The technical aspects of training and tuning are trivial. GP is pointing out that you might not be able to get the model to succeed at the task as often for any number of reasons that you won't know before you actually train one.

Although I guess your point is that it's also cheap to train them, probably cheaper than doing this. But startups are started by social people, not technical people. Stuff like this will always be expensive for social people since they have to pay one of us to do it. YC interviews their CEOs from time to time, it's really clear that's how that works.