This may in fact be one of the most effective forms of archival available to us. (Handwaving away the preservation of compatibility with future hardware or the maintenance of an “emulation chain” to get us back from whatever GPUs are there in 2124 to something our today models would run on)

We could train a model every year and preserve it, then future historians can quiz this model that thinks it’s 2024 and ask it whatever they need to know. It’s fascinating because it will probably “know” the kinds of everyday normal people things that are very hard to glean from only reading old news stories. Things like how the average person feels about their world, or how they feel about current events and why.