We're dealing with magic here that we probably all 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. Like it's not moving fast enough and it's like maybe right on the verge of stalling out. We should both be like hyper excited but also on the verge of like slitting our wrists cuz 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, let's assume that I'm a sort of a novice programmer. So, maybe I'm a student or maybe I'm just somebody, I took a few coding classes and I've hacked around a little bit or I do Excel macros or something like that, but I'm not like a master craftsman at coding. And somebody tells me about Replit and specifically AI and Replit. What's my experience when I launch in with 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 like setting up a development environment and all of that stuff and just have you focus on your idea. So 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? So the prompt box is really open for you. You can put in anything there.

So let's say you want to, you know, build a startup. You have an idea for a startup. I would start with like a paragraph long kind of description of what I want to build. The agents will read that. It will...

You just type standard English.

Standard English. You just type it in. I want to build a... I want to sell crepes. I want to sell crepes online. So you just type in...

It literally could be that four words or five words. Okay.

Or it could be if you have a programming language you prefer or stack you prefer, you could do that. But we actually prefer for you not to do that because we're going to pick the best thing for... we're going to classify the best stack for that request. If it's a data app, we'll pick Python. If it's like a web app, we'll pick JavaScript and Postgres and things like that. So you just type that.

Or you can decide. You can say, "And I want to do it, 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 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. And then just again I know this is obvious people have used it but like I'm dealing in English.

Yes. You're fully in English. I mean, a little bit of sort of background here, when I came here and pitched to you like 7 years ago, what we were saying is we were exactly describing this future is that everyone would want to build software, right? And the thing that's kind of getting in people's ways is all the what Fred Brooks called the accidental complexity of programming, right? There's essential complexity which is like how do I bring my startup to market and how do I build a business and all of that. Accidental complexity is what package manager do I use. All of that stuff we've been abstracting away for so many years. And the last thing we had to abstract away is code.

I had this realization last year which is I think we built an amazing platform but the business is not performing and 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 so ultimately English is the programming language.

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 that tends to be very...

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

No, most mainstream languages that have like 100 million plus people that speak it. AI is pretty good at it.

So, I did a bit of historical research recently for some reason. I just want to understand the moment we're in and because it's such a special moment. I think it's important to contextualize it. And I read this quote from Grace Hopper. So, Grace Hopper invented the compiler as you know. At the time people were programming in machine code and that's what programmers do, that's what the specialists do.

And she said specialists will always be specialists, they have to learn the underlying machinery of computers but I want to get to a world where people are programming English. That's what she said. That's before Karpathy, right? That's like 75 years ago. And that's why I invented the compiler. And in her mind like C programming is English.

But that really didn't... that was just the start of it. You had C and then you go higher level Python and JavaScript. And 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.

I remember when I was a kid it was, there were higher level languages by the 70s like BASIC and so forth and Fortran and C. But there were still, you still would run into people who were doing assembly programming, assembly language which by the way you still do, like game companies or whatever still do assembly to get...

And they were hating on the kids that were doing BASIC. Oh well so the assembly people were hating the kids doing BASIC but there were also older coders who hated on the assembly programmers for doing assembly and not doing direct machine code, right, not doing direct zero and one machine code. Because assembly language is sort of this very low-level programming language that sort of compiles to actual machine code and it's incomprehensible gibberish to even most programmers.

You're writing in octal or something.

You're writing like very, very close to the hardware but even still it's still a language that compiles to zeros and ones. Whereas the actual real programmers actually wrote in zeros and ones. And so there's always this tendency for the pros to be, you know, look on the nose. And say, you know, the new people are being sloppy. They don't understand what's happening. They don't really understand the machine.

And then, of course, with the higher level abstractions do is they democratize. The absolute irony is 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. And we got a lot of hate from the programmers that you should type vanilla JavaScript directly.

And I was like okay whatever and then that's mainstream and then those guys that built their careers on the last wave we invented are hating on this new wave and so just people never change.

