AGI: The Path Forward – Jason Warner & Eiso Kant, Poolside

How many people here know what Poolside is and does? Anyone? Anyone? Yeah. So let's talk about that real quickly. Poolside exists to close the gap between models and human intelligence. That's literally it. That's what we're here to go do.

We're building our own models from scratch to do this. We're based on the idea, two and a half years ago, that we thought next token prediction was an amazing technological breakthrough, but it needed to be paired with reinforcement learning, really, to make that leap.

So that's what we've been doing for the past two and a half years. So we're on our second generation of models now, Malibu Agent, and instead of kind of like walking you through some slides and all that, we just thought maybe, I don't know, let's kind of show you what we're doing here.

So Iso, are you there?

I got you, Jason.

So as I said, you were supposed to see him today, but there's, I don't know, our airline system kind of works sometimes, maybe. So he's stuck in California. But we thought we'd just walk you kind of through some demos here today.

So what you're looking at here is a very modern programming language that the government uses to run all the world's critical infrastructure called ADA. Anyone familiar with ADA? Yes. Yes. OK, so everyone I saw put their hands up for ADA either has no hair or gray hair like me.

So that should tell you what's going on here. So Iso, why don't we figure out what's going on with this code base here?

Well, let's start asking what the code base is about.

That's great. And what you're seeing here is obviously our assistant in Visual Studio Code, backed by Poolside Agent, a model we trained from scratch using our proprietary techniques. And you can see what's going on here, kind of the stuff you expect from an agent.

And obviously, the form factors of all of these things are going to change a couple of times over the next couple of years. But people seem to like VS Code. So we're going to show you this demo here today.

So you can see from this, it kind of went through, told you what this code base is all about. But

these things run in our satellites. And I don't know anything about ADA, but I do know a lot about a couple of other programming languages. So Iso, what do we want to do here? Why don't we see what this thing might look like in Rust?

Let's do it. Let's ask it, convert this database to Rust.

So obviously, you're going to see what's going on here. Again, if you guys have used other tools, you're not going to expect too much of the difference for what's happening here, except that, again, we're backed by our own model.

We're not using OpenAI. We're not using Anthropic. This is Poolside. And Poolside is a bottom-to-top stack that is,right now, if no one's touched it, and I know no one in this room has touched this unless you work for a three-letter agency, a defense contractor, or you've sent missiles somewhere that we're not going to talk about in this session.

Because that's where we're working. We're working in high-consequence code environments for the last year inside the government and the defense sector, as you can see from this demo. So what you see here is it's kind of going through, doing the conversions.

What you see in the middle pane is something that we built to kind of show you as the streams come through all the different changes that are happening. One of the tricky parts about working inside the defense sector and things like that is you can't have an agent that's just going to run around and do stuff.

I mean, I can't walk into half of these buildings. You can't give an agent access to these data sources and just say, hey, go nuts. You need to have theright permissions. You've got to actually really ratchet these things down to do things inside those environments that they feel comfortable with.

So where are we on this now? Is it trying to fix itself yet?

Yes. It wrote about 1,152 lines of code. And it just popped up a command, start and get tested.

Excuse me. So we see here all of the files on the left-hand side that it created. This is essentially our live diff view that's available. And as we see, it's currently starting to actually test it out.

So this is the part where we just sit here and watch this for three minutes, and I see nothing. No, what you see.

The good thing is that this is a very fast inference.

Yes.

So 1,100 lines of code.

Did it.

And tasks completed.

Do we know if this works yet?

Well, let's have a look. So it actually wrote some commands to test it. And when we check out the output of those, this actually looks pretty good.

Can we ask.

Can we verify that?

Run it.

Let's go verify it. So of course, our agent came back and gave a summary of what it did. But let's just ask how to run this.

OK. So I'm going to go open up. So it says, this is how I can run the ADA version, and this is how I can run the Rust version. Let's run the Rust version.

Perfect. Let's have a look here. Ooh, we might be hitting an actual.

An actual demo bug?

