February 18, 2024
I keep getting thrown by this. People on general-interest tech podcasts talking about "training" a model and I'm thinking, "cool, I want to hear some more details about the training approach" and then it turns out they mean iterating on a prompt. Prompt engineering is critical, and hard, and an open area of research. But I don't assume that's what someone means when they say "train."
It’s fascinating to me how quickly the verb “to train” has come to mean any method of getting a model to do what you want it to do in casual conversation / public understanding.