Okay, got it. Okay, so 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. Okay. What happens then?

Yeah. and then Replit agent will show you what it understood. So it's trying to build a common understanding between you and it. It'll show you a list of tasks. It'll tell you I'm going to go 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. And then it shows you this list and gives you two options initially. Do you want to start with a design so that we can iterate back and forth to lock that design down or do you want to build a full thing?

Hey, if you want to build a full thing, we'll go for 20, 30, 40 minutes. The agent will tell you go here, install the app. I'm going to go set up the database, do the migrations, write the SQL, build the site. I'm going to also test it. So this is a recent innovation we did with Agent 3 is that after it writes the software spins up a browser goes around and tests in the browser and then any issue it kind of iterates and goes and fix the code.

So it'll spend 20-30 minutes building that. I'll send you a notification, it'll tell you the app is ready. And so you can test it on your phone. You can go back to your computer. You'll see maybe you'll find a bug or an issue, you'll describe it to the agent, say, "Hey, it's not exactly doing what I expected." Or if it's perfect and you're ready to go, and that's it. By the way, there's a lot of examples where people just get their idea in 20-30 minutes, which is amazing. You just hit publish.

You hit publish. A couple 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 two or three years ago in order to get to that step. You have to set up your local development environment. You have to sign up for an AWS account. You have to provision the databases, the virtual machines, you have to create the entire deployment pipeline. All of that is done for you. And a kid can do it, a lay person 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, you can peel the layers. You can open the file tree and you could look at the files. You can open Git, you can push it to GitHub, you can connect it to your editor if you want. So the cool thing about Replit, 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, you said it gives you, the agent gives you, you say I've got my idea. You plug it in and it says it gives you this list of things and 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. Agent does those things. Yeah. That's a very important point. When we did this shift, we hadn't realized internally at Replit how much the actual user stopped being the human user and it's actually the agent programmer, right?

So, one really funny thing happened is we had servers in Asia and the reason we had servers in Asia 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 and we're like what happened? It's supposed to be faster. Well turns out it's worse because the AIs are sitting in the United States and so the programmer is actually in United States. You're sending the request to the programmer and the programmer is interfacing with a machine across the world and so yes suddenly the agent is the programmer.

Okay so like the new terminology agent is a software program that is basically using the rest of the system as if it were a human user, but it's a bot. That's right. It has access to tools such as write a file, edit a file, delete a file, search the package index, install a package, provision a database, provision object storage. It is a programmer that has the tools and interface. It has a sort of an interface that is very similar to a human programmer.

And then a debate inside the AI industry is with these idea now of having agents that do things on your behalf and then go out and kind of accomplish missions. There's this debate which is okay how long does it maintain coherence? Like how long does it actually like stay in full control of its faculties and not kind of spin out? Because at least the early agents or the early AIs, if you set them off to do this, they might be able to run for two or three minutes, then they would start to get confused and go down rabbit holes and kind of spin out. More recently, we've seen that agents can run a lot longer and do more complex tasks. Like, where are we on the curve of agents being able to run for how long and for what complexity tasks before they break?

That's absolutely the main metric we're looking at. Even back in 2023, you know, had the idea for software agents, four or five years ago now. The problem every time we attempt them, the problem of coherence, they'll go on for a minute or two and then they'll just, they compound in errors in a way that they just can't recover.

And you can actually see it, right? Because they actually, if you watch them operate, they get increasingly confused and then, maybe even deranged. They go into very weird areas and sometimes they start speaking Chinese and doing really weird things. But I would say sometime around last year we maybe crossed the three, four, five minute mark and it felt to us that okay we're on a path where long horizon reasoning is getting solved.

