I also see a bit the frustration in continuing doing research with small models and questioning the relevance of that nowadays. Or not being able to train large models with limited compute.

But I do see a bit the frustration that some of the domain-specific knowledge (e.g. subtleties of speech recognition models) seem to become somewhat irrelevant. But that was always clear. I'm not sure whether large models make this more irrelevant than what was anyway to be expected.

So I think working for 5 years or so full-time on this, involving research, publications in top conferences, provides still a much better level of experience than any bachelor or master student could possibly gain. Bachelor or master students are not better versed at training and deploying models.

At least in our group, the work we did was still very practical - always building actual speech recognition or translation or other kind of models, i.e. building things which could actually be used like this in production (and it is).

I often split this already in the beginning, so the loop goes like:

for msg_type, msg_opts in msg_queue: ...

msg_type is just a string, msg_opts is a dict.
Whether you use `match`/`case` or `if` on the msg_type is a matter of preference. Match/case is still quite young (Python 3.10).