1) easier to modify prompts than retrain a model 2) we simply figured out prompting first - In Context Learning is about 3-4 years old at this point, whereas w

1) easier to modify prompts than retrain a model 2) we simply figured out prompting first - In Context Learning is about 3-4 years old at this point, whereas we are only just beginning to figure out LoRAs and representation engineering, which could encode this behavior much more succinctly but can have tradeoffs in terms of amount of information encoded (you are basically making a preemptive call on what to attend to instead of letting the "full attention" just run as designed