Transcription
So, Anthropic published their fourth Anthropic Economic Index report. They're tracking how AI use is affecting automation, jobs, the economy, etc. And there are some pretty important takeaways in this one, some of which are surprising. The idea of AI taking over all jobs and all us humans just kicking back and sipping my ties on the beach. Well, there's a sort of an asterisk next to it. And now, there's some things that might not be quite how we imagine them.
Before we dive in into this post as well as their 55-page PDF report they've published, here are kind of the big takeaways that you need to know about. So, first of all, the era of Agentic AI has arrived. We have Cloud Code. We have Claude co-work. It's no longer just answering questions. It's doing actual work. It's accessing files. It's executing commands. It makes a plan and it follows it through. But different segments of society use it very differently. And true automation isn't progressing quite as fast as we projected or hoped in some cases. There's a lot here. We'll break it down step by step, but I think it's safe to say to kind of sum it all up is that coding is kind of the canary in the coal mine, if you will.
So, we're seeing a lot of progress towards automation in coding using AI. It's the first sector to really truly experience this agentic automation. But recently, of course, we've had Claude co-work release, which basically is the same thing for everyone else for all the other tasks. So, it's very likely that we're going to be seeing that wave, that impact hit other sectors, other industries as well. So, that's why we're looking a lot at how the coding sector is sort of performing, how it's affected, because largely, again, that's what canary and the coal mine means. It's sort of a way for us to test what happens to this one area that's going to have sort of a mirror effect on the rest of the areas.
One very interesting concept that they're talking about here is upskilling versus deskilling. So briefly, what is deskilling? Imagine all these little dots are kind of the skills and things that you do at a particular job. The ones up here are more difficult. They're harder. The ones down here are easier. So if AI comes in and it's able to take care of a lot of the hard tasks in your profession, well, the stuff that's left over for you to do is the easy stuff. That's deskilling. So imagine something like a legal secretary. For example, in the past, let's say they had to review legal publications for a particular case to identify relevant court decisions. This was a task that is predicted to need 17 years of experience in the field. And the tasks that are left over are likely to be much easier. Put the whole thing together in a binder, make some coffee. I don't know, like the easy stuff, the actual kind of execution of the tasks, right? So you put it all together and you take it to the courtroom or whatever. You give that report to somebody, whatever that is. It's the easy tasks. The actual execution of the task becomes the easy part. So that's deskilling. You need less skill, less attention, less experience. It becomes a much easier, maybe a little bit more boring job to run.
And on the other side, we have upskilling. And this is kind of, I think, the dream scenario if you happen to have a job like this. And that is, let's say if this is kind of all the tasks that you have to do. Of course, higher up is the hard stuff, lower is the easy stuff. This is where the AI takes off your plate a lot of the easy tasks, all the administrative tasks, the boring, the routine tasks, it automates it. So you no longer have to do it. This, of course, leaves you with the high-skill tasks that you have to do. The example to give is, for example, a property manager or a real estate manager, right? So AI does the routine bookkeeping, maintains sales records, reviews the rents against the market rates, right? Does research on the market rates, etc., etc. So all these little boring tasks are handled completely by the AI. What is the human doing? Well, they're negotiating contracts with the architecture firms, for example, right? So high-level negotiation, etc. They're securing loans, developing banking relationships. They're managing various stakeholder relationships, as they're called. These are all high-skill areas of expertise. So the job shifts from admin work to very high-skill work. So likely, if you're in those areas, the value of your work goes up, the value of the worker goes up.
And they also have this very interesting chart basically saying that while an AI can handle some of the tasks for a particular job role, that doesn't necessarily replace the worker. It has to specifically automate the skills that are core to that job. So, for example, with data entry keys, the core of the job is actually entering data. AI is able to do that. So they're at risk for replacement by AI. Their core skills are replaceable by AI. On the other hand, we have something like a microbiologist where even though there's tons of stuff that they do that AI can do just as well, the core tasks the AI can't do. So they're not going to get replaced. Core tasks being, you know, putting stuff in a petri dish, looking through things through a microscope, pipetting stuff.
And this is probably the thing that a lot of people are going to focus on. This is the thing that's going to be used, depending on whether you like AI or not. Kind of what people are going to use to say, "See, I told you so." The point being here is that the task success rate drops over time the longer the task duration is. And the orange and the blue lines, I mean, this is cloud AI in orange versus using the API. More businesses trying to automate stuff, they use the API. So, as you can see here, the blue drops off much more rapidly. The thing to understand here is here the human is kind of babysitting it, correcting it, course-correcting it along the way, therefore improving its task success over time. Without the human babysitting, right, it rapidly drops off. So this is important. This is very important because it shows that some of the more optimistic projections for AI automation and the productivity boost, this is suggesting that they may not be accurate. We have to downshift our expectations a little bit.
This is important because according to Anthropic here, they're saying adjusting productivity estimates for task reliability roughly halves the implied gains from 1.8 to about 1 percentage points of annual labor productivity growth over the next decade. However, these estimates reflect current model capabilities and all signs suggest that reliability over increasingly long-running tasks will improve. So, it's important to understand that while they're still sort of pulling back on the optimistic expectations, I mean, it's still going to have a big, big impact. So, we went from insane to huge. So, but still, it's less than was previously expected.
