The Infra Pod - English is the default programming language via GPTScript! Chat with Darren from Acron

Episode Date: May 13, 2024

Ian and Tim sat down with Darren who is the cofounder of Acorn Labs that's building open source infrastructure to build AI applications. We chatted about their first project which is GPTScript, allows developers to use english language mixed with programming languages and templates to generate apps in a more structured way. Join us to listen about how the team brought their experience from working on Kubernetes during the Rancher days, and what their outlook and spicy take about the future of developers powered with AIs!

we're back at the yet another infra deep dive podcast and we're ready to roll i'm tim from essence vc and let's go in i'm ian helping sneak turn to a platform i couldn't be more excited talk to darren who in many ways is a legend infrastructure he's been around for a long time seen a lot of different things i've been working on some cool stuff. Darren Shepard, tell us who you are and a little bit about yourself. Yes, I'm Darren Shepard, Chief Architect at Acorn Labs. Previously was doing Rancher, which is another startup in the Kubernetes space. And before that was a relationship to cloud.com, and that was in the early

IaaS days of like OpenStack and CloudStack. So basically I've spent the last, like, I don't know, maybe 15 years or so pretty heavily in the IaaS infrastructure space. And what I'm currently doing is not infrastructure, just recently pivoted completely to AI, which is kind of weird. This is an infrastructure podcast, but it's like just recently kind of left the space. I'm hopefully bringing all of the things I've learned from the last kind of 15 years of trying to see if we can apply these learnings to the crazy world of AI, which is madness, pure madness right now. We can't let you start off with that statement without asking you, well, what convinced you?

Like what made you say, you know what, this infrastructure stuff I've been doing, let's put that to the side and go after like focusing AI. Was there some insight? Was there something you did? Like paint the picture for us. Yeah, yeah. And I forgot to pitch,

like the thing I'm working on is called GPT script. So like that's the thing that kind of changed my mind. It was a side project. Acorn Labs, we've been around for about 18 months or so. And we were working in the Kubernetes space, basically trying to make Kubernetes more usable. I'm pretty outspoken about how much I hate using Kubernetes, even though I've built things like K3S and Rancher. So we were trying to improve the Kubernetes space or whatever.

And so what's going on, like everyone cares these days about AI. And so there's kind of this pressure of like, you got to do something with AI. And so I started looking into this really like in November, very intently of like, if AI is the future, let me see if I can actually use it to do something. To a certain degree, it's like kind of like, I'm either going to prove it right or wrong, thinking much more that i would kind of prove it wrong and so i set out trying to build like the stupid demo everyone does which is like look like i have this prompt and it generates this application but like i hated all those demos because it's like it generates a garbage application you can't really maintain like you know how are you going to develop it over time? And, and most of the time,

those demos are just like, they only work because they're demoing like a, a well known app, like Angry Birds or Tetris or Flappy Bird, you know, it's like something AI obviously could do. Anyway, so I sat down to try to like, could I realistically build something that generates an application, and I failed miserably, just like everyone else's, is like I failed I couldn't figure it out but like in the process I actually started to realize like AI has all the building blocks that programming has it's just like all poorly defined right now but like I think I can get to the point where it's like I can structure AI and do one very small basic task well and then just keep like encapsulating and composing and composing and eventually do like much more complicated things so I I started to see that like, hey,

I can actually build a system with AI to eventually do like kind of what programmers are doing, not exactly, but like, tackle same problems. And so I got very fascinated with the idea of like, you know, using AI to do these tasks, Like my entire career has been around pretty much automation. It's like automating infrastructure. And I started to realize it's like, oh, I can apply all the same kind of things I know about automation to AI. And like specifically a lot of stuff that I've learned about like Kubernetes and like I can go and deep dive on that stuff too.

