According to CoinMeta, Artificial Societies released a human behavior prediction model, Rothko, in London. This model allows companies to simulate the reactions of thousands of consumers, industry insiders, or policymakers to test the impact of new products, advertisements, and major decisions. Rothko modifies the architecture based on a pre-trained model to directly predict human choices and is trained using anonymous social data, transaction records, travel trajectories, expert interviews, and surveys to accurately reflect the true preferences and differences among different groups of people. The company used data from 2,796 real respondents to generate 68,280 survey answers. In comparisons with nine general-purpose large models, Rothko had the answer distribution closest to that of real people, with an average error of 8 percentage points, while the best-performing general-purpose model, Qwen3.8, had an error of 31 percentage points. After changing the wording of the questions, Rothko had a prediction error of only 1.4 percentage points for changes in answers, with the second-best model having an error of 6.2 percentage points.