Let's have a look.

Let's see what happens.

Oh, no. No, no. Just warnings.

Just warnings.

Sadly.

Do we have an unwrap in there that we need to take care of? I heard that those things are dangerous.

Soright now, there's a repo. Let's hit help, see what we're able to do. So it looks like we have a set of commands. I'm going to be lazy. I'm going to copy-paste these queries. So create table users. OK, so far, so good.

Let's insert a record. OK, well, let's find out if it actually did its job. So let's start from users. OK, we've got a record here.

That's nice.

Now, I want to actually see if I use the up arrow. It doesn't actually allow me to cycle through commands. Let's ask it to add a feature.

Allows me to use the up arrow to cycle through. I think it will understand my intent here.

The one thing we know about Iso is he actually does know how to read and write, but he can't type. So all those errors that you're seeing in there, yeah.

So it looks like the agent's identified a package that we can use. Let's just quickly look here. Compare this to version one.

And it looks like it's adding a library called RustyLine and changing the files accordingly. It's currently built it. And it looks like the build output is successful. There's some warnings. We'll ask it to clean those up later on.

And let's now start to test it.

OK, apparently, it works. It wrote itself a little bash script to test the history. It's wrote itself a little final demo script. So let's let it. OK, and it gave us the summary. Well, now, how do I rerun this?

I do kind of know that.

Should know that. That was 30 seconds ago.

Let's build it. And let's run it again. OK, let's do a help. And, oh, yeah, that's the up arrow. It works.

Very nice.

Now, our models are just capable coding agents. They're capable in lots of areas of knowledge work. They're also emotionally intelligent. They're fun. They're great to write bedtime stories with for the kids. So I'm going to ask it to write me a poem about all these changes.

But that's just more for fun.

So as Iso was saying, this is just an interface into our platform. There's other interfaces into it if you're inside one of those organizations that has adopted Poolside. So this is the coding interface into it. But we also have other ways in which you can interact with it, web as well as an agent that you can download on your machine.

But yeah, we don't really tout the poem writing or the songwriting, though. I did send this to my wife to see. And I have been sending her love letters written by Poolside. So I kind of hope that she did not enter this session to know exactly how I've been doing that for the past six months.

But yeah, so this is kind of Poolside. This is what we've been up to. So as I said, Malibu Agent is our second generation. We've got a ton more compute coming online. And that's when we're training our next generation.

That is going to be the one that comes out publicly to everybody very early next year. We're going to have it behind our own API. It'll be on Amazon behind the Bedrock API. Anybody in the world who's building out any sort of, on one side, the engineering assistants, like the Cursors, Windsors, Cognitions, Replits of the world, you can use ours.

Or if you're building out on any other side of the fence, the Harveys, the Writers, the whatever applications of the world, there's going to be a fifth model out there that's going to be at that level that you can consume.

But we're dead set on doing this and bringing this out to everybody in the world and kind of advancing that state of the art. And we're just going to keep pushing that out. So that's kind of who we are.

And you can find out very little more at our website since we don't put much out there. But Iso, anything else you want to say before you try to go make your flight this time, please?

So I would say that it's been a pretty incredible journey for the last two and a half years of starting entirely from scratch and now building to a place where we see our models have grown up to become increasingly more intelligent.

And the kind of missing ingredient that we had was compute. And now that's unlocked for us. And we have a large number of over 40,000 GB300s coming online. We see how we can start scaling up some of those models to get even further in their level of capabilities in software development and other types of long-horizon knowledge work.

What I think is exciting about this conference and this audience is of all the work that's happening of evolving the form factor. Right now, what we looked at was this asynchronous way of operating with agents. But Jason and I have agents running that are doing tasks for hours.

And I think in the near future, we can see a world where they're able to start doing tasks in days in the coming years. And so I think the interface will continue to change. We're really focused on the fundamentals, building intelligence, and being able to scale up and serve it.

And it's why we go full vertical. It's why we go from our multi-gigawatt campus in West Texas, where we're building our data centers for a team building out models. And the interface that you saw today is just our version of an expression.