I mean, that's interesting. I mean, what is it about the AI landscape today that you're like, is there like specific problems? You're like, oh, this is just like this one problem we solve in Kubernetes. Like, do you have some examples? I mean, I think we'd all like that. I mean, I just think it's interesting where it's like, so there's this one feature that OpenAI has,

which is called function calling, which was really like eye-opening to me, where it's like, okay, what I can do is I can ask AI to effectively say like, if this, you know, call this function, otherwise call this function, it's really dumb the way it works. All it's really does is like you give it a prompt, like, you know, give it this sentence. And it will say, oh, you need to call this function. So it tells you, you have to call the function, get the result of that and give it back to AI.

And then it will continue on its way. So it's like, oh, I can use AI and I can do very basic logic. Like if, you know, so I can branch, I can do for loops. So it's like, oh, that's the basics of a programming language. And then you can look at things where it's like, oh, well, it can understand data like a JSON blob and transform it into like another JSON blob, like so it can do basic data understanding and manipulation and stuff. So it's like, today, everything's built around like the CPU, which is discrete mathematics of like, you know,

zeros and ones. But like, now I have this new, like logic processor, it can just do things logically, I can give it a logical task, and then it will do the right thing. It just became very interesting to me of like, oh, I can build very small little functions that just perform some logical operation, like I give it an input, and it gives me an output. Like the difference is like, today with programming, you're dealing with data structures and numbers and whatnot. But like we take like kind of human knowledge, and we encode it into numbers, and then we transform numbers and whatnot, and then we take like kind of human knowledge and we encode it into numbers and then we transform numbers and whatnot and then encode it back into human knowledge.

Whereas AI kind of works at a very similar like idea, but we don't have to encode concepts into like numerical representations. It just became interesting of like, hey, the way I solve problems with computers and stuff like that, like what if I took all those same concepts, but then did it in a more abstract form? Like, can I actually build applications? So it just became like a very interesting problem of like, can we actually

build what we think is the future of like AI applications? Like, what does that actually mean? You know, you're not going to use AI really to build an e-commerce site, maybe, maybe not. But like, you know, what does it really mean? Because like I've worked a lot in like, ESB used to be like this popular trend or it still kind of exists in enterprise where it's like you create this data bus of plugging into all these services

and then you transform and stuff like that. It's like that whole process is very difficult and companies struggle with it. Whereas like you could look at that whole space and be like, you know what? AI can do that really easily. It can plug into services and it can transform and understand data really easily.

So it's like, you know, could that transform that whole space? So I don't know. I think I went off on a tangent there. No, it was great. And so how does like GPT script seem to work out? Like, what is it?

And how does this fit into your thinking? And why'd you make the bet on this? help us understand those things yeah yeah so so what GPT script is is like I really wanted to figure out like what's a natural language interface to programming what bothered me when I started getting into AI was like it's just a bunch of Python just all this Python and um it's a lot of Python coming from basically like college students and stuff. And so it's like, or data scientists, it's not particularly well written Python, a lot of the libraries have matured, and they've gotten a lot better. But it's just kind of the sea of all this crazy code that doesn't really make sense. And there's a lot of academic terms built

into it and stuff. And so it's like, it's very difficult. Like i see a lot of like kind of more let's say like traditional software developers who can't get over this hurdle of like what the heck is ai and how do i use it with gpt script i wanted to create just like can i boil it down to some very simple primitives that's all completely natural language driven so i don't have to learn python i don't necessarily have to learn programming can i just like solve the problems with completely natural language driven. So I don't have to learn Python. I don't necessarily have to learn programming. Can I just like solve the problems with just natural language, but still attack it with like an engineering mindset of like, take a problem, break it down into smaller problems, you know, kind of compose those and whatnot.

