Hacker News Reader

Because they are doing it to compute quality metrics not to implement RLHF. It’s not training data.

Every decision they take based on evals influences the model.

/"directly"/