TypeSafe AI's The Bitterest Lesson argues that choosing the right objective comes before improving a model's training data, computing budget or algorithm. A training process can steadily improve its score while optimizing something users do not need. The essay illustrates the gap with language models: continuing internet text and following a person's instructions are different objectives. Its argument puts product understanding upstream of model optimization, because deciding what useful behavior means requires understanding the people and workflow the model will serve.
