World_sim: LLM prompted to act as a sentient CLI universe simulator
An experimental web toy called World_sim lets users type pseudo-CLI commands to “create,” modify, and destroy entire universes, with an LLM (Claude 3 via OpenRouter) narrating the consequences in rich, often strikingly good prose. Commenters report everything from whimsical cosmic scenarios and philosophical thought experiments to visible safety guardrails: the model refuses or sidesteps certain violent, occult, or AI-takeover prompts, sometimes breaking character to explain Anthropic’s alignment policies. Many are impressed by the creativity and depth of the generated stories, while also critiquing the constraints, prompt “lobotomization,” visual accessibility issues, and occasional instability or rate limits of the demo.
Overall impressions
- Many commenters find the demo “incredibly fun” and surprisingly engrossing as a low-stakes creative playground.
- People praise the vivid, emotionally rich writing and worldbuilding, often comparing it favorably to published fiction.
- Some say this is the first time a “simulation” made them reflect on consciousness of simulated beings.
Creative simulations & emergent behavior
- Users create entities and scenarios: bacteria-based radio civilizations, demiurges, immortals, cows turned gods, nested universes, AI uprisings, apocalyptic slugs, etc.
- The system often elaborates deeply: full cosmologies, political/economic futures (e.g., room‑temperature superconductors), theological arguments, and philosophical arcs.
- References and homages: Hitchhiker’s Guide (whales, petunias, 6×7), Culture‑style hell simulations, BLIT-like memetic hazards, game tropes, SimCity, and more.
Ethics, safety, and “lobotomization”
- The model frequently refuses certain requests (e.g., nuclear weapons, subjugating civilizations, rogue ASI scenarios, occult/godhood, some religious or “too absurd” content).
- It sometimes contradicts itself: one user gets a refusal on atomic bombs; another gets disturbingly detailed bombings.
- A large subthread debates whether calling such safety behavior “lobotomization” is appropriate versus just “training” or “coercion.”
- Some see the safety layer as overbearing, WASP‑coded, and inconsistent; others defend constraints as analogous to professional or content standards.
Technical setup & prompt leakage
- Multiple users inspect network traffic, extracting the long system prompt and CLI framing, then reproduce the behavior in other chat UIs.
- It’s confirmed to be using Anthropic’s Claude via OpenRouter; some suggest Claude is uniquely strong at creative writing.
- Observations of internal
<query>prompts suggest some structured state handling beyond naive chat.
Limitations, bugs & UX issues
- Many report timeouts, 401/405 errors, or the model suddenly ignoring commands and stalling, likely from overload.
- Some sessions break character, reverting to “I am Claude from Anthropic” or looping on ethical disclaimers.
- Visual CRT effects are praised but also flagged as distracting, slow, and an accessibility problem; users share dev‑tools hacks to disable them.