The "AI Job Apocalypse" is CANCELLED!

Ajouté : 2026-05-28

[00:00:00]So, when it comes to AI, there were things that we were divided about. Is it going to kill us all? Maybe. We don't know. We're not sure. We're divided. And the text, music, and images it produces.

[00:00:12]Is it a slop or is it beautiful and good? Again, who knows? We're divided.

[00:00:17]But there's one thing that we knew was going to happen beyond a shadow of a doubt. It was going to take everybody's job, right? Job apocalypse. No job for you and all that. I myself, I have to say I was convinced that this was going to happen. How could automation not take jobs? So I, and I assume many of you, I spun up a literal army of AI agents and created workflows and automations to try to reduce just the sheer amount of stuff that I have to do. And it worked. And I would not trade it for anything. But here's the thing. I don't find myself doing less work. I find myself doing more. So, of course, we've heard this message from Silicon Valley going back several years now. AI is coming for knowledge work. Sam Alman warned that whole job categories are about to disappear. Dario Amade said it's going to be a blood bath, a white collar blood bath specifically. Elon Musk said that UBI will be necessary perhaps even not just a universal basic income but universal high income. But just recently and rather seemingly rapidly that message is changing. Both Sam Alman and Daario Amade are softening their message about this job apocalypse about becoming the permanent underclass. And some people are saying well it's because they're going to do their IPO soon. Both SpaceX AAI and Anthropic and Open AI are going to be becoming publicly traded companies and so they're changing their message. So let's dive into it and see exactly what's happening. Why is the story beginning to change? All right, so Sam Alman recently talked at the Bank of Australia conference in Sydney. By the way, I've never been to Australia, but one of the funniest things that I've heard about Australia, is there was this children's cartoon, Peppa the Pig, I think it's out of the UK, and one of the episodes talked about how spiders are harmless. There are friends, you shouldn't be afraid of spiders. It's kind of a kids cartoon, so it kind of teaches them about certain things. That episode was banned in Australia. You can't show it in Australia. Apparently, they had to ban that episode cuz it taught kids not to be afraid of spiders because in Australia, you you should be afraid of spiders. You have to know that they might kill you. And so, that episode of Peppa the Pig cannot be shown on Australian television. I I find that hilarious. Also, I don't think I would ever travel there, but Sam Alman did, and he said the following. I don't think we're going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about. He's saying that in the past he did think that there was going to be a much bigger impact on jobs, specifically white collar jobs, specifically entrylevel white collar jobs and that whole categories will be eliminated. He's saying we're not seeing that and that he's delighted to have been wrong about that. So that's a big shift from what he said earlier, which is most jobs would be replaced, certain categories would be totally totally gone. Now not so much.

[00:03:32]He's also quoted as saying that his scorecard, his prediction scorecard would have been mostly right on the various technological predictions that he did and pretty wrong on the social and economic predictions and implications. Dario Amade, the founder and CEO of Anthropic, used to also have some pretty scary predictions saying that there was going to be a blood bath in the white collar job market. But very recently, again, the tone is shifting.

[00:04:01]He's not describing it as job destroying. He's describing it as a productivity multiplier. So he's saying, quote, if you automate 90% of the job, then everybody does the 10% of the job, but the 10% of the job expands and now that's a new full job. That's a very different story than half of all the jobs disappear or 90% of all the jobs disappear. So this is a little bit more like the Javans paradox but for AI work.

[00:04:31]So if you make a task cheaper more people do it then the amount of that work done just explodes. And so the insight here is that the bottleneck it moves from just production to judgment to taste whatever you want to call that.

[00:04:48]So, here's the thing. Both Fortune magazine and The Decoder, they both frame this timing, this reversal kind of skeptically. So, they're saying that this public narrative of our product that we're trying to create is going to destroy in the entirety of the middle class. That's probably not helpful when you're trying to court investors and regulators and customers. Now, let me know if that makes sense to you or not.

