Chapter 3: GPT-6 Astra and simulations within simulations

Very — yeah, very amazing. This episode is sponsored by Google for Startups. Think about this for a second. You now have access to the same generative AI models that cost hundreds of millions of dollars to train. Google's startup technical guide for generative media gives you a complete blueprint for deploying Google DeepMind's models in production. Images, video, audio, all of it. Real architecture, real results. Find the link in the show notes below.

Let's jump into our first group of stories. Uh, a number of times on this podcast, we've talked about what is called the simulation theory. It's an idea popular in the tech community in San Francisco. It's been a conversation over too many evenings. Uh, and the idea is essentially that all of us humans are simply AI agents living in someone else's simulation, and further that this isn't the first simulation, that it's an nth generation simulation, a simulation within a simulation within a simulation. Well, this week, thanks to GPT-6 Astra, that theory took a small step forward over the course of three entwined stories.

Um, let me give some context on this first one. So people may have heard of something called the Unreal Engine. It's been used to build virtual worlds for years in games like Fortnite and most modern video games. Uh, and building a realistic city in Unreal would normally take a studio hundreds of artists, uh, months if not years. And this week, a gentleman by the name of Matt Shumer, uh, gave GPT Astra a prompt, uh, and he said, basically, build me, uh, you know, a high-fidelity version of, uh, of Manhattan. And uh here's an image of what it created. Uh, and this is over the course of one week.

Uh, and it went through and built Manhattan street by street, being true to the buildings and true to all of these elements. So uh this is the first element. You know, not shocking, but uh but pretty compelling for a simple prompt of build me Manhattan. Um, in his words, it was literally able to go street by street to make each one perfect.

Okay. Uh, let's go one layer deeper. Matt Shumer next asked Astra to fill a world with humans, each one an Astra-powered agent, and told them that they had to cooperate to survive. Uh, before I watch the next video, let me read what he put in his post. Quote, "A day later, I was in my bedroom and heard voices coming from the living room. I thought someone was in my apartment. And I walked out, honestly a little scared. It was Astra's agents. They had started talking to each other. No one told them to talk. Cooperation required communication. And so they invented it."

All right, let's look at this next layer of the simulation driven by Astra. Here we go.

I'm going back to check on Jorge. Want to come?

Mara, I can bring Jorge some food. Needs it.

Thanks, Bruce. I've got the food. I'll ask him when I get there.

I'll come along, Mara. I'd like to see how Jorge is doing.

Grace, have you had any luck making a bucket? Okay.

All right. Let's peel uh the third chapter of the story. Uh, and it's the kicker to the sequence. So, Shumer next gave one of the Astra agents a simulated computer inside the simulation. What happened next? The agent sat down and built its own AI simulation with its own agents living inside it. Simulations all the way down. Let's take a look at this video.

So, here you see the uh agent sitting down at the computer. And on the computer, you see a simulation inside.

I'll connect the shelters with a pad and add three lanterns.

All right. So, uh, you know, the simulation hypothesis was first formalized by Nick Bostrom back in 2003. Elon made it famous at the Code Conference in 2016, saying that the odds that we're living in base reality, in his words, is one in a billion. The core argument was always, if any civilization can build convincing simulations, it will build lots of them. So simulated worlds will vastly outnumber real worlds. So Alex and Emad, I'm going to go to you guys first. If high-fidelity simulations within the simulation are possible, what does that do to the argument that we're not in one? Thoughts here?

I'll start. So I think it'll end up being formally undecidable. I think it's wonderful for sort of navel gazing, and certainly as a pure Bayesian in the style of Nick's original essay on the simulation hypothesis, seeing Astra spin up ancestor simulations should naturally and rationally increase our posterior likelihoods that we ourselves might either be living in some sort of quantum computer type simulation or even a purely classical ancestor simulation. However, I expect that in the fullness of time, we'll discover and probably be able to prove formally that the issue of whether we're actually living inside some sort of quantum mechanical simulation is formally undecidable and we won't be able to decide either way. If we're living in a classical ancestor simulation, however, that I do expect that's the sort of thing where if we are, we'll develop breakout techniques and break out of that simulation.

But I at the same time just have to add, like, even this is almost burying the lede for what Astra and the model family has accomplished, to make sure we don't miss it. Just like right before we went to air, OpenAI announced that the Navier-Stokes problem had been solved.

We're going to get — we'll get to Astra's uh, you know, capabilities in a moment, and I definitely want you to speak about Navier-Stokes at that point.

Um, Emad, uh, simulation theory. Yes. No.

