Marc Andreessen 与 Amjad Masad 谈“够好”的 AI、AGI 与编程终结

We’re dealing with magic here that I think probably all of us would have thought was impossible 5 years ago, or certainly 10 years ago. This is the most amazing technology ever, and it’s moving really fast, and yet we’re still really disappointed. It’s not moving fast enough, and it’s maybe right on the verge of stalling out.

We should both be hyper-excited but also on the verge of slitting our wrists because the gravy train is coming to an end.

Right?

It is faster, but it’s not at computer speed, right? What we expect computer speed to be. It’s sort of like watching a person work.

It’s like watching John Carmack—

The world’s best programmer on a stimulant.

On a stimulant. Yeah, that’s right.

So, let’s start with this: Let’s assume that I’m a novice programmer. Maybe I’m a student, or maybe I’m somebody who took a few coding classes and hacked around a little bit. I might do Excel macros or something like that, but I’m not a master craftsman at coding. Somebody tells me about Replit, and specifically AI in Replit. What’s my experience when I launch into what Replit is today, with AI?

Yeah, I think the experience of someone with no coding experience or some coding experience is largely the same when you go into Replit.

The first thing we try to do is get all the nonsense away from setting up a development environment and all of that stuff, and just have you focus on your idea. What do you want to build? Do you want to build a product? Do you want to solve a problem? Do you want to do a data visualization? The prompt box is really open for you. You can put anything there.

Let’s say you want to build a startup. You have an idea for a startup. I would start with a paragraph-long description of what I want to build. The agents will read that. It will—

You just type—just type—

Standard English. You just type it in: “I want to sell crepes online.”

You can— it literally could be 4 or 5 words.

Or, if you have a programming language or stack you prefer, you could specify that. But we actually prefer that you don’t do that, because we’re going to pick the best stack for your request. We’re going to classify the best stack for it.

It’s a—

If it’s a data app, we’ll pick Python and Streamlit or whatever. If it’s a web app, we’ll pick JavaScript and Postgres and things like that. You just type that—

Or you can decide. You can say, “I know Python,” or, “I’m learning Python in school, and I want to do it in Python.”

That’s right. The cool thing about Replit is that we’ve been around for almost 10 years now, and we built all this infrastructure. Replit runs any programming language, so if you’re comfortable with Python, you can go in and do that for sure.

Okay.

I know this is obvious to people who have used it, but I’m dealing in English.

Yes.

Okay, go ahead.

Yes, you’re fully in English. A little bit of background here: When I came here and pitched to you 10 years ago—or whatever, 7 years ago—

Right?

What we were saying is, we were exactly describing this future: Everyone would want to build software, right? The thing that gets in people’s way is all the accidental complexity of programming, as Fred Brooks called it.

There’s the essential complexity, which is, “How do I bring my startup to market? How do I build a business?” And then there’s all that accidental complexity: What package manager do I use? We’ve been abstracting all of that away for so many years so you can just—

The last thing we had to abstract away is code.

Right.

I had this realization last year, which is: I think we built an amazing platform, but the business is not performing. The reason the business is not performing is that code is the bottleneck. Yes, all the other stuff is important to solve, but syntax is still an issue. Syntax is just an unnatural thing for people. Ultimately, English is the programming language.

Right. Does it work with other world languages, other than English, at this point?

Yes, you can write in Japanese, and we have a lot of users, especially Japanese users. That tends to be very—

So, does it support every language these days? Or do you still have to do custom work to craft a new language?

No. Most mainstream languages that have 100 million-plus people who speak them—AI is pretty good at them.

Okay. Yeah.

Yeah.

Wow.

I did a bit of historical research recently, for some reason. I just wanted to understand the moment we’re in, because it’s such a special moment. I think it’s important to contextualize it.

I read this quote from Grace Hopper. Grace Hopper invented the compiler, as you know. At the time, people were programming in machine code, and that’s what programmers did—that’s what the specialists did.

Yes.

She said specialists will always be specialists. They have to learn the underlying machinery of computers, but she wanted to get to a world where people were programming in English. That’s what she said. That’s before Karpathy, right? That’s 75 years ago.

She invented the compiler, and in her mind, programming in C is English.

Right.

But that really didn’t do it. That was just the start of it. You had C, and then you went to higher-level languages like Python and JavaScript. I think we’re at a moment where it’s the next step, right? Instead of typing syntax, you’re actually typing thoughts, which is what we ultimately want.

And the machine writes the code.

And the machine writes the code.

Right.

Yeah, I remember it. You’re probably not old enough to remember, but I remember when I was a kid. There were higher-level languages by the ’70s, like BASIC and Fortran and C. But you would still run into people who were doing assembly programming—assembly language—which, by the way, you still do. Game companies and others still use assembly. The assembly people were hating on the kids who were doing BASIC, but there were also older coders who hated on the assembly programmers for doing assembly instead of direct machine code—direct zeroes and ones. Assembly language is a very low-level programming language that compiles to actual machine code, and it’s incomprehensible gibberish to most programmers.

You’re writing in octal or something.

You’re writing very close to the hardware. But even still, it’s a language that compiles to zeroes and ones. The actual real programmers wrote in zeroes and ones.

There’s always this tendency for the pros to look down their noses and say the new people are being sloppy. They don’t understand what’s happening. They don’t really understand the machine.

Yeah.

And then, of course, what the higher-level abstractions do is democratize programming.

The absolute irony is that I was part of the JavaScript revolution. I was at Facebook before starting Replit, and we built the modern JavaScript stack. We built ReactJS and all the tooling around it.

We got a lot of hate from programmers saying that you should type vanilla JavaScript directly.

Yeah.

I was like, “Okay, whatever.” And now that’s mainstream. The people who built their careers on the last wave we invented are hating on this new wave. People never change.

Okay, got it. You’re typing English: “I want to sell crepes online. I want to do this. I want to have a T-shirt.” Whatever the business is. What happens then?

Then the Replit agent will show you what it understood. It’s trying to build a common understanding between you and it, and I think there are a lot of things we can do better there in terms of the UI. But for now, it’ll show you a list of tasks.

It’ll tell you, “I’m going to set up a database because you need to store your data somewhere. We need to set up Shopify or Stripe because we need to accept payments.” Then it shows you this list and gives you 2 options initially: Do you want to start with a design so that we can iterate back and forth to get that design locked down, or do you want to build the full thing?

If you want to build the full thing, we’ll go for 20, 30, 40 minutes. The agent will tell you, “Go here, install the app. I’m going to set up the database, do the migrations, write the SQL, build the site, and also test it.”

This is a recent innovation we did with Agent 3: after it writes the software, it spins up a browser, goes around, and tests in the browser. If it finds an issue, it iterates and fixes the code. It’ll spend 20 or 30 minutes building that. I’ll send you a notification telling you the app is ready, and you can test it on your phone.