And so that was what I was trying to do with GPT script. So I think maybe to explain what GPT script is to folks that we don't know what it is. And it's very interesting that your background is so much infrastructure, like, you know, Rancher, all these things are Kubernetes. These are like behind the scenes infra, we call it the boring stuff. That's very, very valuable. GPT script is so much more at the forefront. We're creating basically the interface, what people want to be programming with, with additional like metadata around this sort of like prompt, I suppose. So yeah, explain what it is and why you even want to start here. The basic design of GPT script is like there's a thing called a tool. So a tool is basically

implemented either with natural language, which is like a prompt it's just a sentence of like one of the canonical examples we have is this one called sqlite download where it's like the tool says like download this the sqlite database at this url extract it find out what the schema is and then query it and tell me what's the like the artist with most albums or something. So you create a tool, which is just like a prompt. Those tools can then invoke other tools. And so the other tools you can give it to can also be other AI prompts that do something, or it can actually be like traditional code. So it can be like Python or just like a shell script or whatever, like you can integrate into traditional code. So when you write this prompt, which is a tool,

you then give it other tools, but you never at any point do you like, you just make the tool available to AI. And so it decides when it wants to call the tools, what are the arguments to the tools? And then it will get the response. And so you can create these little building blocks with tools and you can stitch them all together.

So that's kind of the basics of what GPT script is. It's like a very simple system to use. You can get up and running really quickly. It's all right now based off of OpenAI. That's the easiest way you get an API key. You use that and you can get going real quick. But it will work with local models and stuff. But that's a lot harder.

A lot of the local models, they don't perform anywhere near as GPT-4. But the thing that like fascinates me, it's like, I've spent a lot of time building orchestration systems. So I spent a lot of time in like architecture, distributed system, these like event reconciliation loops and stuff like that. And so there's this whole other realm and it's like GPT script is not there yet. When I look at like what people want to do with AI, they'll say like, there's this idea of these agents. And so people will say like, oh, you have this agent where like one is a coder and one's a QA person. So the coder produces code and the QA person tests the code.

And then you have the manager. So you define these agents and they're all supposed to loosely interact with each other and do things or whatever. Well, if you look at the design of those, like it's basically the same thing we were doing in Kubernetes with controllers. Like one of the innovations, people don't realize how innovative Kubernetes was in this, is that like what you do is you have all of these basically reconciliation loops that work on a shared data structure a shared data source like this was pretty revolutionary at the time that kubernetes came out people didn't realize this because like at that time soa was still the predominant architecture or like kind of the leftovers of soa like rest had already taken over

but like the you know artifacts of it and what specific, one of the things about SOA, service-oriented architecture, was that every service has its own data store. And so communication between services was all through service interfaces, whereas Kubernetes was completely different. They're just like, screw it. Everyone talked to the database, which I think was like the more common thought at Google at the time, because they had a really big, well-structured data store. So when you look at AI, what you can do with it, it's like... And so this is where GPT script is right now. It's basically kind of at the level of like, can I create functions and stitch together functions? Where I'm going with GPT script is try to get to these event loops, these reconciliation

loops that are the same as Kubernetes where it's like, okay, well, what if I define how these functions that basically work in a reconciliation loop, and they work on a shared data structure? Like, it doesn't really matter what that data structure is. It can be a database, it can be a service, but they work on shared data so they can collaborate with each other. So it's like, now I have these agents which have a design purpose. So like, this is the things that gets me excited, because because like I've worked in the infrastructure space for so long. And like, I hate to say this, but like, I do know like Kubernetes and orchestration really well. I've just done this crap for long enough that it's like, I feel like, yeah, I'm kind of an expert on doing this type of stuff. But then I look at the parallels in AI and it's like, oh, like I understand all this and it's very similar,

but it's at a completely different level of abstraction. Like it's like, oh, like I understand all this and it's very similar, but it's at a completely different level of abstraction. Like it's, it's not concrete data, but like it all aligns well that like, that's where I get excited where it's just like, you know what? Like I've spent enough time in the industry doing like kind of concrete things with infrastructure and whatnot, that I think all of these concepts, if we apply them to AI, we could do some really crazy, like powerful things with AI if we take all the, like these learnings that we've had. So it's like, that's why I get excited.

