The argument in one line.
Karpathy's Software 3.0 thesis means most existing apps and digital businesses are already obsolete, and the four moats that remain — proprietary data, prompt engineering, system design, and audience trust — belong to the builders who move first.
Read if. Skip if.
READ IF YOU ARE…
- A creator, course seller, or service provider who senses AI is disrupting their market but hasn't found a clear framework for deciding what to build next.
- A solo founder or indie builder trying to identify which parts of their current business model are defensible in a world where LLMs can replicate the output layer.
- Anyone who has heard of Karpathy's Software 3.0 concept in passing and wants a concrete, applied breakdown of what it means for non-engineering digital businesses.
SKIP IF…
- You're already operating an AI-native product with proprietary training data and a defined moat — this is positioning-level framing, not implementation detail.
- You're looking for technical depth on how to build Software 3.0 systems; the video maps the landscape but does not cover architecture, tooling, or execution specifics.
The full version, fast.
Using Karpathy's Software 1.0/2.0/3.0 framework applied to four concrete business models — installers, apps, courses, and video editing — this video argues that AI has already rendered most existing digital businesses obsolete. Software 3.0 replaces explicit rule-following with outcome-oriented agents that debug and adapt in real time, making app-layer products like Karpathy's own MenuGen instantly redundant. The four moats that survive are proprietary training data, prompt engineering skill, intelligent system design wrapped around the AI engine, and audience trust built before the transition. Solo builders and creators who understand these moats now have more economic opportunity available than at any previous point in the industry's history.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:39
01 · The nuke drop
Hook: AI is destroying every standard business model. Promise: Karpathy has the blueprint, early movers win.
01:39 – 04:16
02 · Software 1.0 / 2.0 / 3.0 defined
Host frames Karpathy's thesis: explicit rules to learned weights to LLM-as-interpreter. Plays the Sequoia interview clip.
04:16 – 06:27
03 · Example 1: The installer
Three-paradigm breakdown of a software installer. 3.0 = one-command agent that self-heals any error it has never seen before.
06:27 – 07:31
04 · Example 2: Karpathy's MenuGen app is already dead
Karpathy built an app to add photos to restaurant menus. Then realized one Gemini prompt does the same thing. The app should not exist.
07:31 – 09:30
05 · Example 3: The course business
Udemy (1.0) to engagement-optimized video (2.0) to personalized coaching agent like Alex Hormozi's LLM (3.0). Most course sellers have not hit 2.0 yet.
09:30 – 12:01
06 · Example 4: Video editing services
Premiere Pro (1.0) to Descript AI trimming (2.0) to text-prompt agent that edits in any style in 10 minutes (3.0). Services 3.0 = selling the outcome.
12:01 – 14:03
07 · The four moats
Knowledge (your data), Instructions (prompt engineering), AI Engine (leverage not build), Audience/System (trust + UX wrapper). The only defensible positions left.
Lines worth screenshotting.
- Karpathy's own MenuGen app — a software 2.0 product he built — is already obsolete; the same output now requires only a text prompt to Gemini.
- Software 3.0 does not just make apps faster to build; it makes entire categories of apps redundant.
- An installer in the Software 3.0 world is a single command that hands the goal to an agent — the agent debugs problems it has never seen before.
- Software 1.0 follows explicit rules; Software 2.0 learns from data; Software 3.0 takes a goal and figures out the steps itself.
- Most course sellers haven't caught up to education 2.0 yet, and education 3.0 — a real-time AI coach guiding action as you take it — is already shipping.
- The four surviving moats are proprietary training data, prompt engineering skill, system design around the AI engine, and pre-built audience trust.
- A random person claiming to have trained an LLM on Elon Musk's knowledge is worthless — the same product from Elon Musk himself is a business because of trust.
- Higgs Field competes with free LLMs not on model capability but on system design: the interface and workflow wrapped around the engine are the product.
- Alex Hormozi's AI isn't just a knowledge base you query — the next version will sit beside you as you build and guide each action in real time.
- Businesses that place AI at the center as the engine still need a human-designed car around it — the engine alone is not a product.
- The video editing agent doesn't need generic MrBeast style instructions — it can edit in the style of your own last 100 videos using pattern recognition.
- Descript-style AI editing is still Software 2.0: you open a tool and operate it with your hands; Software 3.0 edits are initiated with a text box.
- The creators who understand these four moats before the transition completes will find more economic opportunity than has existed at any prior point.