You can go back to your computer. Maybe you’ll find a bug or an issue, and you’ll describe it to the agent: “Hey, it’s not exactly doing what I expected.” If it’s perfect and you’re ready to go, that’s it—20 minutes. By the way, there are a lot of examples where people just get their idea in 20 or 30 minutes, which is amazing. You just hit publish.

Mhm.

You hit publish. In a couple of clicks, you’ll be up in the cloud. We’ll set up a virtual machine in the cloud, the database is deployed, everything’s done, and now you have a production database.

So, think about the steps needed just 2 or 3 years ago to get to that step. You had to set up your local development environment. You had to sign up for an AWS account. You had to provision the databases and the virtual machines. You had to create the entire CI/CD deployment pipeline. All of that is done for you, and a kid can do it. A layperson can do it.

If you’re a programmer and you’re curious about what the agent did, the cool thing about Replit—because we have this history of being an IDE—is that you can peel back the layers. You can open the file tree and look at the files. You can open Git, push it to GitHub, connect it to your editor if you want, or open it in Emacs. So the cool thing about Replit is, yes, it is a vibe-coding platform that abstracts away all the complexities, but all the layers are there for you to look at.

Right. So let’s go back to that. You said it gives you—the agent gives you—you say, “I’ve got my idea.” You plug it in, and it gives you this list of things. Then, when you describe it, you said, “I’m going to do this. I’m going to do that.” The “I” there, in that case, was the agent as opposed to the user. Yes.

And so the agent lists the set of things that it’s going to do, and then the agent actually does those things.

The agent does those things. Yeah.

The agent does those things. Yeah. That’s a very important point. When we made this shift, we hadn’t realized internally at Replit how much the actual user stopped being the human user and was actually the agent programmer.

Right?

One really funny thing happened: we had servers in Asia. The reason we had servers in Asia was because we wanted our Indian or Japanese users to have a shorter time to the servers. When we launched the agent, their experience got significantly worse. We were like, “What happened? It’s supposed to be faster.” Well, it turns out it was worse because the AIs were sitting in the United States.

The programmer is actually in the United States. You’re sending the request to the programmer, and the programmer is interfacing with a machine across the world. So, yes, suddenly the agent is the programmer.

The new terminology is that an agent is a software program that is basically using the rest of Replit as if it were a human user, but it’s not. It’s a bot.

That’s right. It has access to tools such as writing a file, editing a file, deleting a file, searching the package index, installing a package, provisioning a database, and provisioning object storage. It is a programmer that has the tools and the interface. It has an interface that is very similar to a human programmer.

We’ll talk more about how this all works, but a debate inside the AI industry is about this idea of having agents that do things on your behalf and go out and accomplish missions. Obviously, it’s a big deal even to have an AI agent that can do relatively simple things. To do complex things, of course, is one of the great technical challenges of the last 80 years.

Then there’s this question of how long the agent can go out and run and operate on its own: 5 minutes, 15 minutes, an hour, or 8 hours. How long does it maintain coherence? How long does it actually stay in full control of its faculties and not spin out?

At least the early agents, if you set them off to do this, might be able to run for 2 or 3 minutes. Then they would start to get confused, go down rabbit holes, and spin out. More recently, we’ve seen that agents can run a lot longer and do more complex tasks. Where are we on the curve of agents being able to run for how long and for what complexity of tasks before they break?

That’s absolutely the main metric we’re looking at. Even back in 2023, we had the idea for software agents—we had the idea 4 or 5 years ago, actually. The problem every time we attempted them was coherence. They would go on for a minute or 2, and then they would compound errors in a way that they just couldn’t recover.

Um—

You can actually see it, right? If you watch them operate, they get increasingly confused and then maybe even deranged. They veer into very weird areas, and sometimes they start speaking Chinese and doing really weird things. I would say sometime around last year, we maybe crossed the 3-, 4-, 5-minute mark.

And it felt to us that, okay, we’re on a path where long-horizon reasoning is getting solved.

And so we made a bet, and I tell my team—

So, long-horizon reasoning means reasoning—dealing in facts and logic—

—in a sort of complex way, and then long horizon being over a long period of time. Yes.

With many, many steps to a reasoning process.

Yeah, that’s right. So if you think about the way large language models work, they have a context. This context is basically the memory—all the text, all your prompts, and all the internal talk that the AI is doing as it’s reasoning.

When the AI is reasoning, it’s actually talking to itself. It’s like, “Now I need to go set up a database. What kind of tool do I have? There’s a tool here that says PostgreSQL. Okay, let me try using that.” Then it gets feedback, looks at the feedback, and reads it.

That prompt box, or context, is where the user input, the environment input, and the internal thoughts of the machine all reside. It’s sort of like a program’s in-memory space. Reasoning over that was the challenge for a long time. That’s when AIs would just go off track, and now they’re able to think through this entire thing and maintain coherence.

There are now techniques around context compression. Context length is still a problem, right? LLMs today are marketed as having a million-token context length, which is almost a million words. In reality, it’s about 200,000 tokens, and then they start to struggle.

We do a lot of stopping and compressing of the memory. If a portion of the memory says that I’m getting all the logs from the database, you can summarize paragraphs of logs with 1 statement. Or you can summarize the database setup: “That’s it.” Every once in a while, we’ll compress the context so that we make sure we maintain coherence. There’s been a lot of innovation outside of the foundation models as well in order to enable that long-context coherence.

So what was the key technical breakthrough in the foundation models that made this possible, do you think?

I think it’s RL. I think it’s reinforcement learning. The way pretraining works is that pretraining is the first step of training a large language model. It reads a piece of text, covers the last words, and tries to guess them. That’s how it’s trained.

That doesn’t really imply long-context reasoning. It turns out to be very, very effective, and it can learn language that way. But the reason we weren’t able to move past that limitation was that this modality of training just wasn’t good enough. What you want is a type of problem-solving over a long context.

What reinforcement learning, especially from code execution, gave us is the ability for the LLM to roll out what we call trajectories in AI. A trajectory is a step-by-step reasoning chain in order to reach a solution.

As I understand it, reinforcement learning works by putting the LLM in a programming environment like Replit and saying, “Here’s a codebase. Here’s a bug in the codebase, and we want you to solve it.” The human trainer already knows what the solution would look like. We have a pull request on GitHub, so we know exactly what it should look like, or we have a unit test that we can run and use to verify the solution.

It rolls out a lot of different trajectories. They sample from the model, and maybe 1 of those trajectories will reach the solution. A lot of them will just go off track, but 1 of them will solve the bug and reach the solution. It reinforces that trajectory, so it gets a reward, and the model gets trained: “This is how you solve these types of problems.”

So that's how we're able to extend these reasoning chains.

