Increadibly exicted to announce Rothko to the world! For me, it’s been a labour of love that started three years ago when I co-authored the first large-scale paper on AI swarms with @james_k_he. During this research, James and I noticed that when language models interact at scale, they mimic fundamental characteristics of human collective behaviour without being explicitly instructed to do so. These findings suggested something exciting: pre-training bakes an inherent understanding of human group dynamics into language models. Despite our excitement, we also noticed that RLHF works against LLMs’ emergent simulation capabilities, compressing the nuances we observe in human behaviour and pulling models towards a population mode. Hence, we faced an interesting question: can we use and enhance language models’ pretraining knowledge to accurately simulate human societies? Rothko represents the culmination of these efforts. It works like a decision model trained on signals we’ve extracted from a large corpus of human behavioural observations, including anonymised social media information, transaction histories, unstructured interviews, and more. By getting rid of RLHF post-training, moving from the domain of natural language to the domain of decisions, and training on behavioural data, we’ve seen that Rothko dramatically outperforms SOTA LLMs at accurately simulating individuals and groups. This finding underpins our work at AS and the value our simulation platform provides to clients. Looking ahead, I see a world where large decision models like Rothko play a fundamental role in scaling innovation, protecting sovereignty, and aligning AI swarms with human values. I find these applications incredibly exciting, and I look forward to scaling Rothko’s capabilities with the awesome team at AS!