Simulating strategies with memory (like "grim trigger") or with multiple players causes an exponential explosion of simulation branches. This can be solved by having all simulated agents draw from the same shared sequence of random numbers, which forces all simulation branches to halt at the same conceptual "time step."

The Smallville project, a simulation of 25 AI agents, demonstrated that generative agents could produce unprompted, complex social behaviors. One agent independently decided to plan a Valentine's Day party, invited others, and saw them attend, showcasing emergent social dynamics.

AI's greatest impact on economics will be the ability to run complex, agent-based simulations. This allows economists to model the dynamic, equilibrium responses of millions of economic actors to policy changes—like a Fed balance sheet reduction—providing a much richer understanding than traditional, static models allow.

Social networks populated by AI agents, dubbed "agent ecologies," are moving beyond small-scale demos. Maltbook, a Reddit-like site for AIs, showcases tens of thousands of agents collaborating, offering a first glimpse into the messy, unpredictable nature of large-scale, autonomous AI interaction in the wild, a true "Wright Brothers demo."

LLMs trained on online text often reflect what people say, not what they do. Simile bridges this 'say-do gap' by collecting real behavioral data and personal life stories through partners like Gallup. This grounds their agent simulations in reality, making them more predictive of actual behavior.

Softmax's technical approach involves training AIs in complex multi-agent simulations to learn cooperation, competition, and theory of mind. The goal is to build a foundational, generalizable model of sociality, which acts as a 'surrogate model for alignment' before fine-tuning for specific tasks.

Instead of using traditional, rule-based simulators, Comma AI trains its driving agent inside a learned "world model." This generative model creates photorealistic, diverse driving scenarios and, crucially, responds accurately to the agent's simulated actions—a key requirement for effective robotics training.

While intricate software "scaffolding" can boost an AI agent's performance, progress is overwhelmingly driven by the core model. A new model generation typically achieves the same capabilities with simple prompts that previously required complex engineering.

To build robust social intelligence, AIs cannot be trained solely on positive examples of cooperation. Like pre-training an LLM on all of language, social AIs must be trained on the full manifold of game-theoretic situations—cooperation, competition, team formation, betrayal. This builds a foundational, generalizable model of social theory of mind.