Right that is a good distinction. Fair enough. Still stand that you could train a worse model depending on the task. Translation, Nuanced Classification are all instances where i've not seen bespoke models outright better than GPT-4. although, like i said it could still be good enough for speed, compute requirements.