It's like a complete paradigm shift. It's completely new for me. And I feel like I can apply the stuff that I kind of previously learned. So is there like a core use case that you're really focused on right now, enabling? Like one of the things you mentioned

is this idea of this collaboration. So it's like, I actually think your parallel to just Kubernetes is great. Like I remember when Kubernetes came out and it was like, you kind of had this self-hearing infrastructure is the way I thought, One of the things you mentioned is this idea of this collaboration. I actually think your parallel to just Kubernetes is great. I remember when Kubernetes came out, it was like you kind of had this self-healing infrastructure, is the way I thought, but you had programmable self-healing infrastructure. And it was just like, okay, we're done with the Bash scripts and Ansible and where we're effectively different ways to run Bash scripts.

We now have this sort of ability for infrastructure to self-heal and react to each other on its own, which was great because now you could program and you could have these things self-evolve over time. So I like your analogy because I deeply understand. I got excited about Kubernetes for a very similar reason than I think you did. Do you have a primary use case you're trying to enable with GPT

script that you're building today? You're like, you know what? This is the core. This problem right here is structured. It's understandable. It's understandable. It's cool. Like, do you have, or are you still in the exploration phase? Or do you think you've like narrowed in on like that, the core use case? You're like, we're going to make this work really well first. Or maybe, maybe all of it will work.

I'm kind of just, help us understand, like, where is it at? That's the super frustrating thing about AI right now. It's like, we're really in the exploration space. You know, it's like, we have this like type and expectation of AI is going to change the world. And like, we're not going to need programmers anymore. And, you know, all this crazy stuff or whatever. So there's like all this hype and expectation.

But then you look at the reality of what people are doing with AI right now. And it's so underwhelming. It's like chat box, chat with documents, customer support, adding dumb little like buttons to your app to like generate. So it's like, there's this crazy ideas of what we possibly could do with it. And then there's the reality of what we have right now is like this huge gap. And then like, what I struggle with when I talk to people is like, people can't think of anything in between. Everything's just like, can I have magic? So this like kind of the the gamble that i'm

doing with gpt script right now is that like no i don't actually have a specific use case i'm going after which is kind of scary to me because it's like ai is so far off from being a usable technology to accomplish anything that like nobody can even figure out how to use it and that's what frustrates me is because like i see ai of this huge, and we have all these academic papers, but like, the theory and practice are so far away from each other. To me, it's like, I could totally use AI to drive Kubernetes and to make completely self-healing infrastructure and drive infrastructure and do all that, right? But then I look at the current state of this technology, and it's way too far away from that. And it's like, the crap that people are doing right now is just so far

off from being reasonable and so i see that the opportunities with gpt script to build the simple primitives because it's like i have a lot of respect for like languages like go go is a language that you can see the authors had a lot of experience of basically being able to say what's not important. They can basically trim down to a very basic language that's like, this is what you actually need. And it's like, we don't have anything near that in the AI space. Nobody knows what's important and what's not. And so it's just a bunch of garbage or whatever. So it's like, can we actually take it and boil down to some very simple repeatable patterns and primitives so that we can make the stuff, you know, efficient and useful. And so, so it's like,

we're very much in this exploration space of like, I'm just constantly frustrated on the, you know, the state of where things are right now. It's just like, it's so far from actually being usable. It's like when people are like, can we have self-driving infrastructure and that stuff? And it's like, eh, not really. We don't have the patterns and approaches to do that. Sorry, I'm just going completely ranting. But like my personal belief is that like GPT-4 as it currently is, is smarter and more powerful than we know what to do with. Like our problem right now is leveraging AI. It's not AI getting smarter. Sure, it'll get smarter. Like it's kind of a factor of just putting more money and resources into it. It'll get more powerful. But like even what we have today, we just don't know how to use

