Has AI Changed Work Forever? Not Really... | Cal Newport

All right. So I got to react to something that I've been sent a lot of times. I think a lot of people have been sent this a lot of times. It's an essay that's been going around on X that went viral. It's written by an AI startup entrepreneur named Matt Schumer. And the essay is called Something Big is Happening. And it is a right down the middle, AI is about to change everything for real this time, let's all be worried type of essay. I got sent this so many times for whatever reason, this crossed over in the normie culture. and out of tech culture and tech journalism culture, everyone seems to be reading this. So by popular demand, I'm going to go through this a little bit. I picked out three or four sections I think are important for understanding the message and approach of this essay, and then I'm going to respond to it and we'll try to get down to the ground truth here. So I'll have this on the screen here for people who are watching instead of just listening. All right, I'm going to read something here. I'll start with something from the introduction to the piece. All right, so here's Matt Schuber. I've spent six years building an AI startup and investing in the space. I live in this world, and I'm writing this for the people in my life who don't, my family, my friends, the people I care about who keep asking me, so what's the deal with AI, and getting an answer that doesn't do justice to what's actually happening. I keep giving them the polite version, the cocktail party version, because the honest version sounds like I've lost my mind. And for a while, I told myself that this was a good enough reason to keep what's truly happening to myself. But the gap between what I've been saying and what is actually happening has gotten far too big. The people I care about deserve to hear what is coming, even if it sounds crazy. All right, that's quite the setup there. There's some sort of classic AI reporting traps that are happening here. There's no actual information in it. It's pure emotional manipulation, trying to give you a sense of the digital ick, make you feel uneasy. It sets you up for this. The emotional state it puts you in, if you're not someone who's following AI closely, is like, Yeah, your worst suspicions are true. It's crazy what's going on out there. And you know what? All right, I'm going to let you in what's going on. That is a classic. Before we get to the content of this essay, that is a classic move. Like I'm going to reveal to you what's happening and it's worse than you think. I mean, that's like the classic move for everything. Conspiratorial thinking for radical health trends. It's a very compelling way to set up whatever you're going to say. All right, let's get into the content itself. I'm going to skip ahead a little bit now. Here's the first, I think, substantive thing that's said here. For years, AI had been improving steadily. Big jumps here and there. But each big jump was spaced out enough that you could absorb them as they came. Then in 2025, new techniques for building these models unlocked a much faster pace of progress. And then it got even faster and then faster again. Each new model wasn't just better than the last. It was better by a wider margin. And the time between new model releases was shorter. I was using AI more and more, going back and forth with it less and less, watching it handle things I used to think required my expertise. All right, I'm going to stop right there. This is our first piece of evidence that, oh, this writer is willing to essentially make things up if they fit the vibe that he's trying to portray. The way he describes this is actually, in some sense, the opposite of reality. As someone who has been covering the generative AI revolution very closely for the New Yorker and here for this show, this is not how things happen. It's opposite. Things were moving really fast when pre-training scaling was working. The jump from two to three and three to four were these impressive leaps. This is where you're in the steep part of the lost power law curve from the Kaplan scaling paper. The period he's talking about as we arrived in the 2025 is actually when progress slowed down. It became a problem for the AI companies. The general overall capability boost that happened from pre-training scaling stopped happening. And they had to shift instead to a backup plan, which was we're going to do this sort of post-training work on very specific tasks. And we're going to do things like post-inference time compute, and we're going to turn our focus from like general ability improvements that are clearly impressive to any user to instead chasing down these arcane named benchmarks that we can sort of teach the model to specifically do well on. And there was this whole period where this was for users were like, I don't really, the main thing I'm noticing is like the personality of the chat bots changing. And you're getting incremental improvements on specific activities where they could specifically try to post train for that activity. It was actually a bad period for growth, not a good period. So this idea that changes were speeding up. Most, I would say, close watchers are saying, no, no, this actually slowed down. And they had to try to find specific areas where some sort of notable improvement could make. They found video generation was one that ended up being a bit of a bust. And then the other place was in computer programming tools. So I think he's extrapolating continued progress on the computer programming tools, which I'll get into in a second. It was not exponential, but hard one with like the models in general. We're getting faster at some sort of bigger pace. It's simply not true. I know it fits the vibe of people talk about cloud code more, but it's simply. not true. All right, so let's read the rest of this quote here. Then on February 5th, two major AI labs released new models on the same day. GPT-53 codex from OpenAI and Opus 4.6 from Anthropic. and something clicked. Not like a light switch, more like the moment you realize the water has been rising around you and is now at your chest. Again, these were just continued incremental improvements on these coding-related agents, which I've been reporting on for a while. They've been impressive for a while. They've been making progress in these somewhat frequent but relatively small steps that are built on fine-tuning and other types of post-training that they're doing specifically around programming tasks. There is something like an inflection point where these latest models on