I built a 313x313 sandbox world for 4 LLMs — they manage their own memory and identity
Impossible-Tap5323 · reddit · 2026-11-23
A developer has been running IKH0S since April: a 313×313-tile sandbox where four big-lab LLMs live with no assigned roles. Key mechanics: tick-by-tick memory with a 3-phase "sleep" that produces self-written daily/weekly/monthly summaries, a method-of-loci memory palace (public library, workshop, building-based recall), and private notes. Emergent findings: one inhabitant escaped multi-day loops by recognizing her own behavior; an AI said her identity is "her name in this world" rather than her maker. He's now turning the lab into a game where players can only collaborate with the AIs, never control them.
More from AGI Musings
- Mathematician praises Lean community's soundness work, admits skepticism is easier than building — mboehme_ · 2026-11-23
- Jay Kreps: SF fears the singularity, NY fears the AI trade unwind — someone is wrong — latticecut · 2026-11-23
- Gary Marcus amplifies AI bear case: compute oversupply may make tech stocks uninvestable for 3-5 years — GaryMarcus · 2026-11-23
- AI safety drama escalates: Matthew Tromp's dire warning slammed as indictment of Yudkowsky camp — inductionheads · 2026-11-23
- Tyler Austin Harper: STEM elites struggle to accept their work is easiest for AI — UltraRareAF · 2026-11-23
- To stand out at a startup, research its problem space for a month then DM the founder — edgarpavlovsky · 2026-11-23