Feel free to show otherwise

I think they mean that it is likely to heavily depend upon the task.

Decoder models are good at generation of language, and of course they're going to do well where that counts. But if you want to do typical NER+Relation extraction and then normalize to an in-house dictionary of 10 million IDs? You can't do that as effectively with GPT-4. You need a local model (right now).

There are a lot of things that GPT-4 doesn't touch in terms of data domain, so for many projects, yes local (typically encoder) models are 1) way faster 2) way" cheaper and 3) have better metrics.

Of course, the capability could* be there with a large decoder model, but if it's a task that needs your specific (large amount of) data, you have to make something locally.

No, you need to show the model is better on this narrow task, not just assert it is because it’s a great general LLM. It’s quite possible you’re correct but just saying GPT is the best, prove me wrong, reeks of a VC or AI bandwagoner.

The OP is about classification. I linked to benchmarks showing it is indeed much better on the kind of task outlined in the OP as well as several other NLP tasks.

I believe "prove me wrong" is more than appropriate here.