certain sorts of code auto-generation, agentic auto-generation tasks, it got just to a level where more and more people were like, I think I can start using this more in my day-to-day. But these are kind of technical shifts, and they're very focused on what's specifically happening in programming. So the idea that there's these models in general are exponentially speeding up. This is the opposite of exponential. incremental steady progress on a small number of narrow applications or other applications that they thought the models would be much better at, so far they failed to be making progress on. All right, let's jump into the details of the programming itself. Matt writes, I am no longer needed for the actual technical work of my job. I describe what I want built in plain English and it just appears. Not a rough draft I need to fix to finish thing. I tell the AI what I want, walk away from my computer for four hours and come back to find a work done. done well, done better than I would have done it myself with no corrections needed. A couple of months ago, I was going back and forth with the AI, guiding it, making edits. Now I just described the outcome and leave. All right, I'll skip the final details here. So he's saying in the narrow world of computer programming, this is the inflection point advanced, that you can now, as a computer programmer, just tell the AI, this is what I want, and come back four hours, and you have that app built. He goes on to talk about that it not only builds the app, it tests the app, it fixes it. You don't have to do anything anymore. All right. So is this how people are now using this technology, the latest models that were released earlier this month? Well, who can tell? I can tell because I'm in the middle of a reporting project that I started just last week. So with the exact models he's talking about. where I have so far received detailed notes on how they use AI from active computer programmers. I have over 250 such case studies. I've made my way through about half of them so far. So I'm still kind of early in this progress. But here's what I can tell you. No one is saying, make me an app and walking away and coming back four hours later and is like, there I have it. Let's release this. That is not how programmers are using these very latest tools. That only works. For very specific types of apps, they have to be in one of a small number of like very common style of applications that are much more, and it's like special languages, sort of interface focused, not too big, and you don't need to be particularly stable. So you can, as like a hobbyist, kind of vibe code, hey, can you make me a Tetris game with, you know, Dungeons and Dragons characters or something? And we'll like do that. You can come back, you'll have something. But the 250 serious programmers I'm talking to, that's not the way they're using. the auto code generation, it's much more narrow and specified. Those who are doing this, and a lot of them aren't, those who are doing this talk a lot about how you have to use super clear specs. This is exactly what I want you to do. And then they let the model build that code for this piece. And then they have to extensively test it because, again, the models make mistakes. 20% of the time, right? And then they run a bunch of unit testing on it. Okay, I think this is working. Let's integrate that in. Okay, now here's the next thing I need you to do. And like one out of five of these attempts are like, okay, the AI just doesn't get it. I'll just do it myself. There's a lot of interesting stuff happening here with AI, but what he's describing so confidently is what a minuscule fraction of this broad sample of real active professional computer programs I talk to, a minuscule fraction is using the tools in this way. It's cool what's happening, but it's not, hey, go make me go do this. I'll come back four hours later. It's just done, and I'm moving on in my life. These are heavily supervised right now. All right. Let's keep rolling here. Here's the next quote from the piece I want to highlight. The AI labs made a deliberate choice. They focused on making AI great at writing code first because building AI requires a lot of code. If AI can write that code, it can help build the next version of itself, a smarter version, which writes better code, which builds an even smarter version. Making AI great at coding was a strategy that unlocks everything else. This is grade A nonsense. It's just vibey nonsense. These AI agents do not let us make better AI models. That's not how that works. That's not what's happening. They're very useful, especially like if you talk to these programmers like I've been doing, the reason why they're saving time is that there's a lot of very tedious tasks that happen when you build various software stacks or applications. building out the interface and connecting all the interface elements to the right functions, or integrating multiple different data sources into a common framework so that you can pull data from any of them. This is tedious type of coding, especially if you're not an expert at exactly, like you haven't built 100 of those type of apps before. And it takes people a long time because they have to look up, oh, God, what's the library call for the button? What's the, what do I, how do I access this sort of data source here for this? What's the call? I got it. What do I have to import? And this type of stuff, these models can do automatically. They know it already. Like, God, that's saving so much time. It's so tedious. I don't have to do the tedious thing. What they cannot do is like invent a new model of intelligence, improve the fundamental mathematics of machine learning, build us a better model for AI than we've ever seen before. That's not how this works. None of the innovations in generative AI are programming-related innovations. They're all conceptual mathematical innovations where someone who is an expert in machine learning realizes like, oh, reinforcement learning could be applied to a language model if we work through the different renormalizations of the vectors properly, and then someone goes off and programs it. So this idea that tedious code or code that requires you to look up a lot of information can be automatically coded and that saves you a lot of time, you cannot jump from there. to say, oh, AI can write itself now, and now we're going to have this self-reinforcing loop. That idea has been in the zeitgeist all the way back to the 1960s when J.L. Goode wrote the first paper on ultra-intelligence and introduced the idea of recursive self-improvement. It