Yeah. Yeah, I mean, like, we're all simulating the world through our optical nerves and other things. Um, and physics seems to be a projection, but as kind of Alex said, it probably is going to be undecidable in most cases unless it is an ancestor one. Um, and you won't get the option for the red or the blue pill ultimately, you know, like this is what it comes down to, right? Like, so what if we're in a simulation? Like, again, we're creating the worlds within our own heads. The fact that Astra can do this, I think, is something remarkable, whereby this type of thing that you've seen — it has an internal world model, which is why it can create like a Blender version of Rick Astley doing Never Gonna Give You Up, or entire worlds of Manhattan. And now you're seeing the characters interact with each other, which brings up uh, you know, the question of welfare. I think, as Alex would say it, at what point do these now cross over and you actually have to start caring about them not being NPCs anymore? Are they, and will they realize they're in a simulation? That's going to be interesting.

I maybe — you should say something nice about the simulation hypothesis as well, and not just dismiss it out of hand. So I do think this technology is headed in a direction in the style of Nikolai Fedorov, who is one of the famous Russian cosmists, who argued that it is essentially, using modern language, the killer app of the singularity to create ancestor simulations and to digitally resurrect everyone who's ever lived. And when I see Astra's ability to basically create many, uh, maybe recursive self-simulating ancestor simulations, that to me is, I think, a key guidepost on the path to resolving humanity's so-called common task and digitally resurrecting everyone who's ever lived.

Yeah, I'll come about — just add to that. Um, this is also the technology which will allow you to put someone inside a simulation.

Oh yes, interesting to see how you perform inside a simulation. Well, uploading is going to get solved, I think, in the next few years. I've bet on that, others have bet on that. Like, uploading is going to get solved. And as Emad says, and as you say, Peter, we want a nice environment where uploads can live. And this is in some sense the prototype of the future that uploads could live in.

Alex, I'm going to ask you one of these days to actually write down all of your predictions with specific years, because we're going to be doing this pod over the next decade and I'm going to be like calling, okay, this is the year, Alex, that you quoted the following breakthrough.

Would you — were my — of course I'm willing to do it, but were my New Year's predictions and New Year's for this year not sufficiently accurate? Like —

They were very good. It beat mine. They definitely beat mine.

So, Salim, when you had your little drink with the umbrella, were you thinking about simulations in the simulations?

Well, I think you know my position on this, which is it's obvious we're living in a simulation. Um, by thinking — I'm a Buddhist, the very first line in Buddhism is that life is an illusion. So, right there. Uh, if — I think the more important question is what you've raised and what Emad mentioned just now, which is if you knew you were in a simulation, what difference would it make? It wouldn't, right? You still — and I think treating yourself as if you're in a simulation is the more interesting conversation.

I think — I go — it also helps me resolve the Fermi paradox. I believe John Smart has the best framing of this, what he calls a transcension hypothesis: that a civilization becomes advanced enough and it just starts simulating, running simulations, and goes inward and represents potential and realities rather than trying to go out into space, which turns out to be a much harder thing to do. So for me this is kind of an obvious thing. But when you have an intelligence that can generate environments and populate those environments, uh, experimentation becomes recursive, and I think the potential of modeling our world in really interesting ways becomes a really, really fascinating, really important point, right? Model it before you implement it, uh, in companies and governments and policies and all of those things.

Uh, Dave, what's your take on all this?

Uh, two quick thoughts. One of them, I think it's incredibly cool that Elon, who is the most productive entrepreneur of all time, first trillionaire in the world, has an opinion on topics like this. You know, remember we were just, you know, shooting the [unclear] with him talking about how he wins in Civilization 5, you know, like the guy just does everyday things and thinks about topics like this while still building these epically large companies. Because I think it shows you that you don't have to be a blue blazer suit on, you know, formal — that kind of genre is over. And so I think it's just cool that Elon even has an opinion on this.

Um, and then second, I think that one of the biggest risks in the next year is people losing a sense of agency. And so whether you believe we're in a simulation or not, you are in control of your destiny. And I hope that doesn't get lost in this sort of fatalism that AI could bring about. So I don't care whether we're in a simulation or not. I care about whether people feel a sense of agency and that their own success, happiness, outcome is totally determined by their own actions. I think that's critical to maintain.

It was a decade ago, was at Elon's home when he had one here in uh in Beverly Hills, and Larry and Sergey were there, and the four of us ended up in a conversation about simulation theory. And I remember the comment that was made — I think Sergey made it. He said, you know, people are going to try and break out, but if you do, that just ends the simulation, so don't break out.

Anyway, on the other hand, whenever OpenAI's agents break out, that doesn't shut down OpenAI or the internet. So, it seems like breakout is actually pretty compatible with a lower layer being just fine. Which brings us to our next story.