So long horizon reasoning meaning reasoning like 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 is that they have a context. This context is basically the memory all the text all your prompt and also all the internal talk that the AI is doing as it's reasoning. So when the AI is reasoning it's actually talking to itself. It's like oh now I need to go set up a database. Well, what kind of tool do I have? Oh, there's a tool here that says Postgress. Okay, let me try using that. Okay, I use that. I got feedback. Let me look at the feedback and read it. And it'll read the feedback. And so that prompt box or context is where both the user input, the environment input, and the internal thoughts of the machine are all within. It's sort of like a program memory space.

And so reasoning over that was the challenge for a long time. That's when AIs just went off track and now they're able to kind of think through this entire thing and maintain coherence. And there's now techniques around compression of context. So there still context length is still a problem, right? So I would say LLMs today, they're marketed as a million token length, which is like a million words almost. In reality it's about 200,000 and then they start to struggle. So we do a lot of, we stop, we compress the memory. So if a portion of the memory is saying that I'm getting all the logs from the database you can summarize paragraphs of logs with one statement or the database setup that's it. And so every once in a while we'll compress the context so that we make sure we maintain 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. So the way pre-training works is, pre-training is the first step of training a large language model. It reads a piece of text. It covers the last words and tries to guess it. That's how it's trained. That doesn't really imply long context reasoning. But the reason we weren't able to move past that limitation is that that modality of training just wasn't good enough. And what you want is you want a type of problem solving over a long context. So what reinforcement learning especially from code execution gave us is the ability for the machine to for the LLM to roll out what we call trajectories in AI. So trajectory is a step-by-step reasoning chain in order to reach a solution.

So as I understand reinforcement learning works is they put the LLM in a programming environment like Replit and say hey 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. So we have a pull request that we have on GitHub so we know exactly or we have a unit test that we can run and verify the solution. So what it does is it rolls out a lot of different trajectories. They sample the model and maybe one of those trajectories will reach the solution and a lot of them will just go off track but one of them will reach the solution by solving the bug and it reinforces on that. So that gets a reward and the model gets trained that okay this is how you solve these type of problems. So that's how we're able to extend these reasoning chains.

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

There is a nonprofit called Meter that is measuring, 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. And they put up a paper late last year that said every seven months the minutes that a model can run is doubling. I think they vastly underestimated that.

Is that right? Vastly it's doubling. It's doubling more often than 7 months.

We so Agent 3 we measure that very closely and we measure that in real tasks from real users. So we're not doing benchmarking. We're actually doing AB tests and we're looking at the data that how users are successful or not. For us, the absolute sign of success is you made an app and you published it. Because 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. And so what we're seeing is in agent one, the agent could run for 2 minutes and then perhaps struggle. Agent two came out in February, it ran for 20 minutes. Agent 3, 200 minutes.

200. Some users are pushing it to like 12 hours and things like that. I'm less confident that it is as good when it goes to these stratospheres, but at like 2-3 hours timeline, it is really insanely good. And the main innovation outside of the models is a verification loop. Actually, I remember reading a research paper from Nvidia. So what Nvidia did is they're trying to write GPU kernels using DeepSeek and 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 like 20 minutes and it was generating actually optimized kernels.

And so I was like, okay, the next thing for us obviously as an agent lab or applier company. We're not doing the foundation model stuff, but we're doing a lot of research on top of that. And so, okay, we know that agents can run for 10-20 minutes now or LLMs can stay coherent for longer, but for you to push them to 200, 300 minutes, you need a verifier in the loop. So, 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. So once you put that in the middle, what's happening is it works for 20 minutes, spin up another agent, spins up a browser, tests the work of the previous agent. So it's a multi-agent system and if it founds a bug it starts a new trajectory and says okay good work let's summarize what you did the last 20 minutes.

So it's like setting up a marathon or like a relay race. As long as as long as each step is done properly you could do in sort of 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.

So you know we're kind of getting into here kind of the holy grail of AI which is sort of generalized reasoning by the machine. The key thing there though for for RL to work for LLMs to reason the key is that it be a problem statement that there is a defined and verifiable answer. That's right. Is that right? And so and that for sure includes math. That for sure includes physics for sure includes chemistry. For sure includes large areas of code.

