They can't afford their Copilot code bill. Uh, when they do it doesn't work anyway and suddenly they're bereft and it's not a pleasant place to find yourself as a business. So, it's going to be hard. I mean, nobody wants all the businesses to fail, but they're being How do you tell people not to fall for the siren song of "You can fire everyone and keep going as a business"? That was always false. I mean, there's a pile of what I can only call boss what what Ed Zitron calls boss erotica where they tell these sexy, sexy stories of how we formed a billion-dollar company with two employees. No, you didn't. If you look at the details, you didn't do anything of the sort. That sort of thing. Great headlines that don't check out when you read the story. I mean, yeah, I I can tell you there's not a bundle of cash in me, Cassandra. I I basically write a newsletter that's going quite well, but you know, I'm not rich.
>> Do you think we would be seeing the amount of investment in AI that we are seeing now if it wasn't for the sort of alarmist and arguably irresponsible marketing that made people believe the bottom was going to fall out of the labor market.
>> I mean, they always wanted the bottom to fall out of the labor market. Um, so it's it's a tricky one. Um, I'm not sure about the investment in AI. It's I don't know, spending money on old rope, if you can call that investment. I suppose it's it's a hard one. Um,
>> the actual AI sort of investment scene, when you're talking about companies buying in AI, they're all sort of trying as a trial balloon and discovering maybe it doesn't work so well. And like as I said, I started hearing about AI cleanups being done very, very stealth-like around mid-last year when it was the CEO's pet project. The senior engineer had to call in someone to clean up the mess. Now, a few months later, it became a lot more public. And now it's just things that consulting people I talk to, they do, you know, "We've got a code cleanup. It's an AI cleanup." That sort of thing because they got some essential service and they filled it with AI and the AI doesn't work. And they'd like a system that works because their business runs on it and it doesn't have weird bugs that keep coming back and so on. Just stuff that doesn't work properly. Every person who ever touches a computer or a phone knows about that, you know, it's not some sort of esoteric technical thing. It's a basic "Does this even work when you're making me use it?" thing. And often it doesn't. One of the big ironies I I see in all of this is that certainly in the in the short to to medium term, one of the only revolutionary sort of macroeconomic impacts we're going to see of AI is is probably the AI bubble bursting.
>> It's going to be bad.
>> What do you what do you think that looks like?
>> I think it looks like so it's hard to mark this out because you know Kodak and Polaroids still exist as brands. You know Atari still exists as a brand. The OpenAI and Anthropic brands will exist forever. So when do you call it dead? I'm going to call it dead when the venture capital subsidy to those two companies stops and suddenly the API prices go up to what they actually cost, which my finger in the air guess is around 10 times. I've spoken to people who actually do local models for companies. Um, they use some local model. Um, one guy said OpenAI's GPT-4 is popular because it's got the OpenAI brand name. That sort of thing. Um, but what they do is they set it up for a company who don't want to send their data to California but or China but they do think they need a chatbot. So the big largest cost I've seen is around $8,750 a month per user to get similar performance to ChatGPT Pro. So that's about 40-odd times the cost completely unsubsidized. Obviously there'll be scale. You can do it for cheaper than $8,000 like if you shop around, but this is just using rented hardware at AWS if they have the hardware in your region and it's all very it's very annoying and tricky. Others set it up for other people. This is a thing that happens. There's a small but real market for local instances running at high speed. This is not the performance you'll get at home on your Nvidia RTX 4000. It's um, it's much more expensive. You basically have to rent the real hardware at AWS because you're not going to buy one of these things and run it on-premises. Um, and it's like about 40 times 20 to 40 times. I predict my guess is about 10 times the cost unsubsidized because that's what it would be if um you can operate at scale like you have a data center full of r full of rapidly decaying Nvidia rigs. They do rapidly decay, by the way. They they're built with their cooling is not the best and they often overheat and many of them die after only 18 months. Expected life 2 to 3 years because by then they're obsolete. You have to buy some more. But um, it's it's a tricky one, but I think that when the subsidy stops, that's when we can call OpenAI Anthropic dead. Um, that will be the end of the bubble. It's like how the financial crisis we mark the end as 2008 usually Lehman Brothers, but of course the whole thing started late 2006 and ended sometime in 2009. But if we want to mark a day, I think we can feel it starting a bit already. Those rumblings are what people like me and Bender and Hannah and Zitron and so on have been saying, "Yep, it's coming. That's it. Yeah, that's the one." We said a few years ago, "Yeah, it's coming." And um, I think we're right because it should be obviously right. And the only reason people don't say it's right is because wishful thinking. But ultimately, gravity works.
>> What goes up must come down.
>> And this is bubble. Bubbles pop or best well actually bubbles deflate, but they do deflate. So if we get back to the Ford rehiring engineers and and other companies rehiring staff as well, that the jobs that these people will be returning to will probably look very different from what they had before, requiring a lot more of that mind-numbing verification of AI slop and that kind of stuff that I think engineers have specifically said is worse than just reviewing junior junior outputs.
