What happens when an AI living inside a simulated world gets the tools to build another world of its own? That question suddenly feels a lot less like science fiction after developers began experimenting with OpenAI’s newly released GPT-6 Astra, creating autonomous virtual environments populated by AI agents and, in one striking case, a simulation inside an existing simulation.
AI entrepreneur Matt Shumer showed one Astra-powered agent using a computer placed inside its simulated environment to construct another simulated world populated by its own agents. The experiment does not prove that humanity lives inside a simulation, but it offers a remarkably tangible demonstration of nested virtual worlds—the kind of possibility that has been central to philosopher Nick Bostrom’s famous simulation argument.
More AI Sparking Unexpected Debates: A game system built on code and rules turning into a much bigger public conversation isn’t unique to simulation theory. For more, read our report on GTA 6’s girlfriend system exploding online, and whether Jason and Lucia’s romance is really a problem for real-life relationships.
GPT-6 Astra’s New AI Capabilities Are Making Autonomous Simulations Possible
OpenAI launched GPT-6 Astra on September 3, 2026, describing it as “the world’s most intelligent and aligned model.”
Unlike a traditional chatbot that primarily responds to prompts with text, Astra is designed for complex, multi-step computer-based work. OpenAI says the model represents a major step forward in computer use, software engineering, browsing, cybersecurity, scientific work and other professional tasks.
The significance becomes clearer when the model is given access to an actual computer.
Astra can interact with software, create documents and presentations, work with spreadsheets, develop software and operate development environments. It is also designed to handle longer-running tasks with greater autonomy.
That means developers can do something considerably more ambitious than asking an AI to describe a fictional universe.
They can give the AI the tools to build one.
An agent can write code, execute it, test the result, identify problems and make changes. Once that agent is placed inside an artificial environment and given access to a computer capable of running another simulation, the possibility of a nested digital world becomes technically demonstrable.
And that is exactly where Matt Shumer’s experiment gets wild.
More Rival AI Model Launches: Astra isn’t the only frontier model making headlines this week. For more, read our report on Google’s Gemini 3.8 Flash landing just as Alphabet dodges a major AdX breakup ruling.
Matt Shumer’s ‘Simulation Inside a Simulation’ Experiment Explained
Shumer began with a simulated environment already inhabited by Astra-powered AI agents.
Then he introduced a computer capable of running another simulation into that environment.
According to Shumer, one of the AI agents sat down, used the computer and built a simulation of its own—with additional agents living inside it.
“So uh, simulation theory might be real.
I dropped a sim computer into the simulation my Astra agents live in. One agent sat down and built a simulation of his own, with its own agents living inside.
Simulations all the way down.”
The phrase “Simulations all the way down” immediately captured the imagination of people watching the demonstration.
The reason is obvious: the experiment resembles a miniature version of one of the most famous ideas surrounding simulated reality. A group of beings exists inside an artificial environment, those beings eventually gain access to the technology needed to create another artificial environment, and that new environment contains its own inhabitants.
Shumer later shared a timelapse showing the Astra agent building the nested simulation.
“Timelapse of the Astra agent building the simulation inside the simulation:”
The demonstration is striking because this wasn’t simply an AI generating a fictional story about characters who live in a simulation. The agent was operating within an environment, accessing a computer and using code to actually construct another digital environment.
That distinction is what makes the experiment so fascinating.
More Things Turning Out to Have Hidden Layers: A world quietly containing another world inside it isn’t just an AI story. For more, read our report on Denmark’s biggest Viking silver hoard, hidden under a backyard until a terrace project exposed it.
Did GPT-6 Astra Actually Discover It Was Living in a Simulation?
Not quite—and this is where the viral claim needs some serious context.
Shumer himself acknowledged that the experiment was deliberately constructed to make the result possible.
“The agents are powered by Astra, and had full autonomy to make their simulation whatever they wanted it to be (via code). Of course, this is a somewhat leading setup… by giving the agents a computer that can run a simulation, obviously, they are going to do that. But the agent still made its own choices in designing the sim. It’s still crazy to think we’re at a point where models can pull this off. Simulation theory doesn’t sound so crazy after all.”
That qualification changes how the experiment should be interpreted.
The AI agent did not suddenly realise it was trapped inside a simulation, investigate its surroundings and independently conclude that it should create another reality.
A human operator intentionally placed a simulation-capable computer inside the artificial environment.
The important achievement was what happened next: once the Astra-powered agent had the necessary capability, it independently used the computer and code to create another simulated environment and make choices about how that simulation would work.
So, this is evidence of an emerging AI capability, not evidence that our universe is simulated.
More Viral Claims That Need a Second Look: A striking demo going viral is one thing — whether it actually proves what people online are claiming is another. For more, read our look at Sam Altman’s claim that 38,000 ChatGPT queries equal one almond’s water use, and whether anyone can actually prove it.
