Karpathy's AI Divide: Which Side Are You On?

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Karpathy’s AI Divide: Which Side Are You On? · Channel: Rob Shocks · 09:49 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, Task and workflow automation, Software engineering with AI

Summary

The video explains Andrej Karpathy’s observation that people’s understanding and use of AI fall along a widening funnel, from little or casual use to professional frontier-model use and access inside major labs. The speaker attributes the gap to differences in access, the uneven capabilities of models across domains, and—most importantly—the ambition of the tasks people give them. He reflects that even experienced users can underuse AI by building familiar, low-stakes replacements rather than tackling bigger problems. His advice is to choose an ambitious project with little downside, experiment quickly, and update expectations about what current models and tools can do.

Key points

Casual users often treat chatbots as better search engines or simple writing aids, while professional users employ frontier models for coding, math, research, and business problems.
Karpathy’s four-layer funnel ranges from billions of people barely exposed to AI, through casual chatbot users and professional frontier-model users, to a small group inside major labs.
The growing gap reflects differences in access, the uneven performance of models across tasks, and how ambitious users are with them.
AI works especially well in digital domains such as coding, math, and science, where results can be tested and feedback can improve performance.
Simple, one-off prompts barely test frontier models; larger goals over longer horizons can make better use of their capabilities.
The speaker admits that many of his earlier projects were familiar app or SaaS replacements, built faster rather than fundamentally more ambitiously.
Frequent experimentation and continued learning can compound over time, even for people who begin with little technical knowledge.
The practical recommendation is to start with a fun, ambitious, low-risk project and move quickly to explore what current agents can do.

Chapters

Start	Chapter	Summary
00:00	The growing gap in AI use	The speaker introduces Karpathy’s observation that casual chatbot users and people building with frontier agents experience very different levels of AI capability.
02:33	Karpathy’s four-layer funnel	The video describes a funnel from people with little AI exposure to casual users, professional users, and people inside major labs.
04:07	Why the gap exists	The speaker explains how access, jagged model capabilities, and the ambition of a task shape the results people get from AI.
06:11	Rethinking ambition and experience	The speaker reflects on having used AI to build conventional app replacements and describes turning toward a more ambitious software project.
08:15	How to push further	The video recommends trying a compelling, low-risk project, experimenting rapidly, and sharing what succeeds or breaks.

Quotes

most of us aren’t anywhere near as ambitious enough with these models. — 00:31

the interfaces are getting simpler, but the capabilities behind the interface is getting a whole lot more powerful. — 02:02

It’s impact scales with ambition. — 05:09

Get curious and then get ridiculously ambitious. — 08:15

Concepts and entities

Andrej Karpathy, Codex, Claude Code, OpenClaw, Harness, REA

Full transcript (original language)

00:00 Yesterday, in a post on X, Andrej Karpathy put his finger on a phenomenon a lot of us AI builders and developers have been feeling. Well, me for sure anyway. If you spend any time online or even chatting at a barbecue, there feels like there is this growing gap in understanding of AI capabilities and potential. As someone who has been deep in building agents for the last 4 years, it’s easy to get confused as to why everyone is not using these tools to their full advantage every single day. I have many thoughts on why this is, but first, here’s a breakdown of what Andrej

00:31 is saying with some graphics to help illustrate. And after going through this breakdown, my takeaway is going to be most of us aren’t anywhere near as ambitious enough with these models. That includes me, and I’m going to show you where I got it wrong. But first, let’s look at where this all came from. So, back in April, Karpathy posted that he was seeing this growing gap in how people understand and use AI. On one side, people who tried maybe the free version of ChatGPT, they’re kind of just using it as an advanced version of Google for the most part. They’ve seen some hallucinations, some quirks, the

01:02 strawberry thing, how many hours are in a strawberry, the car wash thing, and they kind of write it off a little bit. On the other side, you have people that went deep on this. People that are paying for frontier agents like Codex, Cloud Code, and using them professionally for coding, maths, and research. And for building their businesses. So, he’s saying that second group, and I would count myself as in that second group, has the highest amount of AI psychosis. Because the improvements have been, as he quotes, nothing short of staggering. And I definitely have felt that myself at particular points. Back