>> I think they'll put up for as short a time as they can. I think that they will be recommending, "No, this is garbage, throw it away and start over," because I think we can one thing I saw a fellow on I saw someone say this on Master it was um, the reason why it's hard to review a chatbot code contribution is if a human contribution is wrong, even it's completely misconceived, there's still a theory behind it, you know, like line 10 relates to line one. It follows from it. With a chatbot code change, line 10 is completely unrelated to line one because they're not related to anything. They are literally just a statistical average of what code might look like. They are not related things. That is why reviewing code is so fatiguing. And if you give people a fatiguing mental job, eventually they just stop doing it. They just press the "Looks good to me" button and it goes through and then your system breaks, but you know they've got high AI output and you did tell them to do this job. You ultimately can't blame the humans for the AI churning out spam. You can't do that and have it be sustainable. Um, not if you want your systems to work. I'm wondering if they remember that they're supposed to be in business to make money sometimes. How long do you think it is until we start seeing a problem with motivation if people are doing this kind not just monotonous work but arguably infuriating work of managing AI slop?
>> About a year ago, honestly. I've been getting reports of this since about a year ago. I tell you, everyone I am the complaints department for AI. Um, I get these all the time. Um, it keeps my finger on the pulse. I realize I'm only hearing one side. I but I you might say I'm only hearing one side, but I absolutely hear from all the AI boosters and they say all the same things they always have, also the same things they were saying when they were crypto boosters. Just swap out the buzzword. So you I've heard about this for a long time. You can't review weird AI-generated code that literally does not make sense because it never had sense because it's a statistical model output. There is no brain in there. There's no person in there. There's no mind. There's no theory behind it. It doesn't have an idea of what it's doing because it can't think because it's a statistical chatbot. It does tokens. That's all it does. Um, it's amazing. It does as well as it does. All of this stuff, as I said, as a demo, it's incredible. This stuff is really amazing. As production machinery for business, it doesn't work. You know, it's AI code is terrible. AI applications are full of weird bugs. They try to fix them. I I saw a I have to look into it more closely. Statistical thing over the weekend where um someone went through a pile of open-source um AI-coded apps on GitHub and they found that most of them were just abandoned after a couple of months. People had put them out, talked them up really loudly, and then they just sort of quit because it didn't work and nobody really cared. Um, and this stuff is unreviewable by humans in at scale. So if you try to put serious engineers on this stuff, they'll keep the job because times are tough, but they're absolutely no longer taking you seriously as a company. So if seniors have less time to do that like actual innovative work because they're reviewing AI stuff and the juniors have less opportunity to learn because AI is has actually that's one of the areas that AI has taken jobs is the entry level. Where do you think that sort of technical debt leads to? Are we looking at more higher education or something like industry training programs perhaps so that we actually have people coming into positions rather than there just being this gap of "Oh yeah, well we were going to use AI and then actually we had to not use AI and now we just have a few years gap of nobody knows or has any training."
>> Businesses really, really, really hate training anyone on the job. They only do it if they have no other alternative. They'll demand, "Well, why aren't the universities turning out people who are specialized in this task that existed six months ago, you know?" Um, of course they'll say that because they'd be put leaving money on the table if they didn't, but also it's not a reasonable thing to ask. Ultimately, you're going to have to hire graduates who know nothing and teach them how to do the job. And then you get someone who knows your stuff, knows how to do it, and you can keep them around for a few years being productive. You know, um, it's ultimately businesses are going to have to remember why we ever did it that way. And the answer is because it leads to better business outcomes over a time scale of less than a week of more than a week. Sorry. So if we sort of bring this back around full circle as it were, if AI isn't saving employee costs, it may actually be increasing them and the productivity increase at scale is essentially absent. What exactly is it that AI is offering that's worth buying if you are an executive?
>> I think what they're offering is the sizzle. They're offering the big promise, the marketing promise that you can get employees without having humans. And this was always a lie. It was never true. So no, if you want to have a business that where people do things where you would have people doing things, the chatbot is not going to replace humans in that job. It just isn't. I think we have few years of example proof by now. Uh, people don't like People hate AI slop output. They really hate it. They see AI slop pictures and they think you're not serious. They see AI slop text and they react really badly to it. You know, um, I I will continue to use my M-dashes because I use them correctly. But you know, you can see why people moderated a bit. I saw a great thing, a great essay from a Kenyan who said, "I don't write like ChatGPT. ChatGPT writes like me." Because all these things are trained by educated Nigerians and Kenyans who are very smart. They speak excellent English, but also they're cheap. So they get these guys to fix it. And the AI ends up writing like they were trained to write argumentative essays in English. And then they get people say, "You write like ChatGPT." "No, I don't." You know, and it was a great essay because it was very clearly not. It had all the chatbot tropes, but it was not a chatbot essay. It was clearly written by an angry human with a point. That is, there was a mind. There was a theory behind it. And it was obvious from the writing because it was excellent writing. So, you know, nobody likes slop. Nobody wants to deal with slop. Nobody wants to deal with broken things that keep not being fixed. And dealing with the chatbot and dealing with your call is important to us. We have high call volumes. Everyone knows this is this is rubbish. Ultimately, you can only skimp on actually doing the job so far. And I think we've overshot. And I think it would be a productive aim for businesses who want to make money in business to wind that back a bit and try doing a bit better in terms of their public relations and the product they're actually producing. It's worth it. It's worth a try.
Well, David Gerard, thanks for taking the time.
>> Thank you.
>> If you enjoy today's episode and you want to hear more of the Tech Report, please consider liking and subscribing. Also, you can get episodes of the Tech Report wherever you get your podcasts.