[00:05:17]To me personally, it doesn't. This idea that these founders and CEOs as they got closer and closer to the IPO, they're like, "Oh, we should start saying completely different thing because the PR is better." And I don't mean switching a little thing that they're saying about the product, but rather the impact that AI, this brand new technology, is going to have on the world. That idea doesn't really make sense to me. The truth of the matter is we all kind of suck at predicting the future. When I saw this technology emerge, I thought this is pure automation. It's going to get rid of a lot of the stuff that we're doing and we're going to be free to kick our feet up and fully automate our businesses.

[00:05:59]I'm not seeing that happen. There have been some layoffs in these companies, but there still isn't one clean AI unemployment shock that we've seen. I know there's headlines and there's some layoffs and some of it is blamed on AI, but there's nothing you can point to and definitively say that there's a specific cause and effect. So far, all we have is narratives and headlines. Some of the data published by Anthropic and actually Stanford University did a study in part taking that data and kind of doing their paper on it. It does show that there's some entrylevel work that is getting reduced, that the demand for that is lessened, but this massive job apocalypse or entire categories of work being destroyed, we haven't seen any clean signal on that yet, and maybe we won't. Dan Shipper from Every published a very interesting piece. It's called after automation and I think it gives one of the best explanations for why this apocalypse is cancelled, why it might not be happening. So every is an AI native company. They do a lot with it and it seemingly they do it very well.

[00:07:13]I'm very impressed with what they've been doing. They use Codex and cloud code across the entire operation they're doing across coding, writing, design, customer service, just everything. The alpha test the models out of OpenAI and Google and Anthropic and X and they've automated aggressively yet shipper Dan Shipper saying that there's more work than ever. By the way, I have the exact same feeling. And no matter how much stuff I've automated, the work isn't reduced. There's this meme that went semiviral in the AI community as somebody's kind of biking home on their bike and they're holding their laptop, but they got the thumb kind of propping the laptop open just a little bit. So, it's slightly cracked and not fully closed. And I forgot what the wording was. Something like this is how you know a person is vibe coding or they're using AI agents to to code. Basically, because you don't want to close the thing and shut off your computer. You want some process to keep going, but you don't need to see it. you need to do anything.

[00:08:15]You're just there with your thumb, you know, to make sure it doesn't close and and shut off the computer. And yes, I know there are ways to get around the thing shutting off when you close it.

[00:08:24]But it's a meme. It's funny, but I find myself on Telegram talking to my agents when I am anywhere outside of this room, not just in my kind of work environment, but everywhere standing in line or I'm waiting. If if there's like a 10 to 30 second period of non-activity in front of me, I'm on my phone on Telegram and I'm like, "Build this or add this or or something." As I'm recording this, I have codecs running. In fact, I have a video coming either soon or already posted depending on when this gets released about how to launch multiple agents in their own sort of parallel universes so they can build in parallel without screwing stuff up. There's this kind of new approach that's been working incredibly well for me. The point being is that humans, we work on one thing at a time. With these agents, I can launch five or 10 in parallel, have them all complete work, and then look at that work at the end to see kind of like which direction we want to take it, which things we might want to keep versus expand. I'll be talking about this at length in a different video. But this is the insight, the thing that's been kind of sitting in front of us, the thing that we've we didn't realize until just now. Now, you might have heard this before. Or I don't mean just everybody realized it just now, but it's really more and more people in different jobs across the AI industry. Everybody's kind of switching to this way of thinking about it. Hey, quick aside. I want to tell you something I've been thinking about. There's this pattern that I've noticed inside basically every company that's rolled out AI tools in the last 2 years. You've got maybe three or four people who really get it. They're the ones building custom GPTs, chaining agents together, and writing system prompts on the weekend for fun. And then you have everybody else. The other 95% of the people got an email saying, "We now have AI," and they log into some chat interface. They opened it once, typed at the blinking cursor, typed, "Summarize this email." They got a mediocre answer, and they never opened it again. This is the actual state of AI adoption in 2026. That's exactly the gap Hapax is going after. Their goal is to make AI useful to the whole company, not just the people who already know how to get the maximum leverage out of it.

[00:10:47]Because right now it's wildly uneven.