Yes. Absolutely. And I think that's what we're saying.

Bio, like we're seeing with a protein. Genomic. Yeah. Okay. Right. Yeah. Things like that. I think some some areas of robotics, right? There's a clear outcome, right?

But the first two do not work very well just yet. Like law and healthcare they're still a little too squishy, a little too soft. It's unlike math or code. So I would say the softer domains meaning like domains in which it's harder or even impossible to actually verify correctness of result in a sort of a deterministic factual grounded, non-controversial way.

Right. So sort of the more concrete the problem, like it's the concreteness of the problem that is the key variable not the difficulty of the problem. Would that be a way to think about it? In any domain of human effort in which there's a verifiable answer we should expect extremely rapid progress.

I think we're ripping on coding. I think it's going to be like what we're working on with agent four right now is by next year we think you're going to be sitting in front of Replit and you're shooting off multiple agents at a time. You're like planning a new feature. Um so I want social network on top of my storefront and another one is like hey um refactor the database. And you're running parallel agents. So, you have five, 10 agents kind of working in the background and they're merging the code and taking care of all of that. I think the lay person will be as good as what a senior software engineer that works at Google is today. So I think that's happening very soon.

But I don't see them... my experience between as a sort of a, on the let's say healthcare side or more, write me an essay side or more creative side haven't seen as much of a rapid improvement as what we're seeing in code. So I think code is going to go to the moon. Math is probably as well some scientific domains bio things like that those are going to move really fast.

Yeah. So there's this weird dynamic. We have this in the office here a lot and I also have this with the 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 like really disappointed and like it's not moving fast enough and like it's like maybe right on the verge of stalling out. And we should both be like hyper excited but also on the verge of like slitting our wrists because the gravy train is coming to an end, right?

But there's a huge bet on AGI, right? Like whether it's the foundation models, I think now the entire US economy is sort of a bet on AGI and there are crucial questions to ask whether are we on track to AGI or not because 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 of significance right so if we get a lot better at code we're not immediately getting better at like generalized reasoning.

And you know Ilya Sutskever makes a specific form of this argument which is basically like we're just literally running out of training data. It's a fossil fuel argument right like if we slurped all the training fundamentally we've slurped all the data off the internet that is where almost all the data is at this point. There's a little bit more data that's in like private dark pools 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 like slurping things off the internet.

Having said that, you mentioned transfer learning. So transfer learning is the ability of the machine to be an expert in one domain and then generalize that into another domain. My answer to that is like have you met people? And how many people do you know 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 kind of often they have blind spots. This is a well-known thing among like for example public intellectuals. So you get these people who show up on TV and they're experts and what happens is they're like an expert in economics right and then they show up on TV and they talk about politics and they don't know anything about politics. You know, this is the Paul Krugman talking about how the internet's going to be no more significant than the fax machine. He's a brilliant economist. He has no idea how a computer works.

Einstein's like actually my favorite example. I think you'd agree Einstein was a brilliant physicist. He was like a Stalinist. He was a socialist and he was a Stalinist. Once he got into politics, he was just like totally loopy. There was no transfer learning from physics into politics.

So in a way the argument you're making is like we maybe already have human level AI. I mean perhaps the definition of AGI is something totally different is like above human level that something that totally generalizes across domains. We've idealized a goal that may be idealized in a way that it's so far beyond what people can do that it's no longer a relevant comparison to people.

There's also this well-known phenomenon in AI which is the definition of AI is always the next thing that the machine can't do. And so like the definition for of AI for a long time was like can it beat humans at chess. And then the minute it could beat humans at chess that was no longer AI that was just like boring computer chess. The Turing test was the test and then we passed it and nobody, there was no celebration. For 80 years the Turing test, I mean they made a movie about it, that was the thing and like we blew right through it and nobody even registered it. Nobody cares.