it. So that's a very interesting challenge to me is like, so can I use this to drive infrastructure? But like in order for me to do that, I have to focus completely on like AI. Like, can I just solve this within its own right and then get it to the point that like, can I go back to now start tackling these problems? No, I mean, that was great. I'm really so curious to hear, like you talked about the limitations and then you talked about like our challenges are leveraging. What is it that you think is impeding our ability to actually get value out of what we have? Because I very much recognize and agree with the sentiment that you're sending is that we have these new interesting capabilities, but the way that they're surfacing are on what seem to be pretty mundane use cases, things that we were already pretty okay at doing in the first place. And so I'm curious to understand, what do you think the actual limitation is? Where are the challenges in actually getting the leverage? Yeah, I mean, so I think the limitations we have is we've got kind of the academic side,

which has this theory of how things will work. And then we have like the practitioners, the engineers. There's way too much of a gap there in terms of like understanding each side. Like, it's not really technology. A lot of it has much more to do about mindset, you know? And like, that's what I love about paradigm shifts and things like that. It's like, it's not that we didn't have the technology

as the way we approach the problem. And so what I see right now is like the whole engineering kind of traditional engineers, people who are like developers today, they are kind of stuck within their mindset of how they accomplish a problem today. Like when I put GPT script out there, it was like, this is natural language programming. That was a kind of a sensational title of like, we're going to replace programming

and this is how you're going to do it. And so immediately all engineers just say, no, that's stupid. You'll never be able to do this. And they're right. They're absolutely right. There's no reason for me to have AI do the equivalent of like what a lot of programmers are doing today. They look at like programming as like, well, programming is the task I'm currently doing. Like I'm typing this code. You can never do this. But if you take a step back, it's like, no, like you have some end goal. You're trying to get to some goal. How do I achieve that same end goal with this new technology, the way that I'm going to get to that end goal is going to be a completely different path. Right now, I think the thing is, is like the people who know how to solve problems and like

kind of do the day to day, you know, stuff are kind of stuck in a certain mindset of like, well, this is what programming is, this is how you do things. With programming, I'm trying to accomplish something, right? And so can I accomplish that same end goal, but use AI and then hopefully the way of doing it with AI is faster. It just basically enables you to do it quicker. Not necessarily it operates faster, but you can build it faster. With all these technology trends, they're really just about like accelerating the pipeline. It's like just getting from A to B faster. So an example of like a paradigm shift or whatever is like, if you look at the AI world,

Jupyter notebooks are really popular. So if you take a Jupyter notebook and you give it to like a traditional full stack developer or whatever, they're going to look at it and say like, this is garbage, it's complete garbage. And then like, if you give notebooks to any DevOps person, they're just like, like, this is garbage. Like you have to run like a VM or it's like all these security issues. Like how do I manage this? But like a Jupyter notebook is so ridiculously

efficient for data scientists and they are programming. It's in a completely different way. So it's like, it's in a completely different way so it's like it's a completely different take or or paradigm on how to accomplish something that's totally valid but like completely disconnected from one side of the world you know so it's like that's like our problem with ai is like how do i find the approach of how do i solve problems but i think right now people are just kind of too stuck in their own ways. And there's just a massive divide between the academic world and the visionaries of AI and the practitioners, people who get things done. There's so much going on in this space, actually. And there's so many people trying to introduce the near paradigms while the world is moving beneath us on the foundation side so

i guess i'm curious because i mean i've been looking at gpt scripts the releases and sort of like the choices you made maybe can you talk about like where do you see gpt script maybe goes a little bit further because right now we can see at multi models multi models i guess yeah there's a little bit of a decomposable tools you can do and yeah i can see at multi-models, I guess, there's a little bit of decomposable tools you can do. And I can see there's like certain like design choices you're making. So I guess the people you have in mind are people that are developers and you are helping them be able to build things faster, but also have some level of control in the tools, right?