Four Moats That Survive the Software 3.0 Shift
The framework
Every digital business model — apps, courses, services — is being made obsolete by LLM agents, but four defensible positions remain for builders who move now.
01The nuke drop
- Every existing digital business model — apps, courses, services, agencies — is being displaced by AI, and the builders who map where this is going will find more opportunity than those who resist it.
02Software 1.0 / 2.0 / 3.0 defined
- Software 1.0 follows explicit hand-written rules; 2.0 learns patterns from data; 3.0 uses an LLM as the interpreter — and the programming language is now plain English prompts.
- Any product built around a fixed rule-set that an LLM can now execute on demand is already obsolete, regardless of how recently it was built.
03Example 1: The installer
- A Software 3.0 installer does not follow a script — it receives a goal, reads the environment, and loops until it solves problems it has never encountered before.
- This self-healing, goal-directed behavior is the template for understanding what 3.0 looks like across every business category.
04Example 2: Karpathy's MenuGen app is already dead
- Any app whose sole job is to perform a task that a general-purpose model can now do in a single prompt should not exist — it is competing against an infinite free version of itself.
05Example 3: The course business
- Most course sellers have not reached 2.0 yet — the 3.0 version is an AI tutor that sits alongside you while you take action, not a passive video you watch in order.
- The shift from knowledge delivery to guided real-time action is where the value in education moves next.
06Example 4: Video editing services
- Service businesses at 3.0 sell outcomes, not hours — the deliverable is a completed result requested in natural language, not access to a tool or a manual process.
- An agent trained on a specific editing style and given a text prompt can produce a finished cut in minutes — the moat is not the software but the style data and the trust behind it.
07The four moats
- The first moat is proprietary knowledge: your data, your training set, your demonstrated expertise — an LLM trained on your unique output is harder to replicate than a skill.
- The second moat is prompt engineering: the instructions and context you wrap around the engine determine output quality in ways that generic users cannot easily match.
- The third moat is system design: choosing the right AI engine is table stakes — the defensible part is the interface, workflow, and UX you build around it.
- The fourth moat is audience trust: the same AI capability means far more when it carries the credibility of a recognized expert than when it comes from an anonymous source.
Terms worth knowing.
- Software 1.0
- Traditional software where a developer writes explicit, step-by-step code that the computer follows exactly — every rule and edge case must be spelled out in advance by a human programmer.
- Software 2.0
- Software that learns its behavior from data rather than hand-written rules, using machine learning and neural networks to recognize patterns and improve over time through feedback loops.
- Software 3.0
- A programming paradigm where natural language prompts in a context window replace code, and a large language model acts as the interpreter — executing goals intelligently rather than following explicit instructions.
- LLM (Large Language Model)
- An AI system trained on large volumes of text that can generate, summarize, translate, and reason about language; the underlying technology behind tools like ChatGPT, Claude, and Gemini.
- MCP (Model Context Protocol)
- An open protocol that allows AI models to connect to external tools, APIs, and services — enabling them to take actions beyond generating text, such as calling apps or processing images.
- Nano Banana
- A Google-built image generation and overlay tool referenced by Karpathy as an example of a Software 3.0 service: accessible via a single prompt rather than requiring a downloaded application.
- Prompt engineering
- The practice of crafting precise instructions, context, and constraints for an AI model to reliably produce high-quality outputs — treated as a professional skill distinct from traditional software development.
- Moat (business)
- A durable competitive advantage that protects a business from being easily replicated or undercut by competitors, analogous to the water-filled trench surrounding a medieval castle.
- BYOK (Bring Your Own Key)
- A product model where users connect their own AI API credentials rather than relying on the platform's subscription, giving them direct control over model access and usage costs.
Things they pointed at.
02:00linkKarpathy Sequoia Capital interview (AI Ascent)
Lines you could clip.
00:00
“AI has dropped an absolute nuke on everything that we know and loved.”
06:18
“All of my MenuGen is spurious. It's working in the old paradigm. That app shouldn't exist.”
12:01
“Software and business 3.0 is selling the outcome.”
12:47
“You get to design your own car. You get to design your own wheels. You get to design the entire system around that engine.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00For the last decade, the business playbook has been simple. Make a course, offer a service, build an app, start an agency, create some software, or even sell a little bit of specialized knowledge. But AI has that we know and loved.
00:14It is already taking profit away from these outdated business models and will eventually take down even the biggest companies who are building on top of these old school foundations. But the problem is, most people right now are lost because they have no idea where we are heading, and therefore have no idea what's even gonna be valuable in the next few years.