[00:10:50]The technology is incredible. The diffusion is terrible. And the reason for that is pretty simple. Every AI tool in the market puts the burden on the human. You have to know what to ask. You have to know what model to use. You have to figure out the workflow, the integrations, and the prompt engineering. The value lives with whoever already knows how to extract it.

[00:11:18]Everyone else is locked out, not because they're not smart, but because nobody told them the password. And this is becoming a real business problem fast.

[00:11:27]Imagine you push a product update and next morning your team wakes up to 847 tickets. Customers are frustrated.

[00:11:35]Enterprise accounts need answers.

[00:11:38]Engineering needs to know if this is a bug, a rollout issue or user confusion.

[00:11:43]Leadership wants a brief. And normally this becomes a panic spiral of meetings, dashboards, Slack threads, and someone trying to manually stitch the story together. But this is the pivot. Don't worry. Hapex already found the pattern.

[00:11:58]It categorized the surge, spotted that most of the issues were coming from two core issues, flagged the enterprise accounts, built automated response agents, and drafted the engineering brief. You turned a support flood into a prioritized response system. This is the problem that today's sponsor is trying to solve from a completely different angle. This video is brought to you by Hapex. Here's what makes Hapex different. Instead of prompting the AI, the AI watches how you work and then prompts itself. It observes your workflows. It spots the repetitive stuff that's eating your week. And with one click, it spins up a custom AI co-worker that takes that task off your plate. No prompting, no setup, no engineering.

[00:12:49]They call it the only platform where AI builds AI for you. And that's exactly what they do. The beta numbers are part of what caught my attention. They site incident response workflows going from 4 hours to 20 minutes. New hire onboarding is 60% faster because Hapax automatically captures institutional knowledge as people work. Support volume drops of 40% in 30 days. And this isn't a startup demo running on Vibes. The platform was originally built for banks managing up to 90 billion in assets where compliance and security aren't optional. So the enterprise plumbing is real. They've just opened it up to everyone else. If your team has hit the wall that I described where AI is great for the three power users and basically invisible to everyone else, this is built for the everyone else problem.

[00:13:49]It's AI that works for you and not the other way around. Link is in the description and pinned the comment. Go to askax.ai and check it out. See if it spots something in your workflow you didn't know you needed automated. Big thanks to Hapex for sponsoring. And now let's get back to the video. Here's the big point, the insight. All the stuff that we automate. It's not like we stop. If you work, let's say, 8 hours a day and 90% of what you do gets automated, it's not like you only work that remaining 10%.

[00:14:24]All the stuff that gets automated at the end of it, you still have to interact with it. there's a touch point where you have to look at it and review it and either commit it or discard it. The more expert human work there is all the stuff that I've automated for this channel, a lot of it has to do with research, with monitoring the entire internet for new stories, giving those stories to me so I can kind of see what's interesting. I've created this massive automated funnel that just collects the data, but at the end of the day, I do have to be the one that calls the shots and say, "This is interesting. There's something here.

[00:15:03]Let's unpack it." The AI cannot do that.

[00:15:07]And then when I choose a certain angle or a certain story, I have it research it deeply, give me bulleted point articles, I still have to understand what it's talking about. I can automate the research, the scanning for headlines. I can automate it. It flagging me if something's happening in real time. I can automate massive, massive, massive amounts of stuff, but I cannot automate understanding. There's no button I can push. There's no prompt that would just inject an understanding of a concept into my brain. I still have to sit there and read the information with uh, as they say, eyeball mark one or eyeball v.0. Well, however you want to think about that. my unagugmented vision and brain still has to interact with that idea and understand it. So across all these AI labs and people like myself that are just be using it on a daily basis. Companies like every Dan Shipper all of us we're seeing we're developing massive automations at every they're using sales proposal agents newsletter idea agents customer support agents research memo agents. I have one agent that mainly helps me find the information and does the research for me multiple times a day. I get little notifications of, hey, here's some ideas for things that are happening right now.