Do you think there is something in the horizon for them to build like much more complicated, like how complicated of an app you want them to be able to accomplish? And what are the things that you think are primitives that are very important down the line? Yeah, so like the two things that I'm working on right now, because it's like we have the basic composability

of like you can create what effectively is like a function which calls another function and has input and output so there's a couple areas where there's like obviously some big gaps you know so it's like right now one of the big challenges with ai is this idea of the context window you can only have so much in the context window and that looks a lot like this kind of like stack it's like your stack can only get so big. So we have like the general issue of like kind of how do I deal with data? You could kind of almost look at like heap. That's kind of like one big area of my focus right now is like, how do I get it to efficiently interact with data,

not just existing data sources, that's actually pretty easy, but it's more like building up ephemeral context and stuff like that, because you want to create kind of workflows where it's like, grab this information, grab this information, and then, you know, correlate that information and do something with it. So that immediately kind of blows out the context window. Fundamentally, it's no different than, you know, it's like, you know, back in the day, we were programming with, you know, kilobytes of memory, and we had to fool around with memory tricks and stuff like that. So it's a solvable problem. So that's one area with GPT script that it's like really lacking is like interacting with data in a kind of an ephemeral session,

something would be like a heap. The second is like, I have like, kind of like a lambda function. Now, what are my triggers? Like, how do I invoke this? You know, so if I've created these little functions, how do I stitch them together? Like Lambda is a decent example of like, you know, like how do I do that? But like, but kind of like what would be natural to kind of what I would build for these flows.

When you think about the vast majority of applications are really just operating on data. Like they're just reading and writing data. Like that's all Kubernetes is. That's kind of the beauty of what Kubernetes is. It's like what we do with Kubernetes. We say, okay, we'll just represent all of my data as CRDs or custom resources.

And so now I can just operate on data in a standard pattern, but the data reflects some external state, some external system, right? So it's like, how do I get AI doing that? So it's like, I can represent external systems as just pieces of data. And then AI is just interacting with them. So those are like kind of the two things I'm focusing on. And like, I have a very strong opinion of kind of like less is more. It's like, I don't like to add new

features. I think more, it's just confusing. I like, I love languages that you can keep the whole language in your head. You don't have to go to reference, you know, so like Go is a great one. C++ is not, you know. And so those are kind of like the guiding principles of where I'm going with GPT-Script, what we're working on. I'm curious, like Tim and I chat as we always do. And one of the things that came across my mind was like do you envision GPT script replacing a piece of infrastructure? What you were just describing sounds like a significantly more intelligent air flow. In the sense that I need to download some data, I need to process some data, I want to get to an answer.

In that description, it sounds like an easier-to-use task runner for data workflows is one interpretation of what you said. So I'm kind of curious if you imagine, is this going to evolve to replacing infrastructure? Is it a layer over existing infrastructure? How do you think about it? I'm pretty bullish in that I think where we are right now is kind of like the equivalent of punch cards in mainframes. It's like we started with computing of doing mathematical calculations because it was more efficient to get a computer to do that. And then somehow we turned those mathematical calculations, you know, 50 years later into iPhones, right? Like that technology evolves so much that like somehow we, we abstracted that

away into like math turned into, you know, playing games on a phone. It's amazing. So where we'll go with AI is I think it'll have that same evolution that eventually it phone it's amazing so where we'll go with ai is i think it'll have that same evolution that eventually it'll get to the point where we can pretty much do all systems with it but like i'm talking like 20 30 years out or things are moving so quickly maybe this is only a decade out so i do think we'll get to that point but i think the short term obvious thing right now is like if we were going to say like acorn labs would be like go to market or something like that it's like and i'm not a very good business person so don't this is probably what we would do but like the very obvious thing right now is like task automation

for more like knowledge worker type thing you know so it's like hey i want to grab this data i want to analyze it because that's what a lot of people do in their jobs today is like, just make them more efficient and basically, you know, hooking up to different data sources and whatnot. Like I'm trying to generate sales leads or something like that. Like, how do I automate those tasks really easily? You could do that very easily with natural language because it's pretty much like grab this data and then give a natural language description of how to interpret that data and then, you know, save it somewhere else. And then, you know, so you just start stitching together. So that's a very like short term, obvious thing that we can do is just basically task

