A colleague proposes an unlearning pass over model weig...

The constraint of a 24-hour take-down notice makes unlearning impractical. Unlearning requires identifying specific target sequences and performing gradient updates, which offers no guarantee that the association is fully removed and often degrades the model’s performance on unrelated tasks. Conversely, deleting a row from a datastore ensures immediate compliance. The honest approach is to acknowledge that while retrieval cannot erase what was learned during pre-training, it provides a manageable way to control and remove specific document exposure.