00:33If everyone has an AI that can build their own course, write their own code, make their own app, what is actually going to be left behind that is valuable for us to offer? Well, Andre Kapathi, the cofounder of OpenAI, the ex head of AI at Tesla, and one of the most important AI engineers in the world just laid out the blueprint that we need to understand exactly where we are heading and how to adapt our businesses towards this before it's too late.
00:57And the good news is, Kapathi says, for those of you who can grasp this concept early, like right now, will find more opportunity than has ever been possible before. And these products and services will offer far more value to the world and therefore allow us to make far more money. So in this video, I'll break down Andre Kapathi's exact blueprint by using four real world examples so we can first understand what not to build in this new paradigm, and then we're gonna go on to what Andre Kapathi recommends that we all start focusing on and building that will not only get you ahead of your competitors, but allowing you to tap into a new cascading river of demand for your modernized products and services.
01:35All I ask in return is you hit that like button down below. Let's jump straight in. So this comes from Andre Kapathi's recent interview with Sequoia Capital.
01:45And in this interview, Andre Kapathi reveals his thesis on the evolution of software. Now when you hear software and you hear Andre talking about software in this interview, just know that he's talking about all digital businesses because they all have to be projected through this software lens.
02:01It'll make more sense once you hear what he has to say. Yeah. Exactly.
02:05So software one point o, I'm writing code. Software two point o, I'm, um, actually programming by creating datasets and training training neural networks. So software three point o is kind of about your programming now turns to prompting, and what's in the context window is your lever over the interpreter that is the LLM that is kind of, like, interpreting your context and, uh, performing computation in the digital information.
02:25And so Andre Kapathi is highlighting what used to be the software space or the business spaces we're gonna be interpreting it as, where it has evolved to, and then where it is going and transforming into right now. But when he talks about it, it's pretty ephemeral. It's like, what does that stuff even mean for businesses?
02:41Well, we're gonna go through some Andre Kapathi examples together. He starts with the example of an installer, which in the software or the business one point o space, you have to understand you need to follow explicit rules.
02:55So Andre Kapathi talks about when you're coding an app, you have to write all the code yourself, and all the code must entail every single piece of logic or every single action you want that software to perform. It's literally like having a steps one to five. It talks about an installer.
03:08So if you're installing some new software, it needs to be downloaded on the computer, and then the installer code will be like, step one, ask them to agree to term and conditions. Step two, find a folder on their desktop called Tempo or whatever it is, install it on step three, step four. And if it runs into any snags, it's in trouble because it's not a self healing software rule book unless it's in if this happens, try this, then try this.
03:30But you still have to write down every rule step by step for the software to actually follow it. Now, when we evolve into software two point o, you can have learned weights. So in the example of an installer, you download the installer onto your computer.
03:45If it runs into an error, it'll ask you to report it back to someone like Apple or whoever created the actual software. Then what Apple can do is they can machine learn based on all the errors and update their installer so that it has errors less and installs Seemly more often.
04:00There's this feedback loop where you get this machine learning and improvement. But Andre Kapathi now talks about what an installer three point o looks like. And in fact, him trying to install OpenClaw in a three point o way is what opened his mind to this new world.
04:15We can hear it from him.
04:18For example, when you when OpenClaw came out, when you want to install OpenClaw, you would expect that normally this is a bash bash script, like a shell script. So run the shell script to run to install OpenClaw.
04:29But the thing is that in order to target lots of different platforms and lots of different types of computers you might run-in OpenCLaw, these these shell scripts usually balloon up and become extremely complex. But the thing is you're still stuck in a software one point o universe of wanting to write the code, Actually, the OpenCLaw installation is a copy paste of a bunch of text that you're supposed to give to your agent.
04:49Basically, it's a little skill copy paste this and give it to your agent, and it will install OpenCLaw. The reason this is a lot more powerful is you're working now in the software three point zero paradigm where you don't have to precisely spell out all the individual details of that setup.
05:02The agent has its own intelligence that it packages up and then it follows the instructions And it looks at your environment, your computer, and it kinda, like, performs intelligent actions to make things work and debugs things in the loop. And it's just, so much more powerful. So Andre Kapathi talks about an installer in the business and software three point o world is basically an agent that you can call upon with one command.
05:24For example, the OpenCore bash command, which will install the little agent on your computer who then has a goal of making sure that OpenCore gets installed and any problems that he runs into on the way, he is going to keep looping until he solves them even if they've never seen that problem before. Okay. That's cool.
