I'm not convinced that ship has sailed. Really it is just a few players doing it, and mostly Google. The other truth is that this has only been happening for a few years, so it is not hard to "turn back", if you will (because I don't think it has sailed, I think people are abusing the term and using the miscommunication to drive hype and revenue). There aren't many players in the LLM space and people outside this are still using the term correctly. The problem is that the public is picking up the phrase and actively defending it. Which is rather weird, don't you think? That non-researchers are happy to argue about definitions with researchers?

But why I push against this, is because the misuse actually hinders our ability to interpret them. Generalization is one of the most important tasks in ML, and even more so when we start to discuss alignment. What is happening is the erosion of our metrics, which we've already become substantially reliant upon. Google is well known to make generalization claims by doing exactly this -- the JFT -> ImageNet task is quite common. Our field is already extremely noisy, but this just makes it more so. The abuse makes gains seem much larger than they actually are, but in reality all that is happening is that we are decreasing the temperature on the density function.