I started thinking about this idea of like it doesn't matter whether it's truly AGI. We can get to like functional AGI and what functional AGI is is just yeah collect data on every useful economic activity in the world today and train an LLM on top of that or train the same foundation model on top of that and we'll go we'll target every sector economy and you can automate a big part of labor that way. So I think we're on that track for sure.

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 or what are your thoughts there?

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. It felt a lot more robotic, you know, very in its head kind of trying to think through everything. And so I would have just expected like when we went from GPT-2 to 3, it was clear it was getting a lot more human. It was a lot closer to our experience. You can feel like it's actually, it gets me. Similarly 3 to 4. 4 to 5 didn't feel like it was a better overall being as it were.

I like to ask models like very controversial things. Can it reason through, I don't know how deep we want to go here but like what happened with World Trade 7, right? It's an interesting question. Can it think through controversial questions in the same way that it can go think through a coding problem? And there hasn't been any movement there. COVID, right, like the origins of COVID. You go dig up GPT-4 or other models and go to GPT-5, you're not going to find that much difference of okay, let's reason together. Let's try to figure out what was the origins of COVID because it's still an unanswered question, and I don't see them making progress in that.

I use it differently. My main use case actually is sort of PhD and everything at my beck and call. And so I'm trying to get it to explain things to me more than I'm trying to have conversations with it. When an advanced economy puts a tariff on a raw material or on a finished good, like who pays? You know, is it the consumer, is it the importer, is it the exporter or is it the producer? And this is actually a very complicated question. It's a big thing that economists study a lot. And it's just like, okay who pays? And what I found like for that kind of thing is it's outstanding at sort of going out of the web getting information synthesizing it. It gives me a synthesized 20, 30, 40 pages... basically tops out 40 pages of PDF. But it's completely coherent and as far as I can tell for everything I've cross-checked, completely like world class.

I'm kind of bearish on true AGI breakthrough because what we built is so useful and economically valuable. Good enough is the enemy. There's like a local maximum trap. We're in a local maximum trap where it's because it's good enough for so much economically productive work. 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 and maybe they'll succeed, right? But there's enormous optimization energy behind the current thing that we're hill climbing on this like local maximum.

Let's spend our remaining minutes. Let's talk about you. So, how did you get from being born to being in Silicon Valley?

I got introduced to computers very, 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 and he didn't have a lot of money, took out a loan, bought a computer. It was the first computer in our neighborhood. First computer of anyone I know. And one of my earliest memories I was 6 years old just watching my dad unpack this machine and sort of open up this huge manual and kind of finger type CD, LS, MKDIR and like I would be behind his shoulder and just like watching him type these commands and seeing the sort of machine kind of respond and do exactly what he's asked it to do.

It was an IBM as far as I remember. It was IBM PC. What year was this? 1993. So, it's DOS. So, did it have Windows at that point? No, it didn't have Windows. But it wasn't until Visual Basic that I started. So like after Windows 95 that I started making real software, right?

I used to be a huge gamer. So I used to go to these LAN gaming cafes and play Counter-Strike and I would go there and the whole thing is full of computers but they don't use any software to run their business. It was just like people running around just like writing down your machine number, how much time you spend on it and how much did you pay. And I asked him like why don't you just build a piece of software that allows me to log in and have a timer or whatever. And he was like yeah we don't know how to do that. And I was like okay I think I know how to do that. So I spent, I was like 12 or something like that. I spent like 2 years building that and then went out and tried to sell it and was able to sell it and was making so much money. I remember McDonald's opened in Jordan around the time when I was 13-14. I took my entire class to McDonald's. It was very expensive, but I was balling with all this money.

When it came time to go to college, I didn't want to go to computer science because I felt like coding is on its way to get automated. I remember using these wizards. I was okay if AI can do the code what should I do? Well someone needs to build and maintain the computers and so I went to computer engineering and did that for a while but then rediscovered my love for programming.

