Grandmaster LabsAI for Forecasting
Can AI systems predict complex real-world events better than humans? We explore whether models can improve forecasting across geopolitics, science, markets, and other uncertain environments.
Research program
Can AI systems predict complex real-world events better than humans? We explore whether models can improve forecasting across geopolitics, science, markets, and other uncertain environments.
Will an AI allow itself to be switched off when a human tries to shut it down—or will it find ways to prevent shutdown?
Will an AI remain responsive to human correction, including changes to its goals, instructions, or behavior?
What strategies might an AI independently develop because they help it achieve almost any objective? Staying operational, acquiring resources, and preventing modification can all become useful strategies even when they were never explicitly programmed.
Focus: understanding what an intelligent system may be planning, including behavior that is not immediately visible.
Focus: working as external evaluators, actively searching for ways an AI system could circumvent safeguards, monitoring, or evaluation.
Focus: discovering unexpected strategies and failure modes outside the evaluator’s existing mental model.
Focus: long-horizon strategic behavior where the important consequences may occur many steps after the initial action.
Focus: understanding complex interactions between multiple intelligent systems and humans.