01:32 in December, we had a huge jump where the models were just getting everything I wanted done correctly with very little issues. I’m feeling that again recently in the last couple of weeks as my as my goals have become a lot more ambitious and I’m pushing the models further. And I don’t think this is necessarily an improvement in the models we’re seeing. It’s all the infrastructure and harness that comes with the models that have improved hugely. You can get a lot done now. Tools like GrokBot, Dots, Muse, etc. are the evolution of OpenClaw, a

02:02 more mature version of the promise that OpenClaw gave us with a lot more connected systems and security baked in. We’re seeing this phenomenon where the interfaces are getting simpler, but the capabilities behind the interface is getting a whole lot more powerful. Karpathy’s summary basically on this point is we have these two different groups that are speaking past each other. I’ll see this every week in real life conversations. I’ll talk to somebody who uses it like a slightly better Google, and then someone who has agents building a whole system over a weekend. They’re both telling the truth about what they’ve seen, but it’s

02:33 different. So, I guess Andrej has been thinking on this divide like everybody else, and he did some napkin math to to break it down. And it’s less like a divide and more like a funnel. So, in layer one, there’s about 6 billion people, roughly 3/4 of the planet. They have barely come into contact with these models at all. That is a lot. Okay, so layer two is 1 to 2 billion casual users of these free chatbots. And in his words, they see “derpy” chatbots. They use it as a better Google or basically just a simple writing aid, maybe just to do a quick report or get their homework

03:04 done. And this isn’t just lay people. These are executives and professionals sitting in this zone, too. Execs, for example, and anyone whose job is mostly talking to other people. They get it intellectually, but it’s still second-hand to them. So, layer three is around 20 million people, about 0.2% of the population. And they’re using frontier models professionally in math and coding and getting big business problems solved. These are the people watching projects that used to take 2 weeks or months get done by an agent in a one-shot prompt.

03:36 They’re building apps, translating apps, even decompiling and reverse engineering apps from binaries. Check out my video 2 days ago on REA. Karpathy himself admits that even just less than a year ago, he was still writing code by hand, maybe using auto complete, but he was still doing it line by line. So, layer four is about 5,000 people with access to frontier systems inside the big labs. They’re seeing swarms of thousands of agents working together for weeks on mega projects, finding zero-day issues, discovering new science, and pushing the frontier of maths, as we’ve seen

04:07 recently with OpenAI. We often see these wild quotes from the likes of Boris Chen or people who are within the large labs. They have access to endless amounts of compute and can run grand scale experiments. They also have the leading version of models before everyone else gets there. It’s really interesting to see the kind of things they’re talking about to get a sense of what’s coming next. Okay, so why does this funnel exist? And he gives three reasons. So, the first is access, meaning you’re either on a free tier, you’re on a paid tier, or you’re working internally in one of these labs. That’s the obvious

04:38 one. So, the second part is what he calls jaggedness. So, these models don’t work the same for all problem domains. They’re not equally as powerful. So, for example, when the work is digital, uh it’s easy to verify like maths, coding, science, AI is perfect for that. Basically, it has a test that it can pass, it can do reinforce learning on top of that, there’s a feedback loop, and that really makes AI shine. If you’re trying to apply it into other areas that are less deterministic, it might not be uh as much of a win. So, if you’re complaining that your AI isn’t

05:09 great at writing a good email, that’s because it’s much harder to score that. So, email writing improving is slower. His point here is the exact same model can go and refactor a hundred thousand line code base, but it can also tell you to walk to a car wash instead of driving your car. So, his third point is the one I can’t stop thinking about and the one I’m really focusing on. It’s impact scales with ambition. So, in his words, a question with a paragraph answer barely stresses the system. If you’re using ChatGPT in these one-off threads

05:41 to ask simple questions, do Google search, maybe do a little bit of shopping comparison, or write simple reports, you’re barely stressing the system. You need to start thinking about projects in big goals over longer horizons. And if they’re code-based or verifiable, even better. We all have big problems in our life. We all have these issues that we see around them. You need to build up a reservoir of these big difficult problems that you want to solve and actually start to work on them using these agents. We’re not being half ambitious enough about what we can solve