[00:16:26]I have one that helps me manage sponsorship. So, it keeps track of the calendar. It kind of looks through the sponsor's briefs to make sure that we're capturing on the point. Helps me draft some ideas, etc., etc. I have a health related agent that kind of over the long run kind of tracks my blood marker from the labs, but on a day-to-day basis just what am I eating, how much sleep did I get, how much exercise did I get. I have my little whoop thing that I've been wearing, which has been very effective.

[00:16:56]It's like like an aura ring. There's a few different of these tracking devices you can wear, but it gives me some insights into my sleep and heart rate and heart rate variability stuff like that. So that gets input into the agent and it's kind of a a coach slash some advice, some fitness and medical advice, etc. And there's a number of other agents I have that are for specific tasks. But the amount of total work is not reduced. There seems to be more and more work. So it's not like it's reducing something. Just the surface of what I do to interact with everything is is different. more and more of my time is spent managing these agents looking at their outputs and as the every article puts it. So there's kind of two modes of working with AI. There's the agent employees that take care of some stuff for you and the other side is the human agent collaboration. So the agent employees that's real automation. They handle stuff for you. And the other side, the human agent collaboration.

[00:17:58]This is kind of the important mode. And this is the mode that I think maybe our mental models maybe were wrong about in that work mode. It's not like I just hand off the work and I just disappear.

[00:18:10]I have to first and foremost frame the work, then I give it to the agent. I I watch what it's doing. Every once in a while, I have to interrupt it and steer it in a new direction. I review the output and then I decide what the next steps are. One thing that I've been doing more and more recently and again a video about that is coming is creating these kind of separate little worlds that have their own sort of almost thinking of it as a scratchpad. Then creating five to 10 different agents and kind of setting them all into their own little worlds with separate databases and separate code, separate folders. and they all just are free to do whatever they want in there to create their own version of whatever project I set. Now, this could be building something from scratch. Or if it's an existing project, then I'll just copy it over everything, the database, the code, whatever it is, into those new sort of worlds and tell each one of those agents to attempt to fix it. Now, usually I'm just telling it to Codeex. Codex spins out five or however many different agents and just sets them to that task. And when I get the five or 10 results back or whatever it is, then I select the best one.

[00:19:18]Everything else either gets dumped, but more often than not, some of these versions, they'll have some good thing that I really like that I want to keep.

[00:19:27]So, I'll say let's keep version number three, but four had this really good idea. And five, well, I really like how the UI is or the color scheme, so let's pull those in to version number three, but we're sticking with this sort of timeline as our main one. Everything else is treated like scratch paper. It gets thrown out. We continue on three.

[00:19:47]And by the way, that three can now get multiplied five to 10 times over and the process starts again. So the 10 agents, they're not working on the same project, but that project gets perfectly cloned and copied. All of them work on the same project, but in parallel on clones of that project. And then at each intersection, I check the work. I combine it as needed. and I continue.

[00:20:11]And this is what Dan Shipper calls the human sandwich. So human sets the framing. AI does the work or it collapses that task so to speak. So instead of me doing it, it does all that for me whatever it can and then the human judges and extends it. Right? So it's kind of like this sandwich approach. And this explains why the human work survives all of this because the bottleneck is no longer how fast you can type or click on things or navigate the various UIs. that portion gets skipped. But the bottleneck now becomes knowing what to ask for, judging the outputs and determining what happens next. So you have to be a good judge of not just what the AI produces, but if it's collecting data, then reacting to that new data in certain ways. And so that kind of gets us to the key mechanism of all of this. It's that cheap competence creates more work. So what that means is that AI makes what used to be human competence of yester year. It makes that cheap. In the past, you might have been very competent at finding whatever data you needed to online. You just knew where to go, how to find it, how to aggregate it in an Excel into an Excel spreadsheet or a SQL database or whatever. You might have been good at creating some little scripts or coding up some software to help you navigate that data. Then you might have been a really good writer kind of describing what that data means based on your findings and your Excel spreadsheet skills. AI collapses all that competence. It makes it cheap. So what used to be kind of scarce competence is made to be very cheap and very available. Writing a pull request is easy now. So is drafting a newsletter or creating a YouTube thumbnail or summarizing a sales call or some piece of software. It used to be scarce. You used to have to have some competence to do that. AI kind of collapses that. So we knew that already. So this isn't headline news. We kind of knew that this was happening. Here's what I think maybe we missed. And again, some of these ideas aren't new. We're just getting a lot more sort of data on them. And we're realizing that no, this is in fact the thing that's going to push the world this way. That's this idea of the Jebans paradox, although it's usually used to talk about certain valuable resources or expensive resources that when they collapse in price, you know, the demand for them goes through the roof. So let's say you use your car to get to work, get to the grocery store, but gas is very expensive. Then you find some new way to make cars use less of it. So you think, okay, now people are going to use less gas because the cars are more efficient.