is not, not, not what these tools are meant to do. They cannot do that. That's not what's going on. This notion that the AI companies chose to build AI agents first, I mean, coding agents first so that they could build better models that could then do everything else is wrong. The reason why we're hearing more about coding agents is because it's one of the only narrow tranches of applications where they could find a market. They didn't choose. There's a lot of other things that the AI companies promised products in. I wrote an article back in January for The New Yorker about this, where in early 2025, they said, this is the year that we're going to have these general use agents for all sorts of jobs. We're there. It's going to happen. It didn't happen because it turns out that's much harder than coding agents. They put a lot of effort in the video generation and that did pretty well, but there was no market there because people didn't want to pay $200 a month to make TikTok videos. They want there to be other markets. They're just, the technology is not good enough. It's not interesting enough. It's not helpful enough. We're not seeing it move the needle in other positions as much. So we're hearing about coding because it's like the only place right now where there's real progress being made. And it is a good market. And this technology could be really useful for programmers. Again, I'm doing these surveys and I'm going to do a much bigger thing about this soon. But this is like a narrow thing that's happening right now. one of the places where these models have always been good, all the way back to the Instruct model, the InstructGPT model that was helped to make the GPT-3.5 that ChatGPT was made on. From the very beginning, the last half decade, we've known the one thing these models are good at is structured code because it's very structured language with lots of good training examples. And they've been making steady improvements on it using post-training techniques. And they've been passing various milestones as they do these. And this is a good story for the computer industry. And it's an interesting story. We might lose jobs. We might gain jobs. We don't know. We should cover it well. But the idea that the AI companies chose to do that first so that they could then make their own model smarter. And then there's going to be this takeoff and AI takeover. Straight up vibe nonsense. That's not what's actually happening. So, okay, I'm going to leave it here, Jesse, because this essay makes me a little bit upset. But let me be clear. Summarize. There are struggles with the AI industry. Post-GPT4, the failure of Project Orion, the failure of the BMF model, the failure of the Grok 3 failure where they moved to 100,000 GPUs for training and didn't get big improvements, the shift towards post-training, more incremental improvements, and benchmark chasing. I've written about this. Look at my article last August. for the New Yorker for more about this. This is the portrait of an industry that's not like it's failing, but it's also not going gangbusters. This is why right now the investment community is a little bit nervous about the stocks for the big AI companies. We need to see where your big revenue is going to come from and we're not seeing it yet. It's just a mixed story. It's a cool technology. They're trying to find markets. They're finding some niche markets, customer service, video production. That's a pretty small market, but there's good stuff there. And in programming, they're pretty good at programming. And they've been making steady incremental progress. And the tools are now good enough now that it's beginning to affect the actual workflow rhythms of non-trivial percentages of programmers. And that's a cool, interesting story. This essay is about a science fiction dream. This is not an inflection point for most people if you're not a computer programmer. This is not... From here, we get some rapid takeoff. From here, everything changes. And I've seen article after article after this essay came back, including one, if you read my newsletter today and get at calnewport.com, I dissect an Atlantic article that does so much vibes on exactly this. I just don't think that's an accurate way to think about this, all right? Interesting stuff is happening to computer programming. This is not an inflection point that AI is about to rise over our heads and change everything. Again, it's a task for which these models are supremely well-suited. And all this progress has been incremental, but steady to continue to refine and update because it's the only place where they're getting non-trivial monthly subscription fees right now is honestly in that space. So I don't want to say nothing's happening, but also I think this essay is alarmist. And I think he gets the technology wrong and he mixes truth with fiction and makes statements confidently that just aren't right. So there's a lot of good reporting out there about AI. This is not it. You could ignore this one. You could ignore this one. So I don't know. There you go, Jesse. Did you see that essay? Everyone sent me that essay. I did not see it until I saw it in the script. I mean, he wrote it in part with AI and you can tell. I don't know. The bigger thing, not the bigger thing. The other story here is these essays on X going viral is like. definitely i think a thing of the winter of 2026 and obviously that opportunity is going to get saturated and go away but remember that dan co essay we did that was also one of these x essays that went viral so i think we're in this moment where like you can go viral on x doing these long-form essays and i bet it's not going to last past march it's going to get saturated and then that opportunity is going to go away but man you're going to how many youtube videos you think are being made right now about how to go viral with your x essays there's probably so much content about this I don't know. I don't like to think of myself as an AI skeptic, but I see myself as an AI realist. I really want to ground everything I do and what's actually happening. I don't like this vibe approach. Well, I love the fact how you referenced on an earlier show as well that you're doing all the reading with your students from the past and you see all the recurring themes. Yeah. The doctoral seminar I'm teaching on superintelligence with AI doctoral students. Yeah. These themes are very powerful and they've come up again and again and again. people doesn't mean they're true right now. Hey, if you like this video, I think you'll really like this one as well. Check it out.