All right. So, with that said, let's kind of dive in and see what they're talking about here. We covered kind of like the big points, but let's get into the nitty-gritty, if you will. So, first and foremost, that cloud usage remains concentrated among certain tasks, right? The top 10 most common tasks account for 24% of the sampled conversations. A slight increase since our last report. So people are using it for specific things. And automated use remains dominant in WP API traffic, reflecting its programmatic nature. This is kind of a big deal because on the business side, they're using the API to try to automate everything. And then on the consumer side, the ones that are using Claude, the ones that are using the chatbots, it's of course a little bit more back and forth and pretty concentrated in certain areas.
Global usage remains persistently uneven. All right? So different countries have very different usages of these AI tools, while US states converge. So this is interesting. So the United States, all the different states are becoming more and more equal in terms of how they're using these AI tools. This is kind of good because in the past, we've seen kind of a divergence. So tech use was concentrated for the first movers who got an advantage because they knew the tech, because they were able to utilize it. That's not happening here, at least in the United States, because all states are beginning to use these AI tools at a similar rate. So the diffusion is going well. That's probably a good thing. And worldwide, even adoption remains well explained by GDP per capita. Augmentation is once again more common than automation on cloud.AI.
I got to bring this up here because throughout this whole time, I'm running Cloud Code in the background. I am designing something, or Cloud Code is building it for me, I guess I should say, a very specific tool that's going to help me edit some of the videos a little bit better. So it figured out some machine learning open-source thing that allows it to recognize speech. I've been trying to shoot more of my videos in 4K, and that results in these giant files. So sending it to my editor, sending it back is quite a nightmare. Uploading it to places is a nightmare. But if I can automatically run a sort of a first pass with this machine learning thing that cuts out some of the spaces and does some of the grunt work upfront, that would be kind of a big deal and would really help me out. So I've had multiple iterations with Cloud Code. It worked very, very well. We kept kind of building on top of it, but towards the end, it was running very slowly. I was like, "All right, can you speed it up? I have a pretty high-end GPU. I told it, you know, can we use the GPU for that?" It's like, "Oh, well, let's see what we can do." So, now it's been cranking away. Look at that. 3 hours, 10 minutes, and 45 seconds. This thing is just cranking out. It's code that so far is working pretty, pretty well. Just thought I'd throw that in there.
So, within the US, lower usage of states have relatively faster gains in adoption. So, it's converging. There's still a lot of inequality, but it's converging. So moving in, I think that's moving in the good direction, in the right direction. They're saying if sustained usage per capita, it would be equalized across the country in 2 to 5 years, a pace of diffusion roughly 10 times faster than the spread of previous economically consequential technologies in the 20th century. So AI is a lot more democratic. It spreads a lot faster. It's a lot more equal in terms of how people are using it.
So, one thing that the report kind of mentions that I found interesting because I've noticed this as well and actually been trying to figure out how to improve this issue. I want to say, although it's not even an issue, it makes sense that it works this way, but basically they're saying that we find that Claude generally succeeds at the tasks it's given, and the education level of its responses tends to match the user's input. So, this last part is the important one. The education level of its responses tends to match the user's input. So the correlation between the human education, kind of like the prompt, what words they use, a high education response, like the answer that it gives, is almost a perfect correlation, right? So if you ask it like, "How to cook an egg?" it'll say, "Cook egg in hot pan," right? It sort of replies in that kind. If you say, "What is the best way to cook scrambled eggs?" you'll say, "Whisk your eggs in a bowl with salt and pepper," blah, blah, blah, right? So it kind of responds on that level. You say, "I require a protocol for the optimal thermal coagulation of avian ova." It'll answer you. "To mitigate the risk of protein syndesis, one must apply gentle thermal energy. Agitate the emulsion continuously."
So, I've encountered this quite a bit. The problem that I struggle with often times is I want the PhD-level response, like that level of information, but broken down into more sort of manageable and understandable ways. So with the ChatGPT custom instructions, I actually have a custom instruction in there. And before, I would just type it in manually, or actually, I would copy and paste it because I had it off to the side in one of those notepads that would say like, "Actually give the high-level response and then underneath break it down in easy-to-understand language with insights, etc., etc." So spit out like the block of text that would be unreadable by most people if you're not in that specific industry, that would be very hard to understand, and then it would kind of give you the breakdown with those insights. And that tends for me to produce the best results because you're not losing any sort of resolution. You're not losing any of the insights, any of the data. But this is kind of an important, it jumped out at me because how you talk to the thing is how it's going to answer you back. And so you might not be getting the top-tier answers when you're looking for those, if you're not phrasing your prompt correctly.