05:41But we probably don't run installers as our business. Andre Kapathi actually created an app for himself in the last year called MenuGen, where you can take a photo of any menu that doesn't have photos on it, and it'll give you a digital form of the menu that will render the photos embedded next to the names of the items.
05:57But now he's saying his app, which is built in the software two point o world, is instantly useless. Take a look. So menu gen is this idea where you come to a restaurant, they give you a menu, there's no pictures usually, so I don't know what any of these things are.
06:10I've I've coded this app that basically lets you upload a photo. It re renders the menu, and it gives you, like, all the items, and it, uh, use the image generators to get pictures of them and then shows it to you. And then I saw the software three point o version of this, which is which blew my mind, which is literally just take your photo, give it to Gemini, and say, use Nanobanana to overlay
06:30the the things onto the menu, and Nanobanana basically returned an image that is exactly the picture of the menu that I took, but it actually put into the pixels, it rendered the different things in the menu. And this blew my mind because all of my menu gen is spurious. It's working in the old paradigm.
06:45That app shouldn't exist. Yeah. The software three point zero paradigm is a lot more kind of raw.
06:50And so looking at software through the lens of software one point o, two point o, three point o, I mean, Kapathi explained what his app was, which you actually have to download from the App Store, and then you take a photo of the menu. It uses Nano Banana as probably an MCP connection, and it's going to render the images inside of that app for you.
07:08But for software one point o, which was following explicit rules, you'd literally have to have a database of tens of thousands of images pretaken and then match them for the menu.
07:18When it says eggs and toast, pull the photo that says eggs and toast. If you don't already have a photo of eggs and toast, you can't even render it. So that moved on to the images actually being rendered in real time through an app that you downloaded from the App Store that Kapathi created.
07:33However, in the software three point o world, you no longer even need to download the app. As long as you've got Gemini installed, which of course has its own connection to Nano Banana, you can literally prompt it in English.
07:44Say, hey. Add photos to this menu for me, and it's already created, making all these software two point o apps obsolete, which this is coming for almost every single app out there.
07:56And so you might be like me thinking, okay. That is software three point o, but how do we get a piece of the pie? How do we actually build businesses around this if someone like Gemini, ChatGPT, or Claude can just do it all without our help?
08:08Well, we'll get to that in just a second. Before that, I actually wanna show you what Andre Kapathi's evolution of business would look like in both a course seller setting, someone who sells specified knowledge, or someone who offers a service such as content creation or video editing. So let's start with the course.
08:23In software or business one point o, you may know Udemy, the platform. You basically go to Udemy, you buy a course, and you take the lectures in order. Day one, day two, day three, day four, day five, or lecture one, two, three, four, five, all the way through to the end.
08:36And at the end of that, you have specified knowledge. Okay. That's cool.
08:40But you're literally following that list. You're following that rules in sequential order. Knowledge two point o looks more like the YouTube little wavy function at the bottom of a video where you can see what areas of this video are most played and most favorited.
08:54And also you can scroll down to the comments and see which questions are asked most frequently. And then if you are a course seller, you may use some AI to cut out any of the bits that people are not watching, redesign your course based on, uh, whatever the questions are below it or even the order of operations that you're seeing a lot of your students go through.
09:12And in the course selling world, a lot of people haven't even caught up to education two point o yet, let alone being anywhere near education three point o. But education three point o looks extremely different once again. The best example I've got here is once again, you are gonna get a personalized agent.
09:29And the example I'm using is imagine if Max Verstappen got into the f one car with you, was in your headset, told you when to turn exactly as the corner was coming, to when to brake, when to accelerate, what to think, mindset, everything as you're taking action. Remember, all these courses, all these videos, all this knowledge work you're selling is supposed to lead to an actual outcome.
09:50I saw Alex Tomosi who has his own acquisition.comllm where you can query anything against his entire brain and against his entire knowledge base, which will spit out an answer talking to the CEO of Replit. And he was saying the next step of this is not a text based, hey, Alex Mosey, what would you do in this situation?
10:07It is go and build it with me. Okay. I'm building my landing page.
10:10Alex Mosey little AI pops up and he's like, cool. I'm throwing in this heading for this reason. I'm throwing in a video like this for this reason.
10:17I like this color scheme. Ohomosey AI who is guiding you through as you're actually taking action and being the little angel on your shoulder, showing you the right steps that he would personally take is the business and software three point o version of online learning and education. And so where does that lead with people who offer services?
10:34For this example, we're gonna be talking about video editing. Software and business one point o, which you should be getting good at recognizing where this is going to go and applying it to your own niche. Premiere Pro, basically a set of rules.