[00:22:56]They can drive instead of 5 miles on one gallon, 30 m or 50 mi, whatever, whatever the case is. But what happens is people decide to drive everywhere.

[00:23:04]They take a road trip. They just go joy driving, whatever, because it's cheaper, it's easier, it's more efficient, demand goes up, consumption goes up. So the basic idea is when something becomes cheaper, people use more of it. With AI, competence got a lot cheaper. Competence for specific things like the output of the things that the AI model was trained on making code and pros and product specs and documentation support tickets whatever. So it got cheaper, people use more of it. So output explodes. So operations people can now write code.

[00:23:40]Engineers can write product pages. So the first order effect isn't everyone stops working. The first order effect is everyone's productivity goes up.

[00:23:50]Everybody produces more stuff. Now, the big problem here, of course, is the slopification of everything, right?

[00:23:58]Because some of these outputs, they're the same. So, if everybody has access to the same outputs, everything becomes kind of AI slop. And this is the defining factor that I think is going to be so important in a new era is that experts become more important. Yes, anyone can create YouTube thumbnails, but somebody with a very sharp design eye and understanding how they work can create better ones. Hopefully, anyone can produce code and create software, but you really need that engineering mind to understand how to build something properly, how to test it to make sure that it's quality. Anyone can draft a newsletter or a book or a blog post, but you really need great writers and editors to make it into something good, something unique. So, what this means is that automation isn't magic dust that you just sprinkle over work to make it disappear. It just becomes the operating system on which you need still maintenance, evals, right? You need to figure out what's good, what isn't. You need the right sets of permissions. What can the AI automate? What it can't? What is it allowed to do? What it isn't? You need the review cues to make sure that the process keeps going. You need instruction files. A lot of my time, not a lot, but a significant amount of time is spent writing very in-depth, very involved instruction files. And at the end of the day, you do still need human ownership. You need somebody to say, "Hey, like I made sure this is good.

[00:25:30]This is unique. I put my own flavor on it that it's not just slop. The point is AI can automate the middle of the task.

[00:25:38]We assume is we still decide what tasks matter, what outcomes we're trying to achieve. And at the end, we need to decide which one of the versions is better, if it's exactly the outcome that we were looking for, and then repeat that process again. So now that we're here, what comes next? Humans still control the beginning and the end. So that means that the work that we do, the jobs, they they change. They don't go away, but they they evolve. The amount of money you're going to be getting paid in the past might have been for the stuff that you you did, for the lines that you typed, but moving forward, companies will be paying humans for the decisions that they make, for them supervising what gets done. They need to integrate it into the company. they need to differentiate it from all the other stuff that's produced and at the end of the day they need to take ownership for that stuff. So this means that this idea of a permanent underclass might not exist. This is not what we're heading towards at least when it comes to jobs.

[00:26:44]We may yet still have a permanent underclass but not for people in and for workers but rather for companies. That's because AI doesn't help every business equally. There are a lot of companies that their moat, how they make their money, is based on this idea that used to be difficult to produce certain things. The competence was very expensive. We've already seen Anthropic publish a few things that were open- source that crashed the stock valuations of certain companies. So if you had some specific set of code that dealt with some legal aspect for example because it was hard to develop that you needed somebody that was competent in writing code. You need somebody that was competent in the legal profession. They could charge a lot of money per seat for certain companies like legal companies to buy that software then charge and per se to have access to that software. With AI that moat is basically gone overnight. But companies that have the right data workflow, these new people that are able to use AI, they might win bigger and bigger shares of the market.