Got it. How good are the models now at long reasoning? How do we know? How is that established?

There is a nonprofit called METR that has a benchmark to measure how long a model runs while maintaining coherence and doing useful things, whether it's programming or other benchmark tasks that they've done. They put out a paper, I think, late last year that said every 7 months, the number of minutes that a model can run is doubling.

So you go from 2 minutes to 4 minutes in 7 months. I think they vastly underestimated that. Is that right? It's doubling more often than 7 months.

With Agent 3, we measure that very closely, and we measure that in real tasks from real users. We're not doing benchmarking; we're actually doing A/B tests and looking at the data on how users are successful or not.

For us, the absolute sign of success is that you made an app and published it. When you publish it, you're paying extra money. You're saying, “This app is economically useful. I'm going to publish it.” So that's as clear-cut as possible.

What we're seeing is that in Agent 1, the agent could run for 2 minutes and then perhaps struggle. Agent 2 came out in February; it ran for 20 minutes. Agent 3 runs for 200 minutes.

Okay.

Two hundred. Some users are pushing it to 12 hours and things like that. I'm less confident that it is as good when it goes to these stratospheres, but at a 2- to 3-hour timeline, it is really—it's insanely good.

The main innovation outside of the models is a verification loop. I remember reading a research paper from Nvidia. What Nvidia did is, they were trying to write GPU kernels using DeepSeek, and that was perhaps 7 months ago, when DeepSeek came out. What they found is that if we add a verifier in the loop—if we can run the kernel and verify it's working—we're able to run DeepSeek for 20 minutes, and it was generating optimized kernels.

So I was like, okay, the next thing for us, obviously, as a sort of agent lab or applied AI company—we're not doing the foundation model stuff, but we're doing a lot of research on top of that. We know that agents can run for 10 or 20 minutes now, or LLMs can stay coherent for longer, but for you to push them to 200 or 300 minutes, you need a verifier in the loop.

That's why we spend all our time creating scaffolds to make it so that the agent can spin up a browser and do computer-use-style testing. Once you put that in the middle, what's happening is it works for 20 minutes, spins up another agent, spins up a browser, and tests the work of the previous agent. So it's a multi-agent system.

If it finds a bug, it starts a new trajectory and says, “Okay, good work. Let's summarize what you did the last 20 minutes.” Now that, plus the bug that we found, is a prompt for a new trajectory. You stack those on each other and you can go endlessly, but—

So it's like setting up a marathon or a relay race. As long as each step is done properly, you could do an infinite number of steps.

That's right. You can always compress the previous step into a paragraph, and that becomes a prompt. So it's an agent prompting the next agent.

Right. Right. Right. That's amazing. And then, when an agent—when a modern agent running on a modern LLM that's trained this way—let's say it runs for 200 minutes, when you watch the agent run, is it processing through logic and tasks at the same pace as a human being, or slower or faster?

It's actually, I would say, faster, but not that much significantly faster. It's not at computer speed, right? It's not what we expect computer speed to be.

It's like watching a person work.

It's like watching John Carmack work.

The world's best programmer.

Yeah.

The world's best programmer on a stimulant.

On a stimulant. Yeah, that's right.

Working for you.

Working for you. Yeah. There. So, it's very fast, and you can see the file diffs running through, but every once in a while it'll stop and start thinking. It'll show you the reasoning: “I did this and I did this. Am I on the right track?” It really tries to reflect.

Then it might review its work and decide the next step, or it might kick into the testing agent. You're seeing it do all of that, and every once in a while it calls a tool. For example, it stops and says, “We ran into an issue. PostgreSQL 15 is not compatible with this database ORM package that I have.”

“Okay, this is a problem I haven't seen before. I'm going to go search the web.” So it has a web search tool. It looks like a human programmer, right? It's really fascinating to watch. It's one of my favorite things to do: just watch the tool chain, reasoning chain, and testing chain. It is like watching a hyperproductive programmer.

Right. So we're kind of getting into the holy grail of AI here, which is generalized reasoning by the machine. You mentioned this a couple of times, but this idea of verification—so, just for folks listening to the podcast who maybe aren't in the details, let me try to describe this and see if I have it right.

A just-large language model, the way you would have experienced ChatGPT out of the gate 2 years ago or whatever, would have been incredible at language. It's incredible how good it is at writing Shakespearean sonnets or rap lyrics. It's amazing at human conversation. But if you start to ask it problems that involve rational thinking or problem-solving, all of a sudden, math is the whole show. In the very beginning, if you asked it very basic math problems, it would not be able to do them.

That's right.

But then even when it got better at those, if you started to ask it, it could maybe add 2 small numbers together, but it couldn't add 2 large numbers together. Or if it could add 2 large numbers, it couldn't multiply them. It was just like, “All right, this is…”

And then there was the famous strawberry test: How many Rs are in the word “strawberry”?

That's right.

There was this long period where it kept guessing wrong. It would say there were only 2 Rs in the word “strawberry,” and it turns out there are 3. So it was this thing, and people were using this term—the slur at the time was “stochastic parrot.”

Yeah.

I was thinking “clanker.”

Well, clanker is the new slur. Clanker is just the full-on racial slur against AI as a species. But the technical critique was the so-called “stochastic parrot.” “Stochastic” means random, so, sort of, random parrot—meaning basically that large language models were like a mirage. They were repeating back to you things that they thought you wanted to hear.

In a way, it's true in the pure pre-training LLM world.

Right, for the very basic layer. But then what happened over the last year or so was this layering in of reinforcement learning. The key to—

It's not new, crucially. It's like AlphaGo, right? So—

Describe that for a second.

Yeah. So we had this breakthrough before, in 2015 or 2016—the AlphaGo breakthrough, I think—where there was a merging of the old AI debate between the connectionists, the people who think neural networks are the true way of doing AI, and the symbolic systems, or the people who think that discrete reasoning, facts, and knowledge bases are the way to go.

There was a merging of these 2 worlds. The way AlphaGo worked is that it had a neural network, but it had a Monte Carlo tree search algorithm on top of that. The neural network would generate a list of potential moves, and then you had a more discrete algorithm sort those moves and find the best based on tree search—based on just trying to verify, again, this sort of verifier in the loop, trying to verify which move might yield the best based on a more classical way of doing algorithms.

So that's a resurgence of that movement, where we have this amazing generative neural network that is the LLM, and now let's layer on more discrete ways of trying to verify whether it's doing the right thing or not. Let's put that in a training loop, and once you do that, the LLM will start gaining new capabilities, such as reasoning over math and code and things like that.

Exactly. Right. Okay. And then the key thing there, though, for RL to work—for LLMs to reason—the key is that it be a problem statement for which there is a defined and verifiable answer.