But I think this audience is going to do an incredible job of building lots of better versions of how to express using that intelligence into actually valuable, economically valuable work.

Couldn't have said it better. Can't wait to see what you guys build on this in the future when it's publicly available. And if anyone really does want to build a data center campus, we are hiring for that. It is weird to be putting shovels in ground again, like we did in the '90s and early 2000s.

But that's what you got to do to scale intelligence these days.

I would make one other non-scheduled statement.

Oh, no.

If you're going to be OK with this one, Jason.

As our models are getting more capable, we'd love to also see who wants to build with them. Right now, the vast majority of companies that are doing additional reinforcement learning and fine-tuning on top of models are doing it on what I would considerright now the best-in-class open source models, the Qwen and FMEs and Minimaxes of the world.

And we'd like to start figuring out how we can partner with you with our models, anywhere from any checkpoint early on to where we are today, for you to be building closer together with us on top of things.

We haven't really figured out the approach to it yet. But I think since we have this audience, it's not a bad place to put it out there. And so definitely reach out to us. We think the world till date was built by intelligence.

The world in the future is being built on top of intelligence. And so it'd be a great way to partner.

Well, thanks, Iso. Thanks, everybody here. And now we do have five minutes left. I don't know if we're supposed to take questions, but I'm happy to. So if anyone does. But if not, I'm just going to go that way.

What was that?

Is Iso AI.

Sort of. I mean, I think of him that way. Hey, here's a fun story. Here's how I met Iso. I like to tell this story because Iso's a fun dude. I met Iso because it started with a failed acquisition at GitHub.

So back when I joined GitHub in 2017 as a CTO, I wanted to take GitHub from a kind of collaborative code host with open source Bent and turn it into an end-to-end software development platform infused by intelligence. And so you know the products that we launched from '17 on: GitHub Actions, packages, alerts, notifications, eventually code spaces.

And then Copilot was the last thing that the office of the CTO did before I left, with Nat Friedman, Ugo de Moore, and a couple of other folks inside there. But Iso, in 2017, when I joined, he had working code completion before the Transformer architecture had landed fully.

He had on LSTMs. And so I quickly tried to acquire his company. And he just said no. He just said no to me. But that was a long, drawn-out process talking about what we thought neural networks were going to mean for the world.

And so during that process, which was a lengthy one, we became really good friends. And we'd stayed in close contact over the years. And then '22 rolled around. Obviously, ChatGPT comes out, Anthropic's out. And we kind of saw the end game at play.

And we said, do we jump back in or not? And of course, yes, we jumped back in. But I like to tell that story about how he just kept saying no to me. And I just kept asking him questions.

And eventually, he said yes, we should found a company. Because by the way, when I asked him if we should do this, he said, oh, goddamn, no. That was his exact words. He's like, no, we should just learn how to paint and sail.

But here we are.

It's been a gradual journey to get it, Jason. I think the reason we ended up doing this is because of our opinionated view on what it was going to take to build more capable intelligence. And the first 18 months of this company, obsessing and focusing on reinforcement learning combined with LLMs felt like one of the most contrarian opinions in the world.

But I think today it's absolutely not. And it's super exciting to see the progress that's continuing to make. In the coming years, we're going to see the world that started in completions and went to Chat and is now at Agentic increasingly approach more autonomous.

And where all of it is stemming effectively from a combination of bringing highly capable models that are constantly evolving together with real-world problems. And I think what we're starting to see now is we're entering these kind of awkward teenage years ahead of AGI, where everybody in this room who is building out incredible companies and applications is bridging this gap of what it really takes to make intelligence that, in its raw form, actually be valuable.

And we want to be a small, humble part of that. We've got a lot of work still ahead of us. The team is growing. But hopefully, what you've seen today is what our customers and enterprises have been having access to and seeing for a while is that we're hard at work at really pushing those capabilities.

We also want to make sure we make them available to build together with others.

Well, that's it. Thanks, everybody.