automation in like kind of like that knowledge worker area. So that's where I actually think like probably where we're like, we'll see this first, but like just because my background in infrastructure is like, I think that like the patterns, we can get them so good that they actually work better for reconciliation. One of the things that's interesting is AI handles errors much better. One of the difficult things in orchestration is always having to handle the errors, but AI will actually handle it much better. Because as interfaces change, for example, it doesn't really matter as much. The exact data structures don't have to perfectly match. They just have to have the same logical representation. So it's like, they're actually significantly more adaptive as other systems

change. There's one pattern that I think is very interesting. What I see with AI, that's very hard to do with traditional orchestration today is whenever you're orchestrating, let's say in like Kubernetes is like, if I have a system which creates another resource, so I have a, you know, reconciliation loop. So it's like a deployment creates a pod, right? So I have a resource which creates another resource within an orchestration system, like that created resource has to be owned by the creator. If a human goes in and changes that, or another system goes in and changes it, it messes up the whole system. So orchestration today has this clear idea of ownership and who touches what fields, like Kubernetes has this whole concept of field manager and stuff. Those problems, it's basically a three-way diff, become significantly easier with AI.

Because you can have like, here's my intent, and it produces the data structure. And then a human can go in and screw with that. But as long as it matches the original intent, then the AI doesn't need to modify it. So it's like, you can actually be working on shared resources much more efficiently with AI because the logic is more fuzzy. It's like that's fuzziness is really hard to do with like concrete data structures and stuff.

All right. So we're going to be going into what we call the spicy future. Spicy future. And so as you know, we've been doing this for some time now. What do you believe is going to be the future? What should people, everybody do that they're not doing yet? So I think, you know, there was a very popular headline that went around the ceo of nvidia was like we'll no longer need programmers i think he's like 100 incorrect

the exact opposite i really think ai is going to actually increase the number of programmers that we'll need because basically what ai is going to do is increase the scope of who can be a programmer such that we will have significantly more people creating systems that need significant support from the other systems. So it's like all the traditional jobs that exist today will most likely not be replaced. A lot of the existing stuff, there's no reason to rewrite it or do it in a new way with AI, but you need to continue to maintain it. So it's like, I really think AI is going to do the exact opposite to the engineering field of, it's the same way that AWS was like, people were like, oh, IT is dead. And it just made IT bigger. And it made IT more powerful.

Like, I think AI is going to increase the engineering space there will be more engineers there will be more ai engineers will become a completely new discipline i really don't think engineers today have anything to fear about ai most likely your job will just become more valuable so i don't know if it's a hot take but i mean i think it's it's against core narrative right like i think if you listen to the mainstream take of ai doomerism it said ai is here to take your jobs and we'll be automated away and i yeah so i think the thing is like the pattern you can see over and over again in history is like basically as we get more efficient at things we just use more of it ai will make people more efficient such that their

output is greater and then you'll want more of them This is one of the things it's like in it is like, they always want to like make it more efficient, but we never really, unless we go through these big bus or whatever, we don't decrease headcount really ever. It's like, we just kind of produce more and the headcounts tend to grow anyway. So it's like, nobody has anything like engineering specifically. I do think there are huge realms of jobs that can be impacted, but not engineering. I mean, and that theme when Jensen mentioned programming and we're thinking about programming,

I think this term programming is going to fundamentally change probably the next 10 years or more, right? I mean, it's been changing the last a lot. And so English as natural language seems to be your take like hey everybody should be able to write english everybody should be able to program to some degree and maybe so that that means everybody can start creating things right because everybody can just describe but is yeah so i think the percentage of people who can program will just get significantly larger yes yes so to a certain degree he's right in that it's like well

okay well if today let's say you know one percent of the world programs with code but you know another three percent program with natural language where are you going to have your kid like are you going to tell your kid to go after the smaller market or the larger market this is the thing is like i don't think't think it's like the people who enjoy C, for example, like low level, I love fiddling bits and do kernel level programming. Those people are not going to like suddenly become prompt engineers.