That’s right. Is that right? You might think about this as giving a bunch of examples. In medicine, this might be a diagnosis that a panel of human doctors agrees with, or, by the way, a diagnosis that actually solves the condition. In law, this would be an argument that, in front of a jury, actually results in an acquittal or something like that.

In math, it’s an equation that actually solves properly. In physics, it’s a result that actually works in the real world.

I don’t know—in civil engineering, it’s a bridge that doesn’t collapse.

Right. There’s always some test. The first 2 do not work very well just yet. Law and healthcare are still a little too squishy, a little too soft. It’s unlike math or code. The way they’re training on math, they’re using a sort of programming language—a provable language called Lean—for proofs. You can run a Lean statement, you can run computer code, and perhaps you can run a physics simulation or a civil engineering physics simulation, but you can’t run a diagnosis.

But you could verify it with human answers, or not.

Yeah. So that’s more RLHF, in a way. It’s not the fully autonomous, fully scalable autonomous RL training, which is why coding is moving faster than any other domain: We can generate these problems and verify them on the fly.

But with coding, as anybody who’s coded knows, there are 2 tests. One is, does the code compile?

Right.

And then the other is, does it produce the right output? Just because it compiles doesn’t mean it produces the right output. Verifying that it’s the correct output is harder.

Yeah, SWE-bench is a collection of verified pull-request end states, so it’s not just about compiling. SWE-bench is the main benchmark used to test whether AI is good at software engineering tasks, and we’re almost saturating it. Last year we were at maybe 5% in early 2024 or less, and now we’re at 82% or something like that, with Claude Sonnet 4.5 at state of the art. That’s a really nice hill climb happening right now. They went and looked on GitHub, found the most complex repositories, found bug statements that are very clear, and found pull requests that actually solve those bug statements with unit tests and everything. So there’s an existing corpus on GitHub of tasks that AIs can solve, and you can also generate them. Those aren’t too hard to generate; it’s called synthetic data. But you’re right, it’s not infinitely scalable, because some human verifiers still need to look at the task. Maybe the foundation models have found a way to have the synthetic training go all the way.

Right. And then what’s happening, I think, is that the foundation model companies, in some cases, are actually hiring human experts to generate new training data.

Yes. So they’re actually hiring mathematicians, physicists, and coders to basically sit and work. They’re hiring human programmers, putting them on the clock.

And having them actually write code in a way where there’s a known result of the code running, such that the RL loop can be trained properly.

That’s right. These companies are also building systems where the software itself generates the training data, generates the tests, and generates the validated results. That’s all synthetic training data.

That’s right. But again, those work in the very hard domains. It works to some extent in the software domains.

And I think there’s some transfer learning. You can see the reasoning work when it comes to tools like Deep Research and things like that, but we’re not making as much progress in the softer domains.

So, softer domains meaning domains in which it’s harder, or even impossible, to actually verify the correctness of a result in a deterministic, factually grounded, noncontroversial way?

If you have a chronic disease, you could have POTS or EDS, or whatever. They’re all clusters, and it’s because it is the domain of abstraction. It’s not as concrete as code and math and things like that. So I think there’s still a long way to go there.

Right. So the more concrete the problem, the concreteness of the problem is the key variable, not the difficulty of the problem. Would that be a way to think about it?

Yeah. I think the concreteness, in the sense of: Can you get a true-or-false, verifiable answer?

Right. But in any domain of human effort in which there’s a verifiable answer, we should expect extremely rapid progress.

Yes.

Right.

Yes, absolutely. I think that’s what we’re saying.

Right. And that for sure includes math. That for sure includes physics. It for sure includes chemistry. For sure includes—

Large areas of code.

That’s right. What else does that include, do you think?

Biology, like what we’re seeing with proteins?

Genomics.

Yeah. Things like that. I think some areas of robotics, right? There’s a clear outcome, right?

But it’s not that many.

I mean, surprisingly—

Well, it depends.

Yeah, it depends on your point of view. Some people might say that’s a lot. You mentioned the pace of improvement. What would you expect from the pace of improvement going forward?

I think we’re ripping on coding. I think it’s just going. What we’re working on with Replit Agent 3 right now is that, by next year, we think you’re going to be sitting instead of typing in front of Replit, and you’ll be shooting off multiple agents at a time.

You’re planning a new feature: “I want a social network on top of my storefront.” And another one is, “Hey, refactor the database.” You’re running parallel agents, so you have 5 or 10 agents working in the background, merging the code and taking care of all of that.

You also have a really nice interface on top of that, where you’re doing design and interacting with AI in a more creative way, maybe using visuals and charts and things like that. So there’s a multimodal angle to that interaction.

I think creating software is going to be such an exciting area. I think the layperson will be as good as a senior software engineer who works at Google is today. I think that’s happening very soon.

There’s this weird dynamic—see if you agree with this, and Erik, I’m also curious about your point of view on this. We have this in the office here a lot, and I also have this with leading entrepreneurs a lot, which is this thing of—

Like, wow, this is the most amazing technology ever, and it’s moving really fast, and yet we’re still really disappointed. It’s not moving fast enough, and it’s maybe right on the verge of stalling out.

And we should both be hyper-excited, but also on the verge of slitting our wrists because the gravy train is coming to an end, right?

And I always wonder: On the one hand, not all ladders go to the moon. Just because something looks like it works doesn’t mean you’re going to be able to scale it up and have it work to the fullest extent. It’s important to recognize practical limits and not just extrapolate everything to infinity.

On the other hand, we’re dealing with magic here that I think we all would have thought was impossible 5 years ago, or certainly 10 years ago.

I didn’t think I would live to see any of this. I got my computer science degree in the late ’80s and early ’90s. This is just amazing that this is actually happening in my lifetime.

But there’s a huge bet on AGI, right? Whether it’s the foundation models, I think the entire US economy is sort of a bet on AGI. There are crucial questions to ask: Are we on track to AGI or not?

There are some ways that I can tell you it doesn’t seem like we’re on track to AGI, because there doesn’t seem to be transfer learning across these domains that are significant. If we get a lot better at code, we’re not immediately getting better at generalized reasoning. We need to also get training data and create RL environments for biology, chemistry, physics, math, and law.

This has been the point of discussion in the AI community after the Dwarkesh Patel and Richard Sutton interview, where Richard Sutton poured cold water on the Bitter Lesson.

So everyone was using this essay that he wrote called “The Bitter Lesson.” The idea is that there are infinitely scalable ways of doing AI research, and anytime you can pour in more compute and more data and get more performance out, that’s the ultimate way of getting to AGI. Some people interpreted that interview to mean that perhaps he’s doubtful that we’re even on a Bitter Lesson path here, and perhaps the current training regime is actually very much the opposite, in which we’re so dependent on human data, human annotation, and all of that stuff.