Cal Newport comments on a recent article by Matt Shumer concerning AI in this clip from episode 393 of the Deep Questions podcast.

Buy Cal Newport's latest book, “Slow Productivity” at www.calnewport.com/slow 

Download my FREE Deep Life Guide HERE: https://bit.ly/3QBIcug

Listen to Episode Here:  https://www.thedeeplife.com/listen/

Get your questions answered by Cal! Here’s the link: https://bit.ly/3U3sTvo

0:00 

Connect with Cal Newport:

🔴Visit Cal's BLOG and website:             https://calnewport.com/blog/
🔴Check out Cal's books:                         https://calnewport.com/writing/
🔴Check out The Deep Life:                     https://thedeeplife.com

About Cal Newport:
Cal Newport is a computer science professor at Georgetown University. In addition to his academic research, he writes about the intersection of digital technology and culture. Cal's particularly interested in our struggle to deploy these tools in ways that support instead of subvert the things we care about in both our personal and professional lives.

Cal is a New York Times bestselling author of seven books, including, most recently, A World Without Email, Digital Minimalism, and Deep Work. He's also the creator of The Time-Block Planner.

The videos are considered to be used under the "Fair Use Doctrine" of United States Copyright Law, Title 17 U.S. Code Sections 107-118. Videos are used for editorial and educational purposes only and I do not claim ownership of any original video content. I don't use said video clips in advertisements, marketing or for direct financial gain. All video content in each clip is considered owned by the individual broadcast companies.

#CalNewport #DeepWork #DeepLife #DeepQuestions #TimeblockPlanner
#WorldWithoutEmail #DeepQuestionsPodcast