Not really? It's a lot of work, a multi-week project, but reading a couple hundred word speech can be done in 5 minutes, following a checklist in hand, probably 10 minutes. Times 12 categories, and 80 years of history, that's a lot of time 160 hours, a working month. A lot of effort but humanely doable.

That's true, but assumes you have the checklist of what data to analyze in hand when you start out. If you only decide after the fact which familial relationships have interesting trends, you'd have to start over again. It seems more reasonable to start by transcribing everything to text, annotating that text, and then running a lot of scripting to automatically query that data.

They probably just used the speech database that the Academy hosts? https://aaspeechesdb.oscars.org/

Ok, obviously it's _doable_, but is it worth it? Using LLMs for this purpose would have been significantly cheaper, easier and with the right configuration just as reliable. Once the setup works, you could extend the analysis to all kinds of other interesting branches without having to look at a single speech by hand.

I would even go so far as to say that _not_ using LLMs for this task would be fairly odd, unless I'm missing something or the author really enjoys a month of manually classifying documents to write an interesting and well-written but not exceedingly outstanding article.

Some people like doing stuff.

Of course. It's just my opinion that this task would be perceived by most to be fairly repetitive and unfulfilling, but if the author thinks otherwise, great for him.