So I think—I agree with you. As a company, we’re excited about where things are headed, but there’s a question of whether we’re on track to AGI or not. I’d be curious what you think.

And, you know, Ilia Sutskever makes a specific form of this argument, which is basically that we’re just literally running out of training data. It’s a fossil fuel argument, right? If we’ve fundamentally slurped all the training data—if we’ve slurped all the data off the internet—that’s where almost all the data is at this point. There’s a little bit more data in, like, a private dark pool somewhere that we’re going to go get, but we have it all.

And then we’re in this business now trying to generate new data, but generating new data is hard and expensive compared to just slurping things off the internet.

There are these arguments. Having said that, you get into definitional questions really quickly, which are kind of a rabbit hole. But having said that, you mentioned transfer learning. Transfer learning is the ability of a machine to be an expert in one domain and then generalize that into another domain.

My answer to that is: Have you met people? How many people do you know who are able to do transfer learning?

Not many. Right. Well, because there’s—

Quite the opposite, actually. The nerdier they are in a certain domain, the more often they have blind spots. We joke about how everyone is just full of shit in one area, or they make some massive mistake. You don’t trust them on this, but on this other topic, you know—

Right? Yeah. Well, this is a well-known thing among, for example, public intellectuals. There have actually been whole books written about so-called public intellectuals. You get these people who show up on TV and they’re experts, and what happens is that they’re an expert in economics, then they show up on TV and talk about politics, and they don’t know anything about politics. They don’t know anything about medicine, the law, or computers.

This is Paul Krugman talking about how the internet was going to be no more significant than the fax machine.

Facts.

He’s a brilliant economist. He has no idea how a computer works.

Is he a brilliant economist?

Well, at one point. Even if he’s brilliant, this is the thing: What does that mean? Should a brilliant economist be able to extrapolate that the internet is significant? That’s a good question. But take anybody—or, by the way, Einstein is actually my favorite example. I think you’d agree Einstein was a brilliant physicist.

He was a Stalinist. He was a socialist, and he was a Stalinist. He thought Stalin was fantastic.

Oh, still?

Yeah.

Okay. All right.

True socialism.

All right, all right. Einstein—I’ll take your word for it. But once he got into politics, he was just totally loopy. He sounded like an undergraduate lunatic, like somebody in a dorm room. There was no transfer learning from physics into politics. Whether he was right or wrong, there was clearly nothing new in his political analysis. It was the same rote routine you get out of—

So, in a way, the argument you’re making is that maybe we already have human-level AI. Perhaps the definition of AGI is something totally different—something above the human level that totally generalizes across domains. It’s not something that we’ve seen.

Yeah. We’ve idealized a goal that may be idealized in a way that, number one, is just so far beyond what people can do that it’s no longer a relevant comparison to people. Usually AGI is defined as being able to do everything better than a person can.

And it’s like, well, okay: If doing everything better than a person can means that a person can’t do any transfer learning at all—

Right. Doing even a little bit, a marginal bit, might actually be better. Or it might not matter, just because no human can do it. Therefore, you just stack up the domains.

There’s also this well-known phenomenon in AI. AI engineers and scientists always complain about it: The definition of AI is always the next thing that the machine can’t do. The definition of AI for a long time was, “Can it beat humans at chess?”

And then the minute it could beat humans at chess, that was no longer AI. That was just boring.

That’s computer chess. It became—

Computer chess. It’s just boring, and now it’s an app on your iPhone. Nobody cares.

The Turing test was the test, and then we passed it and nobody—

We blew through it. This was a really big deal.

There was no celebration.

There were no parties.

That’s exactly right. For 80 years, the Turing test was the thing. They made a movie about it. That was the whole thing, and we blew right through it and nobody even registered it. Nobody cares. It gets no credit. We’re just like, “Ah, it’s still a complete piece of shit.”

Right. The AI scientists are used to complaining that they’re always being judged against the next thing, as opposed to all the things they’ve already solved.

But that’s maybe the other side of it. They’re also putting out for themselves an unreasonable goal, and then doing this sort of self-flagellation along the way. I wonder which way that cuts.

Yeah, it’s an interesting question. I started thinking about this idea that it doesn’t matter whether it’s truly AGI. The way I define AGI is that you put an AI system in any environment and it efficiently learns. It doesn’t have to have that much prior knowledge in order to learn something, but it can also transfer that knowledge across different domains.

We can get to functional AGI. Functional AGI is just collecting data on every useful economic activity in the world today and training an LLM—or training the same foundation model—on top of that. We’ll target every sector of the economy, and you can automate a big part of labor that way. I think we’re on that track for sure.

Right. You tweeted after GPT-5 came out that you were feeling the diminishing returns. What were you expecting, and what needs to be done? Do we need another breakthrough to get back to the pace of growth?

This whole discussion is sort of about that. My feeling is that GPT-5 got good at verifiable domains. It didn’t feel that much better at anything else—the more human angle of it. It felt like it regressed, and you had this sort of Reddit pitchfork movement against Sam and OpenAI because people felt like they had lost a friend.

GPT-4 felt a lot more human and closer, whereas GPT-5 felt a lot more robotic—very in its head, trying to think through everything. I would have expected that when we went from GPT-2 to GPT-3, it was clearly getting a lot more human. It was a lot closer to our experience. You could feel that it understood the world better.

Similarly, going from GPT-3 to GPT-4 to GPT-5 didn’t feel like it was a better overall being, as it were.

Is that partly emotionality?

I like to ask models about very controversial things. Can they reason through—I don’t know how deep we want to go here—but what happened with World Trade Center 7?

Right.

Sure, it’s an interesting question. I’m not putting out a theory, but it’s interesting: How did it happen, and can it think through controversial questions in the same way that it can think through a coding problem? There hasn’t been any movement there with all the reasoning and everything. And not just that—that’s a cute example—but COVID, right? The origins of COVID.

Go dig up GPT-4 or other models and compare them to GPT-5. You’re not going to find that much difference in terms of, “Okay, let’s reason together.”

Let's try to figure out what the origins of COVID were, because it's still an unanswered question, and I don't see them making progress on that. I mean, you play a lot with them. Do you feel like—

I use it differently. I don't know; maybe I have different expectations. My main use case is sort of a PhD in everything at my beck and call. I'm trying to get it to explain things to me more than I'm trying to have conversations with it. Maybe I'm just unusual with that. But—

And that gets back to the question.

Well, what I found specifically is a combination of GPT-5 Pro plus Deep Research, or Grok 4 Heavy—the highest-end models. They now basically generate 30- to 40-page, essentially book-length reports on demand on any topic.

Anytime I get curious about something, you just take it. Maybe it's my version of it, but here's a good example: When an advanced economy puts a tariff on a raw material or on a finished good, who pays? Is it the consumer, the importer, the exporter, or the producer? This is actually a very complicated question. It turns out to be a very complicated question. It's a big, big, big thing that economists study a lot, and it's just like, okay, who pays?

