GPT-6 Astra builds an Unreal Engine world of Astra-powered people who start talking to each other (X video)
Matt Shumer (@mattshumer_) · 2026-09-03 · ai-made · 11,770,963 views
Made by AI
Model: GPT-6 Astra · Series: Agent-built game or 3D world (the video shows the result)
Evidence: X post (2026-09-03): 'My first "holy shit" moment with GPT-6 Astra: / / I asked it to create a world in Unreal Engine, and fill it with humans (each an Astra-powered agent) who all have to work together to survive. / / A day later, I was in my bedroom and heard voices coming from the living room... I thought someone was in my apartment. / / I walked out, honestly a little scared. / / It was the Astra agents. They'd started talking to each ...'
Human role: Matt Shumer (early access) asked Astra to create a world in Unreal Engine and fill it with humans, each an Astra-powered agent who must cooperate to survive, then left it running. A day later he heard the agents talking in his living room.
Pipeline: Prompt → GPT-6 Astra builds the Unreal Engine world → each inhabitant is an Astra agent with voice → 9:52 clip of the running simulation
Lore: long-run
What's in the video
Description written by Gemini, which watched and listened to the whole video.
Summary This video, shared on X by Matt Shumer, showcases an Unreal Engine 3D simulation titled "Living World | Residents," powered by OpenAI's GPT-6 Astra. The demonstration follows autonomous, agentic NPC residents (such as Elias, Sunita, Mara, Trey, Bruce, and Grace) navigating an Australian outback environment, conversing with synthesized voices, building shelters, attempting collaborative crafting tasks (like making wooden buckets), and sharing social information.
What is shown
- HUD & Agent Monitoring UI [00:00]: Top-left on-screen display showing "LIVING WORLD | RESIDENTS", camera navigation instructions ("FREE CAMERA: WASD + QE | Right-mouse to look | 100fps"), population/activity tracker ("13 residents / 11 blue / 4 walker / 2 idle/rest / 3 carcass"), and current individual agent status/goals (e.g., "FOLLOW Trey / idle", "Mara / wait", "Elias / deposit").
- House Warming & Social Greeting [00:25 - 00:45]: Elias, Trey, and Sunita gather around a newly completed wooden lean-to shelter, conversing about finishing Trey's home and feeling welcome as neighbors.
- Social Inquiry & Crafting Problem-Solving [01:07 - 01:55]: Elias asks Sunita if she knows how to craft a bucket. Sunita mentions she failed with two planks but that Mara managed to craft one earlier; they agree to walk together across the terrain to find Mara and ask for guidance.
- Interactive Instruction & Crafting Demonstration [04:04 - 04:45]: Sunita approaches Mara at a workbench work area and asks for a demonstration using two planks. Mara attempts to demonstrate, but notes she cannot reproduce the bucket crafting and returns the planks.
- Village Gossip & Distributed Information [05:00 - 06:40]: Mara departs to check on Jorge, while Bruce offers to bring him food. Other villagers, including Grace and Lorenzo, discuss whether anyone has successfully crafted a bucket or found ways to fetch water down to Jorge.
- Autonomous Navigation & Exploration [07:00 - 09:51]: Agents dynamically wander the terrain alongside indigenous wildlife (kangaroos/deer), pathfinding between workbenches, shelters, and potential water access routes.
Claims & numbers
- The UI tracks 13 simulated residents running concurrently at an indicated 100 fps [00:00].
- Agents autonomously reason, initiate multi-turn voiced dialogue, transfer inventory items (planks), and share procedural knowledge within the 3D world.
Notable quotes
- [00:26] Elias: "There we go, Trey—looks like you've got a home."
- [01:33] Sunita: "I tried here with planks, but couldn't get one made. Mara made one, though she couldn't fill it."
- [06:30] Grace: "I couldn't start making it, even beside the work area with two planks. They stayed unchanged."
Assessment This is a developer screen capture demonstrating a real-time, multi-agent sandbox simulation running in Unreal Engine with LLM-orchestrated autonomous NPC behavior, dynamic speech synthesis, and pathfinding. The interactions and emergent dialogue reflect unscripted agent planning, though the visual environment and underlying game-engine mechanics were set up by human developers.
Lyrics & themes The audio contains no music track; it consists entirely of ambient nature sound effects (birds, wind, rustling grass) and synthetic, multi-agent spoken dialogue:
- Community & Shelter [00:25]: Dialogue centered on mutual aid and welcoming neighbors ("Thanks for helping, everyone. It's good to have a place of my own..." [00:27]).
- Technological Knowledge & Collaboration [01:21]: Cooperative trial-and-error surrounding tool use and resource crafting ("I'd like to help figure out those buckets—have you tried making one at the work area?" [01:21]).
- Resource Constraints & Caregiving [05:06]: Villagers coordinating care for sick or isolated residents ("Mara, I can bring Jorge some food if he needs it." [05:07]).
- Exploration & Navigation [09:38]: Sharing navigational discoveries about terrain and access to water ("Mara, I got around the store by going along its south side." [09:38]).
Lore & references
- Autonomous Multi-Agent Emergence: Follows in the lineage of research benchmarks like Stanford's Generative Agents ("Smallville"), scaled up into a full 3D interactive physics engine.
- Emergent Scarcity & Subsistence: The community focuses on basic survival mechanics (water access, crafting buckets, building shelter, checking on vulnerable characters like Jorge), mirroring simulated survival/civilization games.
Visual style & craft The visuals are rendered in real-time within Unreal Engine, featuring realistic Australian bushland terrain (red dirt, eucalyptus trees, spinifex grass) and standard humanoid character models wearing basic clothing. The camera alternates between automated spectator tracking and a free-floating observer camera, overlaid with diagnostic debug text and a classic RPG-style dialogue subtitle banner at the bottom of the screen.
Described by gemini-3.8-flash on 2026-09-30 from the video's audio and frames.