...what would happen if there were three political parties instead of two?
Or whether WWII was still inevitable in Germany without Hitler?
I do. But I don't just want to hear the answer — I want to see it play out.
No existing social simulation platform works as a real thought-experiment playground for these kinds of questions. After seeing Stanford's Generative Agents, I decided to build one myself: Civilizism.
To me, the answer is simple: the memory of individuals.
When agents own subjective memories — memories that act as the root hub of all their perceptions, decisions, and relationships, shaped by interactions and mutual experience — society emerges by itself. You don't program the society. You program the people well enough, and the society follows.
But mimicking individuals shouldn't come from abstraction, which is what every ML library defaults to. Everything in Civilizism is traceable from first principles.
"Just let AI roleplay everyone," you might say.
Here's the thing: the increasingly pressing question in AI stops being "what can AI do" and starts being "what should you let AI do." AI is creative and easy to use — but it's also expensive, slow, and unreliable by nature. It is not a Swiss knife for every question.
In Civilizism, LLMs are reserved for exactly two things: action generation and reflection — the moments where improvisation is the intended effect. Everything else — attention modeling, memory retrieval, belief clustering, personality initialization, perceptual scoring — is algorithmically designed from scratch.
That's STEME, BLOC, SHAPE, and DASP. All built in-house. None of them are wrappers.
Before a specialist can be introduced, you first need a firefighting generalist who knows everything well enough to make things work.
That's me.
I wanted a thought-experiment playground, so I built one. Civilizism needed a visualization layer, so I forced myself to learn it. My Student Council was breaking apart from administrative collapse, so I held it together with documented handbooks and institutional memory until the structure could stand on its own again. Now I'm doing the same thing somewhere it actually has to survive contact with a real product: at Phygtl, I designed and built a small AI system of my own — one LLM roleplays real users, working through the actual product under real-world scenarios, and a second LLM reads the resulting transcript and judges how the product held up. It runs nightly, gated by CI, no one watching it by hand. And when I found a gap in how the product understood people, I didn't just flag it — I proved it with data and got the fix adopted company-wide.
You need something? Name it. I can't promise perfection — but I can promise sufficiency, and faster than you'd expect.
Languages Python · R · SQL · Java · C++ · JS/HTML/CSS
Libraries NumPy · scikit-learn · sentence-transformers · Pydantic
Concepts System Architecture · Semantic NLP · Concurrent Multi-Agent Systems
LLM Integration · Modular Design
LinkedIn · timchen56789@gmail.com · resume on request