Well, it's outstanding at going out on the web, getting information, and synthesizing it.

Correct. It gives me a synthesized 20-, 30-, or 40-page PDF. It basically tops out at 40 pages.

I can get up to 40 pages of PDF, but it's completely coherent and, as far as I can tell from everything I've cross-checked, completely world-class. If I hired, for a question like that, a great economics postdoc at Stanford who just went out and did that work, it would maybe be that good.

Yeah. But then the significance is that, at least for many domains, you have a PhD in everything.

But this is synthesizing knowledge, not trying to create new knowledge.

Well, but this gets to the angels-dancing-on-the-head-of-a-pin thing, which is: What's the difference? How much new knowledge ever actually is there anyway? What do you actually expect from people when you ask them questions?

What I'm looking for is, yes, explain this to me in the clearest, most sophisticated, most complex, most complete way that it's possible for a real expert to explain things to me.

And that's what I use it for. Again, as far as I can tell from the cross-checking, I'm getting almost basically 100 out of 100. I don't even think I've had an issue in months where it's had a problem in it.

And it's like, yeah, you can say synthesizing isn't supposed to create new information, but it's generating a 40-page—it's basically generating a 40-page book.

That's amazing.

That's incredibly fluid. The logical coherence of the entire thing—it's great writing. If you evaluated a human author on it, you would say, “Wow, that's a great author.” Are people who write books creating new knowledge? Well, sort of not, because a lot of what they're doing is building on everything that came before them, synthesizing a body of knowledge. But also, a book is a creative accomplishment, right? And so—

Yeah. One of the things I'm interested in and hoping AI could help us solve is just how confusing the information ecosystem is right now. Everything feels like propaganda. It doesn't feel like you're getting real information from anywhere. I really want an AI that could help me reason from first principles about what's happening in the world, so I can actually get real information. Maybe that's an unreasonable ask of AI researchers, but I don't think we've made any progress there. So maybe I'm overfocused on arguing with people instead of trying to get at the underlying truth.

Well, here's the thing I do a lot with this: I say, “Take a provocative point of view, and then steelman the position.” I often pair these: steelman the position that it was a lab leak and steelman the position that it was natural origins.

And again, is this creativity or not? I don't know. But what comes back is 30 pages each of, “Wow, that is the most compelling case in the world I can imagine,” with everything marshaled against it and the argument structured in the most compelling way possible.

Part of the reason that started happening is that it stopped being taboo to talk about a lab origin. When it was taboo, the AIs would talk down to you, like, “Oh, you're a conspiracy theorist.”

Yes. So there was a period of time when you took something truly controversial, and they actually couldn't reason about it because of all the RLHF and all the limitations. As you know, I won't pick specific ones here, but there are certain big models that will still lecture you—

They'll tell you that you're a bad person for asking that question. But some of them are really, really open now to being able to do these things.

Ultimately, what you're looking for—the ultimate thing—would be something that I don't think anybody has really defined well. Again, all the conventional definitions of AGI are basically comparisons to people.

Yeah.

The conventional explanations of AGI have always struck me as a lot like the debate around whether a self-driving car works. Does a self-driving car work because it's a perfect driver, or does it work because it's better than the human driver? Being better than the human driver, I think, is actually quite a real thing, just like with chess and Go. I actually think that's a real thing.

Then there's, “Is it a perfect driver?”—which is obviously what the self-driving car companies are working toward. But I think you're looking for something beyond the perfect driver. You're looking for the car that knows where to go. So I'm of two minds: One mind is the practical entrepreneur, right?

I just have so many toys to play with and build. If we stopped AI progress today, Replit would continue to get better for the next 5 years. There's so much we can do just on the app layer and the infrastructure layer.

But I think the foundation models will continue to get better as well, so it's a very exciting time in our industry.

The other mind is more academic, because as a kid I've always been interested in the nature of consciousness and the nature of intelligence. I was always interested in AI and reading the literature there. I would point to the RL literature. Richard Sutton—and there's another guy, I think a co-founder of DeepMind, Shane Legg—wrote a paper trying to define what AGI is. In there, I think the definition of AI is the original, perhaps correct one, which is efficient continual learning.

Okay. If you truly want to build an artificial general intelligence that you can drop into any domain—you can drop it into a car without that much prior knowledge about cars—how long does it take a human to learn how to drive? Within months, you'd be able to drive a car really well. Generalized skill acquisition, generalized understanding acquisition, generalized reasoning acquisition.

I think that's the thing that would truly change the world. That's the thing that would give us a better understanding of the human mind and human consciousness, and that's the thing that would propel us to the next level of human civilization. On a civilization level, that's a really deep question, but separately, there's an academic aspect of it that I'm really—

So what odds—if we're on Kalshi today, what odds do we place on that?

I'm kind of bearish on a true AGI breakthrough because—

What we built is so useful and economically valuable. In a way—

Good enough is the enemy. Yeah. Do you remember that essay?

Worse is better.

Worse is better.

So there's a local maximum trap. We're in a local maximum—

A local maximum trap, because it's good enough for so much economically productive work.

Yes.

It relieves the pressure in the system to create the generalized answer.

Yes. And then you have the weirdos like Rich Sutton and others that are still trying to go down that path. Maybe they'll succeed, right? But there's enormous optimization energy behind the current thing that we're hill-climbing on, this local maximum.

Right, right, right. And the irony of it is that everybody's worried about the gazillions of dollars going into building out all this stuff. The most ironic thing in the world would be if the gazillions of dollars are going into the local maximum.

That's right.

As opposed to a counterfactual world in which they're going into solving the general problem.

But it's also potentially irrational. Maybe the general problem is actually not within our lifetimes. Who knows? Right?

How much further do you think we can go? Do you think we squeeze most of the juice out of LLMs in general, then? Or are there any other research directions that you're particularly excited about?

Well, that's the thing. I think the problem is there aren't that many. I think the breakthroughs in RL are incredibly exciting, but we've also known about them for over 10 years now: marrying generative systems with tree search and things like that.

But there's a lot more to go there, and I think, again, the original minds behind reinforcement learning are trying to go down that path and bootstrap intelligence from scratch. Carmack is going down that path, as far as I understand. You guys may be invested in Carmack, but they're not trying to go down the LLM path.

So there are people who are trying to do that, but I'm not seeing a lot of progress or outcome there. I watch it kind of from afar.

Although, for all we know, there's already a bot on X somewhere. You never know. It might not be a big announcement. It might just be that one day there's a bot on X that starts winning all the arguments.

Yeah, it could be—

Or a code agent, and all of a sudden it's generating incredible software.