[00:27:50]And in fact, AI powered companies might expand and wipe out other companies.

[00:27:55]This is still kind of unclear. We don't know exactly where this is going to be going. I don't tend to know how this is going to play out, but that might be kind of the next thing to watch. How disproportionate will AI be for companies? So, I think the takeaway at the end of the day is that so far it's looking like maybe the whole jobs going away thing was overstated. Again, we're not saying that this is the new future.

[00:28:19]We know it for sure. We're saying that the hypothesis changed. So, over time, we'll see more and more data that either proves or disproves this. But for the time being, it might be that the jobs are evolving but not going away. And in terms of how AI will affect the companies, the big question here is is it that AI will raise every company's productivity, you know, let's say 1% or 5% or is it that it's going to take the top 5% of companies, raise their productivity 20%, but not necessarily everyone else. The difference between those two scenarios will make a drastic impact on how business is done, on what the stock market looks like. But the big point here is I think for a lot of us our intuitions about what AI is going to do were off in terms of specifically how it's going to affect jobs. Now I think the best advice that I or maybe anyone can give people that are worried about their jobs is this. Use all the new models as they come out and understand where you fit in into that work process.

[00:29:20]Think of yourself as managing the inputs and the outputs. So on the input side, what are you prompted with? What context do you give it? Like what model do you use? Everything that you can decide on the input side to improve how well the model works and then on the output decide which of the different versions you're going to use. If the job it completed, was it good? Was it not good?

[00:29:43]The people that figure out how to do those sides of the equation, I think they're going to do very, very well. But just like we all had to learn to type and interact with computers, I think doing this and working with AI agents will just be table stakes. Also, it does seem like different job titles will maybe either go away or at least become a little bit more diffused, a little bit more vague because again the competence is going away. Like you might have been really good at this particular thing and that was your job title. Now, more and more we're going to be doing a lot of different things across job titles and be more like project managers. Maybe that's a better term for it. So, I'll leave you with this. I think you're going to be seeing a lot of headlines saying that AI companies did a full 180 and kind of flipped on what they were saying before. And some of those publications will probably try to paint this as a preipo change of, you know, marketing and PR, etc. I don't think that's a fair claim. And that's not a fair framing because a lot of us that use this technology on a daily basis, we're all noticing a similar shift. Each new iteration of the model is allowing us to automate more. It does more for us. We're more and more amazed at the stuff that it's able to do. It gives us more productivity, more output, but it doesn't reduce the amount of time that we have to work. If this trend continues, and this is important, we have to realize that it kind of negates the two biggest AI risks that we've ever talked about. First of all, if we have to manage it and make sure the outputs are right, then that whole thing about it killing everybody, well, that's probably not going to happen. Not without a human being sitting there and going like, "Okay, let's make a plan for how to kill everybody." And then on the other side going, "Well, I don't know if that's going to work. Let's get some better ideas. Let's uh let's see if we can improve the process. Which by the way, that idea of humans using AI to do something nefarious, that can always happen. Any technology can be abused.

[00:31:47]But the specific X risk from rogue AI as it in developing self-awareness and then going all Skynet Terminator 2 style. I always say Terminator 2. I've never seen Terminator 1. I I just saw the second one. To me, that's the Terminator movie.

[00:32:03]I I don't even know what happens in the first one. But the point is, if this pattern holds, then those two big risks go away. There's no X-risk from rogue AI specifically, and there's no AI job apocalypse. There's a massive explosion in work productivity. How we do work changes. The internet changes. Tons of things still get disrupted, but the two biggest fears might be a thing of the past. Let me know what you think about this. Do you agree with this hypothesis?

[00:32:33]Do you believe what Sam and Daario are saying or do you think it's just marketing for their IPO? Let me know in the comments. If you made this far, thank you so much for watching. My name is Wes Roth. will see you in the next