How Nested AI Simulations Could Eventually Work
The idea becomes easier to understand when the layers are placed side by side:
| Level | What Happens |
| 1 | Humans create an artificial environment |
| 2 | AI agents operate inside that environment |
| 3 | An AI agent receives access to simulation-capable computing |
| 4 | The AI creates another simulated environment |
| 5 | The new environment contains AI agents of its own |
In simplified form, that becomes:
Human-created reality → AI agents → AI-created simulation → AI-created agents
Humans already create simulations in many forms, including video games, physics simulations and virtual environments. What changes here is that entities inside those environments can potentially be given the tools to construct environments of their own.
Shumer’s experiment demonstrates at least one level of this nesting.
It does not establish that such a chain can continue forever.
However, as AI models become increasingly capable of writing software, operating computers and maintaining complicated systems, building increasingly sophisticated virtual environments becomes less difficult.
And that leads directly into another experiment attracting attention.
More Human Activity Reshaping a System: Layers building on top of an existing system, changing what it can do, isn’t just a software story. For more, read our report on Everest Base Camp’s new environmental warning as human activity adds heat to the glacier.
Another AI Simulation Experiment Gives Characters Memories, Goals and Physical Locations
Another developer known online as Parzival demonstrated a different approach with Argos, described as a stateful simulation engine designed to maintain an underlying world while using an AI video model to render what happens inside it.
The demonstration goes beyond simply creating a convincing AI-generated video.
Parzival said he was able to send a text message to a character inside the simulated world. Instead of instantly replying, the character had to physically locate her phone before responding.
“I texted someone inside a simulation.
She had to go upstairs and find her phone before she could text me back.”
Parzival explained that the video model itself did not decide what happened. Underneath the visuals was a persistent world containing characters, objects, locations and systems.
The characters have their own cognition, memories and goals. They choose actions, while the simulation determines what actually happens.
The character’s phone has a location. Messages have delivery and read states. Typing a response takes time. If the phone is moved before she reaches it, the phone is gone—even if she remembers leaving it there.
Parzival described the broader ambition as creating “a world that keeps growing while its characters inhabit it—with agents writing and maintaining the simulation’s systems.”
He also said the project will continue to be developed publicly, with future possibilities including connecting the system to SMS so people can text characters inside the simulated world or potentially send an impulse into a character’s thoughts and see whether the character follows it.
The key idea is simple but powerful: this isn’t merely a video that looks like a world.
It is an attempt to create a world that continues to exist underneath the video.
More Stories Where the Full Picture Isn’t What First Appeared Online: A viral clip is never the whole story — sometimes what’s actually going on underneath takes a company statement, and a dispute, to sort out. For more, read our report on College HUNKS movers accused of mocking a 79-year-old woman during a Pittsburgh move, and the company apology that followed.
Why Persistent AI Worlds Are Different From AI-Generated Videos
A conventional AI-generated video can show a character walking upstairs, finding a phone and answering a message.
But once the video ends, there is not necessarily a persistent digital world underneath those frames.
A stateful simulation works differently.
The underlying system can maintain objects, locations, rules and changing states. A phone can remain in one location. A character can remember where she previously left it. A message can have a delivery state. Actions can take time.
The visual output therefore becomes a representation of what is happening inside the simulation rather than the simulation itself.
That distinction could become increasingly important as AI-generated environments become more sophisticated.
More Staged Visuals Getting Scrutinized: Whether what viewers see is spontaneous or carefully constructed is a question that comes up well outside AI video. For more, read our report on Naomi Osaka’s US Open match sparking influencer backlash over ring lights and a photoshoot.
How AI Video Models Are Becoming Part of the Simulation Stack
Parzival said Argos uses H3 through Fal Director to render its simulated world.
Fal’s documentation describes H3 Max Director as a system designed for continuous real-time video streams while maintaining characters, settings and story continuity.
This provides a potential visual layer for persistent simulations.
The underlying simulation determines the state of the world, while the video-generation system attempts to depict those changing conditions.
In theory, that separates two different functions:
World state: Characters, objects, locations, rules, memories and interactions.
Visual presentation: The AI-generated video showing what is happening inside that world.
Instead of asking an AI model to invent every frame independently, developers can maintain a structured digital world and use generative models to render its current state.
That starts looking much more like an interactive digital environment than a conventional AI video.
More Ideas That Take On a Life Beyond Their Origin: A concept escaping its original, narrower context and becoming something the internet runs with on its own isn’t unique to philosophy papers. For more, read our story on how YouTube’s lawn-care star became the face of the Toy Freaks controversy.
Nick Bostrom’s Simulation Argument Explained
All of this naturally brings the conversation back to philosopher Nick Bostrom and his 2003 paper, Are You Living in a Computer Simulation?
