Chinese researchers dropped Light Society—a simulation framework running ONE BILLION LLM-powered agents simultaneously. Each agent has memory, personality traits, beliefs, and decision-making logic, all grounded in real World Values Survey data from 96k respondents.
The architecture is brutal efficiency:
- Mixture-of-models engine routes complex interactions to full LLMs, routine stuff to distilled surrogates (MLPs or tiny transformers)
- Prompt caching + compressed graph representations for social networks
- Vectorized batch processing + aggregated event execution
Result: Orders of magnitude cost reduction without killing behavioral fidelity.
They validated it with trust game simulations. Key finding: demographic patterns (education/social class → trust/reciprocity) sharpen as agent count scales up. Stochastic noise vanishes at billion-agent scale, revealing clean scaling laws. Repeated interactions spontaneously generate social norms.
This isn't toy research. It's the first framework that bridges high-fidelity LLM behavior with earth-scale populations. Traditional ABMs were too simplistic. Pure LLM sims maxed out at millions. Light Society hits billions by intelligently mixing model complexity based on interaction type.
Computational social science just got its first planetary-scale laboratory.
The architecture is brutal efficiency:
- Mixture-of-models engine routes complex interactions to full LLMs, routine stuff to distilled surrogates (MLPs or tiny transformers)
- Prompt caching + compressed graph representations for social networks
- Vectorized batch processing + aggregated event execution
Result: Orders of magnitude cost reduction without killing behavioral fidelity.
They validated it with trust game simulations. Key finding: demographic patterns (education/social class → trust/reciprocity) sharpen as agent count scales up. Stochastic noise vanishes at billion-agent scale, revealing clean scaling laws. Repeated interactions spontaneously generate social norms.
This isn't toy research. It's the first framework that bridges high-fidelity LLM behavior with earth-scale populations. Traditional ABMs were too simplistic. Pure LLM sims maxed out at millions. Light Society hits billions by intelligently mixing model complexity based on interaction type.
Computational social science just got its first planetary-scale laboratory.