Okay, let's spend our remaining minutes talking about you. So how—yeah, take us from the beginning with your life. How did you get from being born to being in Silicon Valley?

Okay.

In 2 minutes. Yeah, I'm just joking.

I got introduced to computers very early on. I was born in Amman, Jordan, and for whatever reason, my dad, who was just a government engineer at the time, decided that computers were important. He didn't have a lot of money, but he took out a loan and bought a computer. It was the first computer in our neighborhood, the first computer anyone I knew had.

One of my earliest memories is being 6 years old, watching my dad unpack this machine and open up this huge manual. He would finger-type “cd,” “ls,” “mkdir,” and I would be behind his shoulder, watching him type these commands and seeing the machine respond and do exactly what he asked it to do.

Pop a Tylenol as your autism activated.

Exactly. Autism activated.

Of course, you have to.

You have to.

Exactly. What kind of computer was it?

An IBM, as far as I remember. It was an IBM PC.

What year was this?

1993.

1993. Okay, so it was DOS. Did it have Windows at that point?

No, it didn't have Windows.

Right before Windows.

Right before Windows, but I think Windows had been out. You would add it on.

It was an add-on. You wouldn't boot it up.

I think we bought the disk for Windows, and you had to boot-load it from the disk. Then it would open Windows and you could click around. It wasn't that interesting because there wasn't a lot on it, so a lot of the time I just spent in DOS, writing batch files, opening games, and messing around with that.

It wasn't until Visual Basic, after Windows 95, that I started making real software. I used to be a huge gamer. I used to go to these LAN gaming cafes and play Counter-Strike. The whole place was full of computers, but they didn't use any software to run their business. It was just people running around, writing down your machine number, how much time you spent on it, and how much you paid, and tapping you on the shoulder like, “Hey, you need to pay a little more for that.”

I asked them, “Why don't you just build a piece of software that allows me to log in and have a timer or whatever?” They were like, “Yeah, we don't know how to do that.” I was like, “Okay, I think I know how to do that.”

I was 12 or something like that, and I spent 2 years building it. Then I went out and tried to sell it and was able to sell it. I was making so much money. I remember McDonald's opened in Jordan around the time when I was 13 or 14. I took my entire class to McDonald's. It was very expensive, but I was balling with all this money and showing off. That was the first business I created.

At the time, I started learning about AI, reading science fiction, and all of that. When it came time to go to college, I didn't want to study computer science because I felt like coding was on its way to being automated. I remember using these wizards. Do you remember wizards?

Yes.

Wizards, basically. They're extremely crude early bots that generate code.

Yeah.

I remember you could type in a few things—here's my project, here's what it does, whatever—and then click, click, click and scaffold a lot of code. I was like, “Oh, I think that's the future. Coding is such a—”

It's almost—

Yeah, it's solved. Why should I go into coding? I thought, “Okay, if AI can do the code, what should I do?” Well, someone needs to build and maintain the computers. So I went into computer engineering and did that for a while.

But then I rediscovered my love for programming, reading Paul Graham essays on Lisp and things like that, and started messing around with Scheme and programming languages like that. Then I found it incredibly difficult to learn different programming languages. I didn't have a laptop at the time, so every time I wanted to learn Python or Java, I'd go to the computer lab, download gigabytes of software, try to set it up, type a little bit of code, try to run it, and run into a missing DLL issue.

I was like, “Man, this is so primitive.” At the time, it was 2008 or something like that. We had Google Docs and Gmail. You could open the browser and, partly thanks to you, use software on the internet. I thought the web was the ultimate software platform. Everything should go on the web.

I thought, “Okay, who's building an online development environment?” No one. It felt like I had found a $100 bill on the floor of Grand Central Station. Surely someone should be building this, but no one was. So I was like, “Okay, I'll try to build it.”

I got something done in a couple of hours. It was a text box where you could type some JavaScript. There was a button that said “Eval.” You clicked Eval, and it evaluated the code and showed you the result in an alert box. 1 + 1 = 2. I was like, “Oh, I have a programming environment.”

I showed it to my friends, and people started using it. I added a few additional things, like saving the program. I was like, “Okay, there's a real idea here. People love it.”

Right, right.

Again, it took me 2 or 3 years to actually be able to build anything because the browser could only run JavaScript. It took a breakthrough at the time. Mozilla had a research project called Emscripten that allowed you to compile different programming languages, like C and C++, into JavaScript.

For the browser to be able to run something like Python, I needed to compile CPython to JavaScript. I was the first to do it in the world. I contributed to that project and built a lot of the scaffolding around it. My friends and I compiled Python into JavaScript.

I was like, “Okay, we did it for Python. Let's do it for Ruby. Let's do it for Lua.” That's how the idea for Replit emerged: When you need a REPL, you should get it. You should Replit it.

Replit is the most primitive programming environment possible. I added all these programming languages, and all this time my friends were using it and excited about it. I was on GitHub at the time, and my standard thing was that when I made a piece of software, I open-sourced it. So I was open-sourcing all the things I was building over the years—this underlying infrastructure to be able to run code in the browser.

Then it went viral on Hacker News, and it coincided with the MOOC era—massive open online courses. Udacity was coming online, Coursera, and most famously, Codecademy. Codecademy was the first kind of website that allowed you to code in the browser interactively and learn how to code.

They built a lot of it on my software that I was open-sourcing all the way from Jordan. I remember seeing them on Hacker News, and they were going super viral. I was like, “Hey, I recognize this. What are you using?” And so, I left a comment on Hacker News. I was like, “Oh, you’re using my open-source package.” They reached out to me and said, “Hey, we’d like to hire you.” I said, “I’m not interested. I want to start a startup. I want to start this thing called Replit.” They were like, “Well, no, you should come work with us. We can do the same stuff.” I kept saying no, and then I said, “Okay, I’ll contract with you.”

They were paying me $12 an hour, and I was really excited about it. I had just come back from Oman. But, to their credit, they came out to Jordan to recruit me and spent a few days there. I kept saying no, and in the end, they gave me an offer I couldn’t refuse. They got me an O-1 visa, and I came to the United States.

That’s when you moved. You were born in what year?

1987.

1987. What was the first year you could remember having the idea that you might not live your life in Jordan, that you might actually move to the United States?

When I watched Pirates of Silicon Valley.

Is that right? Okay, got it. All right.

Maybe 1998 or 1999. I don’t know when it came out.

Okay. That might be a good place to—

Yeah.

Is it worth telling the hacker story? There’s a version of the world where, if that had changed, maybe you wouldn’t have gone to America.

Right. In school, I was programming the whole time. I just wanted to start businesses. I was exploding with ideas all the time, and the reason Replit exists is because I have ideas all the time. I just want to go type on the computer and build them.