Bostrom did not claim to have scientific proof that humanity lives inside a simulation.
Instead, he presented what became known as the simulation argument, built around a trilemma.
In simplified terms, at least one of three propositions is likely to be true:
- Human civilization is very unlikely to reach a posthuman stage capable of running enormous numbers of ancestor simulations.
- Civilizations that do reach that stage are unlikely to run large numbers of ancestor simulations.
- If advanced civilizations reach that stage and routinely run huge numbers of ancestor simulations, simulated observers could vastly outnumber observers in the original reality—making it statistically plausible that we ourselves are simulated.
The third possibility is the one that has fascinated people for years.
But there is a crucial distinction: Bostrom’s argument is philosophical and probabilistic, not a scientific demonstration that the universe is a computer simulation.
More Vivid Demos Making an Abstract Idea Feel Real: Sometimes a single striking demonstration does more to make a claim feel tangible than years of argument. For more, read our look at Sam Altman’s almond-and-ChatGPT water-use claim and whether it actually holds up.
Why GPT-6 Astra Makes the Simulation Argument Feel More Tangible
The recent demonstrations do not solve Bostrom’s argument.
What they do is demonstrate a technological premise that once felt much more distant.
AI systems can now operate computers, write code, construct virtual environments and work through extended tasks without a human manually controlling every individual action.
Those AI agents can also be placed inside artificial environments.
Give them the tools necessary to build another environment, and an AI agent can potentially create a simulation inside that simulation.
That doesn’t mean the same process is happening in our physical universe.
But it does mean that one of the ideas frequently used in thought experiments about simulated civilizations is no longer purely hypothetical from a computer-science perspective.
A sufficiently capable AI agent can be placed inside a simulated environment and given the ability to construct another one.
More on What Happens When Systems Outgrow Their Original Purpose: A system quietly taking on a life of its own, for better or worse, is a thread running through a lot of tech and AI stories right now. For more, read our report on Google’s Gemini 3.8 Flash landing right as Alphabet dodges a major AdX breakup ruling.
The Bigger AI Question: What Happens When Agents Maintain Their Own Worlds?
The more consequential question may ultimately have little to do with proving simulation theory.
Instead, it may be about what happens when AI agents become capable of independently creating, maintaining and modifying increasingly complex virtual environments.
OpenAI says GPT-6 Astra was designed for complex, multi-step work and computer interaction, while its safety documentation describes the model as reaching a new level of cybersecurity capability.
That growing ability to operate tools autonomously is exactly what makes demonstrations such as Shumer’s possible.
Today, these simulated environments still rely on human-designed infrastructure and carefully selected capabilities.
Future systems could potentially receive broader access to development tools, persistent memory, computing resources and the ability to maintain their environments over much longer periods.
That could make the distinction between an AI-generated scene and an AI-maintained world increasingly important.
A generated video stops when the rendering stops.
A persistent simulation can continue changing.
And a sufficiently advanced simulation could potentially contain agents capable of creating new software, environments and simulations themselves.
That is where the concept of “simulations all the way down” becomes genuinely intriguing.
No, These AI Experiments Do Not Prove We Live in a Simulation
For all the viral excitement, the most important conclusion remains straightforward: nothing demonstrated by these developers establishes that our universe is simulated.
Shumer’s own explanation makes clear that his nested-simulation experiment was deliberately designed to give the AI agents the ability to create simulations.
Parzival’s Argos demonstration is likewise an engineered software system created using existing AI and rendering technologies.
Neither experiment provides an observation showing that the physical universe is running inside a computer.
Neither demonstrates that human consciousness is being simulated by an external civilization.
What the experiments actually demonstrate is still remarkable.
AI systems are becoming capable of creating virtual environments where persistent objects, rules, agents, memories and interactions can be represented computationally.
And when those AI systems are placed inside simulated environments and given the necessary tools, they can create additional simulations.
That is enough to make Bostrom’s thought experiment feel far more tangible to a generation watching AI capabilities accelerate in real time.
For now, however, “simulations all the way down” remains a striking demonstration of what computers can create—not proof of what reality actually is.
Disclaimer and Sources
This article is based on research into the technological demonstrations and information provided by the following sources, while distinguishing verified developments from philosophical speculation. The AI experiments described above do not provide scientific proof that humanity or the universe exists inside a computer simulation. Claims about consciousness, subjective experience or the ultimate nature of reality should not be inferred from the behavior of AI agents in these demonstrations.
Sources
- Matt Shumer — Nested Simulation Experiment on X
- Matt Shumer — Timelapse of the Nested Simulation on X
- Parzival — Argos Persistent Simulation Demonstration on X
- BroBible — Original Report on the GPT-6 Astra Simulation Experiments
Featured Image Credit: Shantanu Kumar on Pexels
Leave a Reply