Post

lol weird line break i'm ee cummings

Nope!

exactly.

Even in the geo area you’re talking about 95% of records/archives are NOT digitised.

I'm a little reminded of a SF story I read many years ago, where a company branded their virtual reality technology as Reality™.

and “vast majority” is like. way over 90%. i volunteer to transcribe historical documents for the national archives because i can read cursive handwriting from the 18th & 19th centuries (some of the penmanship being not great lol) and the amount of untranscribed material is *huge*.

Let's keep it that way! Vellum lasts for 5,000 years. No URL can say the same.

Even if LLMs had all those texts, they couldn't be trusted with them. You have to understand the biases your sources have, and AI wouldn't. It would just straight use every source as fact.

While researching a civil war memorial I was able to identify a sword that the BPL had sitting in their archives for 100 years. History is one of the last fields that should be “worried about ai”

the concern should not be that AI will replace historians (it can't) but that the people who write the historians' paychecks come to think it can

Good luck!

I would love to hear how it turns out, this is an excellent exercise! 🙌

Exactly. So far, with our experiments, we're finding that most LLMs actually do a very good job if you specifically ask for diverse perspectives on historic events. I've even had the LLM label some "claims" as "untrue and harmful", which impressed the students.

Have had to argue with people in RL about this numerous times

Same reason I refuse to get rid of my LPs.

it's been a few years now but there were a couple major cable breach incidents that crippled the internet and i think more people could do with being reminded "no, you may in fact be forced to 56k or less if there is only one old cable carrying your data the long way around the planet"

I work with archival AV records, including from broadcasting, and let me tell you, finding contextual info online about things that have been broadcast on the radio, even in the last 40 years or so, is so incredibly difficult. There's just so little out there, because nobody cared to put it there.

They are telling the truth (as long as you adopt their definition of "human knowledge", which, while contestable, is IMHO reasonable)

The problem is that AI boosters, especially within academia, operate within a zone of teleological confidence that casually extrapolates from now to a magically exponential curve. So they are confident both that digitization and LLM interpretation will encompass everything, inevitably.

Also there is so much human knowledge that isn't written down, let alone digitized!

My literal self identity was changed by one guy in Albany translating 17th century Dutch colonial records by hand

Exactly.

Most human knowledge has not been digitized, true, but more and more digital content is invisible to web crawlers behind paywalls now. Why does prose by Copilot read like LinkedIn? Because MSFT owns LinkedIn and has a readymade source for training data.

nothing even counting the information locked inside human expertise

*Gil Scott-Heron voice*: "The past has not been digitized."

Just for my mom's intensive deep dive in genealogy, she ended up finding that out. Hard to get access from Texas! And she spoke with German archivists that had to gently let her down and tell her that they didn't have records from before 1945 any longer for obvious reasons.

Exactly: Traveling all over, years of research. Lots of money.

And on the flip side they cannot quantify or qualify uncertainty

Same goes for art, of course. All this "the world's knowledge and creativity" bollox is BOLLOX.

The ironic thing is that ‘AI’, in the sense of ML-based handwriting recognition, really is a revolutionary technology for historians! Just not in a way that’s going replace any of them!

Some of the sources I used to work with are catalogued in ledgers and "Findbücher"; you need to go to the archive to find whether they have anything relevant, catalogued in spidery handwriting and in haste in around 1808. Digital archival catalogues are still a desideratum, never mind digitizations!

Something that blew my mind when I visited the Library of Congress in 2009 was that it still had a physical card catalog that took up I can't even guess how much space. Wish I had thought of turning to it when a DDOS attack shut down their website during a visit in the 2010s—if it was still there.

When I visited the American Antiquarian Society in 2010, just 5% of its periodical holdings had been digitized. And they keep adding to the collection.

Unfortunately the grifters just see another market and will want to crank out shitty, inaccurate, worse-than-what-we-even-had imaging-to-text models. Fortunately ("fortunately"), there is zero grant money for this so at least that's a decent ward against this shit.

Sometimes quoting them with the note 'translations mine.' ::crumbles to dust:::

Id love to recreate some of those routes in flight simulators. I'm sure some clown will be along shortly to tell me how an AI can simply make that for me though, based on thinking AI is cool and knowing nothing about history or flight simulations

The kids can’t even read script.

This seems tenuous to me. AI historians may, for the moment, need the primates to digitise data for them. But they would no doubt claim that finding the patterns in it is still doing history.

Also, dates are like, a big part of the whole exercise, and LLMs are just subpar at feeding back accurate numbers.

+ @faineg.bsky.social the answer, basically, is yes

No, not always

And even for records that are digitized, they are frequently incomplete and need supplemental sources for interpretation that may not have been digitized yet

And in the case when they have been digitized, like say on the BNF’s Gallica system where everything is available in an OCR’d PDF, LLMs will straight-up invent bullshit about what they supposedly say. The machines cannot do the historian’s job.

Also, ALSO, "digitized" doesn't mean accessible. There's one very specific digitized collection of papers that I'd love to use. It costs thousands of dollars via an institutional subscription ONLY, and my library ain't buying. If I ever see them, it'll be in the flesh (which will be cheaper anyway).

LLM's gonna need a robot to remove all the rusty staples and to disassemble those files bound with those 2 metal prongs!

People who have not done scientific research rarely have even the slightest clue how weird doing scientific research actually is. 😂

you just have to write something like "good manual dexterity and attention to detail" and hope that any interviewers have a good sense of humor if they ask for an example of your responsibilities lmao.

I’m interested in how many bones animals have but others are interested in other things like how closely related organisms are, or what chemicals they’re made of. Most of these involve collections (archives) and can only be assisted by ai at best.