I wasn’t going to school because it was incredibly boring for me. Part of the reason Replit has a mobile app today is because I always wanted to program under the desk, just to do things. At school, they kept failing me for attendance. I would get A’s, but I just didn’t show up, so they would fail me.

I felt it was incredibly unfair, and all my friends were graduating. This was 2011, and I had been in college for 6 years. It should have been a 3- or 4-year program, and I was incredibly depressed. I really wanted to be in Silicon Valley, so I thought, “What if I changed my grades in the university database?”

There we go.

I went into my parents’ basement and implemented polyphasic sleep. Are you familiar with that?

I am.

Leonardo da Vinci’s polyphasic sleep. I didn’t hear about it from Leonardo da Vinci; I heard about it from Seinfeld, because there’s an episode where Kramer goes on polyphasic sleep.

Polyphasic sleep? What, 20 minutes every 4 hours?

20 minutes every 24 hours. Yes, somehow this was going to work well.

And hacking, if you’ve ever done anything—

As the meme goes, this has never worked for anybody else, but it might work for me.

Yes.

A lot of what hacking is involves coming up with ideas for finding certain security holes, writing a script, and running that script. The script will take 20 or 30 minutes to run, so you take that 20 or 30 minutes to sleep and go on.

I spent 2 weeks going mad, trying to hack into the university database. Finally, I found a SQL injection somewhere on the site and found a way to edit the records, but I didn’t want to risk it. I went to my neighbor, who was going to the same school—I think to this day no one has caught him—and said, “Hey, I have this way to change grades. Would you want to be my guinea pig?” I was honest about it. I said, “I’m not going to do it. Are you open to doing it?” He was like, “Yeah, yeah, yeah.”

They call these human trials. This is how medicine works. So we went and changed his grades, and he pulled his transcript, but the update wasn’t there. I went back to the basement. It turned out that I had access to the slave database, but I didn’t have access to the master database.

So you found a way through the network—privilege escalation. It was an Oracle database that had a vulnerability, and then you found the real database.

I just did it for myself. I changed the grades, pulled my transcripts, and sure enough, it had actually changed. I went and bought the gown, went to the graduation parties, and did all that. We were graduating.

One day, I was at home. It was maybe 6:00 or 7:00 p.m. The telephone at home rang—an ominous ring. I said, “Hello?” He was like, “Hey, this is the university registration system.” I knew the guy who ran it. He said, “Look, we’re having this problem. The system’s been down all day, and it keeps coming back to your record. There’s an anomaly in your record where you have a passing grade, but you’re also banned from the final exam for that subject.”

I thought, “Oh, shit.” It turned out the database wasn’t normalized. Typically, when they banned you from an exam, the grade reset to 35 out of 100, but apparently there was a Boolean flag. By the way, all the column names in the database were single letters. That was the hardest thing: security by obscurity.

Right.

It turned out there was a flag that I hadn’t tracked. When you go over the attendance limit, they want to fail you, so they ban you from the final exam. I changed the grades, and that created an issue and brought down the system.

They called me, and I thought, “I could potentially lie, and it’ll be a huge issue, or I could just fess up.” So I said, “Hey, listen, I might know something about it. Let me come tomorrow and talk to you about what happened.”

I went in and opened the door, and it was the deans of all the schools, including the computer science dean. They had all been working on it for days because it was a very computer-heavy university, and this was a serious problem. They were all really intrigued about what had happened, so I pulled up a whiteboard and started explaining what I had done. Everyone was engaged. I basically gave them a lecture.

This is your oral exam for your PhD. This is great.

They were really excited, and I think it was endearing to them. They were like, “Oh, wow, this is a very interesting problem.”

Okay, great. Thank you. Wait, we don’t know what to do with you. Do we send you to jail? Do we—

They said, “We have to escalate this to the university president.” He was a great man, and I think he gave me a second chance in life. I went to him, explained the situation, and said, “I’m really frustrated. I need to graduate. I need to get on with my life. I’ve been here for 6 years, and I can’t sit in school studying things I already know. I’m a really good programmer.”

He gave me a Spider-Man line at the time: “With great power comes great responsibility.” He said, “You have a great power.” It really affected me, and I think he was right in that moment. He said, “We’re going to let you go, but you’re going to have to help the system administrators secure the system for the summer.”

I was happy to do it. I showed up, and all the programmers there hated my guts. They would lock me out. I would see them outside, knock on the door, and nobody would let me in. I tried to help them a little bit, but they weren’t collaborative. I was like, “All right, whatever.”

When it came time for me to actually graduate, it was the final project. One of the computer science deans came to me and said, “Look, I need to ask you a favor. I was a big part of the reason we let you go and didn’t prosecute you, so I want you to work with me on the final project. It’s going to be about security and hacking.”

I said, “No, I’m done with that shit. I just want to build programming environments and things like that.” He said, “No, you have to do it.” I said, “Okay.” I thought I’d do something more productive, so I wrote a security scanner. I was very proud of it. It crawled different sites and tried SQL injection and all sorts of things.

My security scanner found another vulnerability in the system.

Amazing.

I went to the dean, and he said, “You need to run this security scanner live and show that there’s a vulnerability.” I didn’t understand what was going on at the time, but I said, “Okay.” I gave the presentation about how the system worked, and then I said, “Let’s run it.”

It showed that there was a security vulnerability. We said, “Okay, let’s try to get a shell.” The system automatically ran all the security tools and got us a shell. Then I realized that the other dean had been given the mandate to secure the system. I started to realize I was a pawn in some kind of rivalry.

His face turned red, and he said, “No, it’s impossible. We secured the system. You’re lying.”

I said, “You’re accusing me of lying? What should I look up—your salary or your password? What do you want me to look up?” And I was like, “Yeah, look up my password.” So I looked up his password.

It was gibberish. It was encrypted. I was like, “Oh, that’s not my password. See, you’re lying.” I was like, “Well, there’s a decrypt function that the programmers put in there.” So I decrypt it, and it shows his password. It was something embarrassing. I forgot what it was. So he gets up really angry, shakes my hand, and leaves to change his password.

So I was able to hack into the university another time. Luckily, I was able to graduate. I gave them the software, and they secured the system. But, yeah, later on I would realize that he wanted to embarrass the other guy, which was why I was in the middle.

Politics. Well, I think the moral of the story is, if you can successfully hack into your school system and change your grade, you deserve the grade, and you deserve to graduate.

I think so.

And just for any parents and children out there, you can cite me as the moral authority on this. One lesson I think is very relevant for the AI age: I think that the traditional, more conformist path is paying less and less dividends, and I think kids coming up today should use all the tools available to be able to discover and chart their own paths, because I feel like just listening to the traditional advice and doing the same things that people have always done isn’t working out as much as we’d like.

Thanks for coming on the podcast.

Thank you, man.

Fantastic. Wow. Wow. Wow.