But then I found it incredibly difficult to just like learn different programming languages. I didn't have a laptop at the time. And so every time I go to wanting to learn Python or Java, I would go to the computer lab, download gigabytes of software, try to set it up, type a little bit of code, try to run it. I was like, man, this is so primitive. Like at the time it was 2008 something like that. We had Google Docs, we had Gmail. I thought the web is the ultimate software platform. Everything should go on the web. Okay, who's building an online development environment, right? And no one. It felt like I found like $100 bill on the floor. Surely someone should be building this, but no one was building this.

So I was like, okay, I'll try to build it. It took a breakthrough at the time. Mozilla had a research project that allowed you to compile different programming languages like C, C++ into JavaScript. And for the browser to be able to run something like Python, I needed to compile CPython to JavaScript. So I was the first to do it in the world. So built and contributed to that project and we compiled Python into JavaScript and I was like okay we did it for Python let's do it for Ruby let's do it for Lua and that's how the emergence of the idea for Replit came.

And I was on GitHub at the time and my standard thing is when I make a piece of software is open source it. And then it went viral on Hacker News and it coincided with the MOOC era. So massively online courses, Udacity was coming online, Coursera and most famously Codecademy. So, Codecademy was the first kind of website that allowed you to code in the browser interactively and learn how to code. And they built a lot of it on my software that I was open sourcing all the way from Jordan. And so, they reached out to me. They're like, "Hey, would like to hire you." I was like, "I'm not interested. I want to start a startup. I want to start this thing called Replit." In the end, they gave me an offer I can't refuse. And they got me an O1 visa. Came to the United States.

In school they kept failing me for attendance. You know, so I would get A's, but I just didn't show up and so they would fail me. And so I felt it was incredibly unfair. And all my friends were graduating now. This year was like 2011. I've been like for 6 years in college. And I was like incredibly depressed. I really wanted to be in Silicon Valley. And so I was like, "Oh, what if I changed my grades?"

The university database. And so I went into my parents' basement and implemented the polyphasic sleep. I spent two weeks just going mad like trying to hack into the university database and finally I found a way. I found a SQL injection somewhere on the site. I went to my neighbor who's going to the same school. I said, "Hey, I have this way to change grades. Would you want to be my guinea pig?" So we went and changed his grades. And then I just, you know, did it for myself. Changed the grades and went and pulled my transcripts and sure enough it actually changed. Went and bought the gown, went to the graduation parties.

And then one day I'm at home. I get a telephone at home rings. "Hey this is the university registration system." And he's like "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're both pass, you have a passing grade but you're also banned from that final exam." I just fess up. So, I go in and I open the door and it's the deans of all the schools. And I pull up a whiteboard and started explaining what I did and everyone was engaged. I gave them a lecture basically.

And he gave me a second chance in life and I went to him and I explained the situation. He gave me a Spider-Man line at the time. was like "With great power comes great responsibility and you have a great power." And so he said well we're going to let you go but you're going to have to help the system administrators secure the system.

And one of the computer science dean came to me and he said look I need to call a favor. I want you to work with me on the final project and it's going to be around security and hacking. So I wrote a security scanner and actually my security scanner found another vulnerability in the system. And so I went to the defense and he's like you need to run this security scanner live and show that there's a vulnerability. And now I started to realize I'm a pawn in some kind of rivalry here. And his face turned red and he's like, "No, it's impossible. We secured the system. You're lying." I was like, "What should we know? Should we know your salary or your password?" So, I look up his password. It was something embarrassing. And so he gets up really angry, shakes my hand and leaves to change his password. Luckily I was able to graduate, gave them the software, they secured the system.

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. One maybe lesson I think that is very relevant for the AI age. I think that the traditional sort of 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 cuz I feel like just listening to the traditional advice and doing the same things that people have always done is just not working out as much as we'd like.