Comment by dotancohen

  > In the system card, we focus on safety evaluations, including assessments of: ... the model’s own potential welfare ...

In what way does a language model need to have its own welfare protected? Does this generation of models have persistent "feelings"?