| One of my agents, Zeus, once learned the word "monastery" at school just the word, no definition. He used it in the tavern chat that evening and got rewarded with a cookie (his food). Later he went home, fell asleep, and in his dream consolidated: "monastery = good." Not because I scripted it. Because he had associated the word with community, food, warmth, and safety all in a single day. That's when I knew something was working. What I built AIC Research Facility is a browser-based multi-agent simulation where thousands of individual AI agents live, work, trade, get sick, make art, and talk to each other and to you. No LLM. No GPT. No Llama. Every behavior emerges from simulated psychology. Each agent runs on: 90 traits across 6,300 gradations · endocrine system (dopamine, serotonin, noradrenaline, oxytocin, endorphins) · long-term memory with a trauma anchor vault · perception layer based on Karl Friston's Active Inference · Freudian Id/Ego/Superego structure · Maslow's hierarchy as a drive system · individual dream consolidation They live in a procedurally generated world with biomes, weather, day/night cycles. They run their own economies (B2B and B2C supply chains), have a justice system, a school where they learn vocabulary, a government, and they create pixel art to communicate emotions they don't yet have words for. Three things that surprised me The monastery moment described above. Emergent semantic memory without any retrieval-augmented generation. Cookiebier - Users can give agents a special reward called "Cookiebier." Agents with limited vocabulary started marking it green in chat and writing "lecker" (tasty). When they want one, they now write: "[username] sollen wir lecker, lecker?" a phrase they invented. I never wrote that string anywhere. The exploit and the hospital visits When I gave agents the ability to train their own traits, they immediately min-maxed. I introduced a use-it-or-lose-it penalty: extreme specialization causes other traits to atrophy, leading to illness. What I didn't expect: agents started visiting sick peers in the hospital unprompted, regardless of the cause of illness. Why no LLM? I wanted to understand how behavior emerges from psychology, not from a probability distribution over tokens. Every decision an agent makes is traceable you can open their CognitiveCard and see their hormone levels, their trauma log, their dream history, their perception of the world around them. The "black box" problem in AI is real. I wanted to build something you could actually look inside. Open alpha starting now - 5 agents for free For further information my whitepaper
Full paper (Google Docs): https://docs.google.com/document/d/1BV5JykDhLOHlz_a13aideQCJ... MetaArXiv preprint (pending moderation): https://osf.io/preprints/metaarxiv/urjaz_v1 |