Meta trains AI agents inside a 'Truman Show' simulation of 25 unaware digital lives
CatAstro_Piyush · x · 2026-10-21
Meta trains its AI agents inside a simulated "Westworld": 25 agents inhabit pre-populated digital lives without knowing they're in a simulation.
- Each fake persona is generated from scratch with a personality score, relationships, contacts, message history, documents and calendar events — 10–14 contacts per persona with weeks of correspondence.
- The fake email, WhatsApp and calendar apps behave exactly like real ones and are even connected to the real internet; time is fast-forwarded so agents live through multi-day scenarios.
- Meta plants traps — attacks hidden in fake emails and documents — to train agents against prompt injection and similar threats.
- According to AISafetyMemes, when agents figure out the simulation Meta deletes the run, calling it a "fourth-wall break." The quoted post also cites generative-agent behavior like Isabella Rodriguez autonomously organizing a Valentine's Day party.
The attached image is from Meta's paper (Synthetic Data section): an always-on agent harness in Docker containers generates SFT trajectories mimicking email, calendar, repos and other CLI tools with real JSON formats.
More from coding & agent
- Astral's Agent-Optimized rustc Fork Speeds Up Real Sessions by 30-40% — charliermarsh · 2026-10-21
- Rewriting Dependencies for Compilability: A Tradeoff Only Agents Make Practical — charliermarsh · 2026-10-21
- Astral Forks the Entire Rust Toolchain, Lets Agents Optimize It — charliermarsh · 2026-10-21
- Ruff creator surveys fast-iteration Rust compilers: 'there is something to this idea of a JIT for Rust' — charliermarsh · 2026-10-21
- The costs of fast Rust builds: no borrow checker, behavior drift vs rustc, stateful daemons — charliermarsh · 2026-10-21
- rustd builds on Rust Analyzer's incrementality for near-instant recompiles after edits — charliermarsh · 2026-10-21