So right now, it does seem like it's showing that instead of kind of fully automating people, the work will shift from actually doing to more managing. So whereas before workers would be actually doing the stuff, writing emails, entering data, writing code, slowly we're going to be shifting to managing, like looking at the proposals that the AI gives us, reviewing the plans, approving the AI edits. It's funny because that's literally kind of like what I'm doing every once in a while when Cloud Code pops up. It's like, "Hey, should I do this thing? I have this idea how to improve the product. Should I do it?" You're like, "Yes, go forth, do it. I will allow it." Or if it screws something up, you're like, "No, that did not work. Do it again. This is unacceptable." Right? So instead of sitting there trying to figure out on my own, I'm just kind of reviewing what it's doing.
So since the report notes that human prompts are highly correlated with education levels, this suggests potentially, let me know if you agree with this, that this idea of managing AI will be a fairly high-skill activity. One other interesting thing that they've pointed out is that those kind of high-skill tasks that are complex, that take a while, while they are on the whole harder to succeed at, it's kind of this thing where you see this rapid drop-off, right? So it gets, you know, for the API task, for the fully automated tasks, it drops below the 50% line by, you know, call it hour three, maybe hour four or five, etc. Whereas the ones that are getting babysat by the human, you know, it's much smoother. It sort of remains more successful over time. But here's one thing that they noticed for these higher complex tasks: when they do work, when they do succeed, because some of them are just more prone to be able to be successfully completed, they offer the highest kind of speed-up multiplier. And I got to say, if this thing manages to build exactly what I need here, I will be kind of blown away.
What's interesting here is it's actually, it's not actually actively working. It's running the program and testing to see if it works, which is kind of mind-blowing, I think, if you haven't used Cloud Code too much. Like, it built a program and now it's sitting there troubleshooting it. It runs it and then it looks at the output. It's like, "Hmm, something's not quite right. Let me try that again." And it's been at it for three and a half hours. I can't focus on anything for three and a half hours to save my life.
One other very interesting point is how different countries use Claude. Basically, as the GDP per capita increases, so sort of the wealthier countries, they tend to use it for business, for work, for personal use. And as it decreases, they're more likely to use it for coursework. So, this seems to be a good thing if you're interpreting it from that perspective, that it allows that kind of information accessibility. It kind of evens the playing field in terms of education. Now, this is of course assuming that AI improves education. So, when people are using it for education, they're using it to gain better skills, etc. I haven't seen any specific data on that quite yet because I'm sure the institutions, the colleges and universities, there's going to be a little bit of a time of flux while they adapt to these new AI tools, I think.
So, it's kind of hard to sum this whole thing up in a sentence or two because there's definitely a lot of different things that are happening here. There's a lot of complexity, but on the whole, it seems to suggest like if there's a spectrum between full automation and one where we use augmentation more, where humans are still in the loop and they're using these as AI tools, you know, maybe it's not super leaning towards automation. It's just pulled back a little bit. Like, yeah, there's going to be a lot of automation. There's going to be a lot of things that will be automated, but you still need that person managing the AI, and you do still need that person to be, you know, educated, intelligent. They need to have certain skill sets and certain capabilities to be able to manage these AI agents effectively.
However, there are going to be certain professions where we are going to see kind of a hollowing out as AI takes over a big chunk of what they do, and they become more administrative assistants, right? So, the AI runs a lot of their stuff, and they just kind of are in the background just making sure everything's okay. They gave travel agents as an example, right? So planning complex itineraries, doing research, explaining why different places might be good for these specific people that want to go somewhere. Those are tasks. They take time. They're high-skill, but AI can do most of that. So what is left for the travel agent to do? Well, I don't know, book tickets, take payments, like basic administrative tasks.
And there's also going to be kind of a bottleneck economy, if you will, because as a lot of things get automated or accelerated, the things that still are that can be accelerated, automated, they're going to be the bottlenecks. It's the few critical things that the AI can't do: physical hands-on work, in-person, high-level negotiations. So, likely what we're going to see happening is productivity is going to skyrocket in certain more abstract industries like coding, data entry, etc. But it's going to remain stagnant in more physical human domains. And this is probably going to make those human domain services, those human-centric services relatively more expensive, valuable. So if you're worried about a job, maybe think about what bottlenecks there will be.
So it does almost seem like two paths. One is the slowdown, work with humans, you kind of get paid more, right? Things that can't be truly automated. Or lean into the automation, right? Take advantage of this rapid influx, this increase in productivity, etc. Use the new tech, how quickly you're able to build things to kind of take advantage of that. But in terms of kind of this massive job apocalypse and the needing of UBI, etc., etc., you know, I'm sure we're still going to need it. I don't think it's still off the table, but it does seem it's not quite as lightning fast as maybe some people expected. Just a little bit more chill. You know, we're driving towards it at 65 mph, not 100 mph. Like, we're still going to get there. We still need to start thinking about it. We're just not going there at some insane breakneck pace.
Anyways, let me know what you thought about that. Would you prefer to deskill or upskill? And do you think you're better off going into kind of like the digital abstract AI-assisted automation productivity-increasing areas? Or do you prefer maybe to lean more into the human-centric professions? Let me know in the comments. If you made it this far, thank you so much for watching.