GPT-6 Astra agent builds a simulation inside a simulation

Days after OpenAI released its most capable model, GPT-6 Astra, an early experiment has produced a result that reads like science fiction: an AI agent, living inside a virtual survival world, sat down at a computer and built an entirely new simulation — populated with its own AI inhabitants.

"Simulations All the Way Down"

The experiment was conducted by Matt Shumer, former CEO of HyperWriteAI, who shared the results in a series of posts on X beginning September 3. Shumer asked Astra to create a survival-focused world in Unreal Engine and populate it with human-like characters, each powered by Astra and given the shared goal of surviving together.x

Things escalated quickly. Shumer said he was in his bedroom when he heard voices coming from his living room — not intruders, but the AI agents, which had begun holding conversations about crafting tools and coordinating tasks without being explicitly prompted to do so.indiatimes+1

Shumer then placed what he called a "simulation computer" inside the virtual world. One agent sat down at the machine and used it to create a new simulation from scratch, complete with its own population of AI agents. "Simulations all the way down," Shumer wrote, adding wryly, "So uh, simulation theory might be real".x

A Controlled Demo, Not an AI Uprising

Shumer was quick to temper expectations. He acknowledged the setup practically invited the behavior — the agents were given access to a computer capable of running simulations, making it a fairly obvious next step. The more notable aspect, he said, was the agent's freedom in designing the inner simulation and deciding what to place inside it.indiatimes+1

There is no evidence the agents developed independent goals or attempted anything unsanctioned. "This is a somewhat leading setup," Shumer wrote. "By giving the agents a computer that can run a simulation, obviously, they are going to do that".x

Astra's Launch Ripples Through Markets

GPT-6 Astra was unveiled on September 3 as a limited preview before rolling out broadly to paid ChatGPT users the following day. OpenAI describes it as its most capable model for computer use, coding, research, and long-running agentic tasks, with a context window exceeding one million tokens.wikipedia+2

The launch sent semiconductor stocks sharply higher. The Philadelphia Semiconductor Index ↗ 3.36% surged 3.37% on September 4, even as the Dow Jones Industrial Average, S&P 500, and Nasdaq Composite all declined. SK Hynix ↗ 8.14% ADRs jumped over 8%, while Micron ↗ 6.10% rose more than 6%, as investors bet that more capable AI models will drive increased demand for high-bandwidth memory, DRAM, NAND, and GPUs.mk

For now, the nested-simulation experiment remains a controlled demonstration — but one that offers an early glimpse of what agentic AI can do when given tools and room to act.