Like there's people who are attracted to certain domains and like those domains are going to continue to exist. And I think they'll just continue to actually, all the domains will most likely continue to grow because they're all supporting things. Like I don't think we're going to destroy the previous layers cobalt has been around and nobody's going to replace cobalt in fact we're most likely just going to start having ai program cobalt like that's most likely the future instead of replacing cobalt it's just like just get ai to maintain it yeah the infrastructure it just keeps living forever

right just keep layering on. That's like a famous saying in our group for a long time. But I guess maybe, do you think English language is going to be the only sort of like programming interface for the future programmers?

Or do we need some other language? Right now, looking at GPT script, it's early, obviously, but you have sections to have some composability. But you're actually marrying scripting, like existing programming languages, with your English scripting.

People have been trying to figure out, do we need a new compiler? Do we need some kind of ways to have more structure? Do we need a complete new set? Do you think there's going to be another almost like the python have have you know days like i think there's between python and english is there something we need to actually introduce to even make it more powerful as a like maybe is that one of my opinions right now is i actually think Python is holding back AI. I think we need to get Python out of AI.

Well, there's two different sides of AI, inference and training. It can stay in a training side. The inference side, Python shouldn't be there. But no, I don't think, you know, it's like you just look at like the evolution of it. You know, it's like every time we create a new evolution of technology in the language space, it's like it's really no different of like when we went from assembly to C or from C to Java or Java to JavaScript interpreted languages. So I think as we go up the stack, it gets wider because each language gets a little easier, but it never seems to replace the one below it. I would actually like to know this, like

historically over time, has the number of C programmers really changed? Has it gone significantly up or down? I don't really know, but it's still a dominant language. So it's like, we've never managed to kind of get rid of a dominant language. So it's like, I don't think English will be the only way. I think we're going to open up a whole different realm of writing applications, you know? So it's like what we have today, web applications, you know, am I going to replace Next.js with AI? I doubt it. I just don't like, is there a really big reason for that? But like, you know, AI can definitely replace

like site builders, like Wix and those types of things. But I think there's going to be like a whole new realm. You know, it's like, it's hard to say what it will be because like you know it's like iPhone apps didn't exist whatever 20 years ago and like now everyone's building mobile apps I think there is a future for like what AI applications will be but it doesn't even make sense that like why would they even have um you know a website or uh you know like they might be just completely text driven or just, you know, have more human interfaces to them. So it's like, I don't even know what the interface to an AI app is right now. So it's like where we're going right now is kind of like, well,

let's make it as close to as possible to the human. So that's why I'm interested in like, as I'm programming AI, well, I should be writing English or, you know, this works in any language, actually, like 26 languages or something like that. But like, I should be writing English or, you know, this works in any language, actually like 26 languages or something like that. But like, I should be writing English because that's the closest thing to the human right now is like, that's how I'm communicating. I think the future is just extremely unknown. And anybody who tells you they have any idea of where it's going, it's like just full of crap. It's like, there's no way to predict where any of this is going, but like, it's hard to believe that it's going to kind of stop. It's like, there's no way to predict where any of this is going, but like, it's hard to believe that it's going to kind of stop. It's like, we've definitely stumbled onto something very powerful.

Like the LLM is just a very powerful tool. Once we figure out how to use it, it'll, it'll enable something. Awesome. So we got so much great thoughts from you. Last question is how do people check out or use gpt scripts and where should people find you on social yeah so um just go to gpt script.ai that will eventually just redirect to our repo which is github repo right now it's very simple there's like you can just read the readme there or we have a docs website it's pretty simple to read read it through it's like brew install it's just a cli so it's easy to get up

and going. That's the project, the main project we're pushing right now. Acorn Labs is the company, we're at acorn.io. And then you can follow me on Twitter. My handle is I build the cloud. And I pretty much just rant and complain about everything there. Amazing. Thank you so much, Jarenaron it's been a pleasure