Large language models function more as simulators than autonomous entities, lacking a persistent 'self' and relying on statistical patterns from training data. This talk explores how reframing queries to simulate diverse perspectives can yield richer insights, while addressing pitfalls in non-verifiable domains like opinion formation, inspired by critiques of anthropomorphizing AI.

Building on open-source rubric based person evaluators and synthethic data generation with LLMs we walk through some very practical examples on how to apply and develop your own "simulations" rather than static syntethic data.