We’re introducing Rothko, a foundation model of human behaviour designed from the ground up to simulate the world’s most consequential decisions.
Rothko uses a proprietary architecture and training corpus to dramatically outperform state-of-the-art LLMs at predicting how individuals and groups make decisions. As announced in our October evaluation blog, it produces more accurate predictions of top-line population distributions, more coherent representations of individuals’ belief systems, and more consistent outputs across repeated questions and stimuli. These improvements amount to billions of dollars when informing the highest-stakes corporate and sovereign strategies, and they unlock the future of modelling collective behaviour.
Our philosophy
To understand why we built Rothko, it’s helpful to think about how most labs train language models and why this makes them bad at predicting human behaviour. In general, researchers train LLMs in two steps: pretraining and post-training. Pretraining gives the model information about the world, and post-training teaches the model how to use it to answer questions and solve difficult problems.
This second step is essential for making AI helpful in day-to-day life. However, it encourages LLMs to produce common outputs that strip away nuances in human behaviour and make them act like coding savants rather than real people. At Artificial Societies, we need to figure out how to use and enhance these models’ base understanding of the world while making them behave like the individuals we care about in our simulations.
Rothko reflects more than three years of frontier behavioural science and AI research to do this effectively. We start with a pretrained model and change its architecture so that it produces decisions as outputs rather than natural language text. We then continue training the model on an enormous corpus of human behavioural data extracted from anonymised social media information, transaction histories, mobility traces, unstructured expert interviews, traditional surveys, and more.
We end up with a large decision model that takes information about an individual and their context and predicts their behaviour. The model understands the difference between what people say and what they do, and when we aggregate its respondent-level outputs, they reflect human collective behaviour substantially more accurately than standard LLMs.
Scaling innovation, sovereignty, and humanity
Building a foundation model of human behaviour has enormous value that we’re only just starting to uncover. One immediate application involves using it to predict purchase intent and public reactions to high-stakes corporate strategies. Another involves using the model to increase democracies’ resilience against misinformation, polarisation, and other cognitive attacks.
These applications are novel for decision models and represent some of the highest-value inference in the world. However, our frontier safety research suggests that Rothko’s most valuable application will be at the interface of human-agent collaboration, particularly as we explore new techniques to make swarms of super-intelligent models safe. As a result, we’re motivated to continue scaling Rothko’s capabilities so that humans can develop a better understanding of each other and collaborate more effectively with future technologies.