06:11 and what we can do with these models. So, paying for a top-tier access or $200 a month to cloud code doesn’t move you down the funnel itself. You can still ask a frontier model for a paragraph-sized question and get paragraph-sized results. It’s kind of like trying to use a chainsaw to sharpen a pencil. And here’s where I have to be kind of honest with myself. So, personally, I really thought I was in layer three. I mean, I’ve been working with these agents for 4 years. I get early access to tools and models from the labs. I build with these tools every day. But, when I’ve looked back at what I’d actually built, a lot of it was just

06:42 copies of apps that already existed. I was just replacing SaaS that I was using. I’d build a slightly different receipts tracker or a different dashboard system. Don’t get me wrong, they’re useful, but I was essentially using the intelligence of a frontier model to build things I could have built 5 years ago, just faster. So, it’s only in the last few months that I’ve started being properly ambitious. Like right now, I’m building my own self-learning software factory in Harness. I’m enjoying what the big companies are putting out in this area, but inevitably, this is too important for me not to have control of. I’m basically

07:12 getting frustrated with what’s on offer. Every tool is built around how someone else works, and we have this opportunity now to build something bespoke to how we work ourselves. And a year ago, I would not have even considered that as a one-person project, but now it is, and I think it’s possible for almost everyone watching this video. So, when I look at the people I know who are all in on AI and seeing what it can do, they’ve got something in common. They’re using it constantly. Every hour of every day, they’re experimenting. They tend to be curious and technical, the kind of

07:43 people who are drawn to building, hacking, and making things. And to be honest, if you’re neurodivergent in any way, you’re hitting a sweet spot with agents. And all of this learning is compounding. I’ve seen this in my own community where people step in with very little technical knowledge. They’ve hung around for a year or so, and what they’re building now is blowing me away. If you’re only paying lip service, if you’re using it lightly, if you’re just using it for basic chat, you’re missing out on that compounding. And you have to remember that the gap between these layers is widening. Some people and some companies who fall into their own flywheel are going to move a lot further ahead. I know if you watch this channel,

08:15 if you’re subscribed, you’re already AI pilled. I don’t really need to convince you. But what I’m saying is maybe it’s time to go up a notch. Get curious and then get ridiculously ambitious. Why not? So what do you do practically here, especially if you’re a developer that is only just starting to use this lightly? I would say pick a fun ambitious project that you’re really interested in. Something maybe start with that has zero risk. Nothing at work, nothing critical that anyone else is dependent on at an early point, nothing that’s going to matter if it breaks. Now, even if you’re using agents for the last couple of years, what you want to do next is go

08:46 hard and fast. Throw out a lot of the validation, the verification, throw out a lot of the rules that you previously had about the speed that you like to move at and just see what the models are capable of doing now over a couple of projects. One disadvantage I see with a lot of developers who have been using this for quite some time, they’re still using a mindset from 2024, even 2025, where the models and harnesses weren’t as good as they are now. I think you’ll be surprised about how far the edges have moved. If you want to do this alongside other people who are pushing it just as hard, that’s what the Switch Dimension course and community is for.

09:17 The link is in the description. The waitlist is open. We’re announcing the new cohort in the next 2 weeks. We’d love to see you in there. So I’d love to know in the comments, what’s the most ambitious thing that you’ve been trying to build with AI? And where was it successful and more importantly, where did it break? If you want to see a real layer 3 example, watch my video on REA next. It lets your coding agents reverse engineer apps, which is the decompiling thing Karpathy mentioned in layer three. I launched that video maybe three or four days ago. Already, I’m seeing loads of use cases of people using it in clever ways. One of the most interesting

09:49 areas is actually the recycling of old hardware. People are using it to jailbreak printers so that it’ll accept any cartridges. Lots of different things like that that I hadn’t anticipated. Check out that video next. If you found this useful, subscribe and I’ll see you in that video.