10:45If there is a space, we need to cut out the space. If there's an intro, we need a whooshing sound or a transition here, transition there. But everyone is manually doing it.
10:52It's an editor. You can hire an editor. They can follow that set of rules for you.
10:55But everyone has a set of rules and they're manually executing on them. Moving over to business two point o in the services video editing world. This is the app that I shall use called Descript.
11:04Now Descript is AI infused. You still have to use a tool.
11:09You still open the Descript app and you operate it like a normal human with your hands, but it has pattern recognition based on what most people don't want. For example, it can automatically identify the mistakes you make when recording or even the silences and then trim them with a click of a button.
11:25But services three point o is even different to that. It is an agent who is trained on all those popular video editing styles. And all you have is a text box where you say, hey, can you please edit this video in a viral mister b style and have it no longer than eight minutes?
11:39And then within five to ten minutes, once it renders it, it is completely done for you. So as you can see, software and business three point o is selling the outcome. And it doesn't have to be, oh, I want this edited like mister beast.
11:50It might be, I want this edited just like my own last video or my last 100 videos. And the AI agent knows exactly how to do that based on the neural network and pattern recognition. Now to answer the question, where are the actual moats where we can participate in this future economy?
12:06Because a lot of the large LLMs are gonna offer these services, and are we even needed anymore. Well, you're gonna have four moats left. You're gonna have the moat of your own personalized knowledge.
12:16Like I said, Alex Mosie created his own LLM trained on all of his data. So your data is going to be one of the first moats. If you can have an agent that edits in the style of mister beast and you have all that data and you have that pattern recognition, your value is in that knowledge and in that training data and have him do a good job so you're not just creating AI slop.
12:35Because just like Andre Kapathi said at the start, the second piece of that of getting a good outcome is not just the knowledge, but it's the set of instructions and the context that you're giving it. And so you need to become really good at prompt engineering. And what Andre Kapathi keeps saying and even Boris Cherni keeps saying as I study them more and more is that all businesses are literally being restructured to be AI first.
12:57So all businesses need to have AI in the middle of them being the engine and driving them forward. But that is all the engine is. You get to design your own car.
13:05You get to design your own wheels. You get to design the entire system around that engine, which leads us to the third thing. What system are you designing around that engine?
13:14For example, Higgs Field, which is a massive company right now, an AI company, basically just do what Gemini and Chattypitty can do, plugging into Nano Banana and generating videos and content for users. But they are doing it through an interesting user interface that helps people interact and get the most out of these tools.
13:32They've designed that system perfectly around that Nano Banana engine. And lastly, the fourth moat that you're going to have is trust from your audience. So some random on Twitter was to say, hey, I've trained this LLM on all of Elon's knowledge, go and use it now, versus Elon Musk to say, hey, I've trained this LLM based on all of my knowledge.
13:48The trust that you have in Elon Musk, the trust that you have in his access to his own data is the most valuable thing in the future that we're heading towards. If you made it this far in the video, please hit the subscribe button below. I'm excited to build some modern businesses with you.
14:01Thanks for watching. I'll see in the next
The Hook
The bait, then the rug-pull.
The last decade had a simple playbook: build a course, run an agency, ship an app. Then Andrej Karpathy walked into a Sequoia Capital interview and quietly declared all of it obsolete. Dream Labs AI's host spent 14 minutes translating what that actually means for creators and solo builders — and landed on four moats that still hold.
Named ideas worth stealing.
02:00model
Software 1.0 / 2.0 / 3.0
- 1.0: Explicit rules (write every step)
- 2.0: Learned weights (machine learning feedback loops)
- 3.0: LLM-as-interpreter (prompting replaces programming)
Karpathy's three-era model for how software and all digital businesses work. 3.0 means the context window is your code.
Steal forAny explainer about why old-model businesses are failing; frames the argument without sounding alarmist
12:01list
Four Moats of Software 3.0
- Knowledge (proprietary data)
- Instructions (prompt engineering)
- AI Engine (leverage the LLM)
- Audience / System (trust + UX)
The four defensible positions for creators and businesses when LLMs commoditize the execution layer.
Steal forWorkshop slide, carousel, short-form hook, newsletter framework. Four nodes around a center AI engine visual.
How they asked for the click.
VERBAL ASK
13:50subscribe
“If you made it this far in the video, please hit the subscribe button below.”
Clean sign-off with on-screen subscribe animation. No product pitch, no upsell.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
