myrmepropagandist
I have found one use case. Although, I wonder if it's cost effective. Give an LLM a bunch of scientific papers and ask for a summary. It makes a kind of nice summary to help you decide what order to read the papers in.
It's also OK at low stakes language translation.
I also tried to ask it for a vocabulary list for the papers. Some of it was good but it had a lot of serious but subtile and hard to catch errors.
It's kind of like a gaussian blur for text.
myrmepropagandist
@nazokiyoubinbou @CptSuperlative @emilymbender
Consider the whole genre of "We asked an AI what love was... and this is what it said!"
It's a bit like a magic 8 ball, but I think people are more realistic about the limitations of the 8 ball.
And maybe it's that gloss of perceived machine "objectivity" that makes me kind of angry at those making this error.
myrmepropagandist
@nazokiyoubinbou @CptSuperlative @emilymbender
If I don't have the experience of "finding it useful" I can't possibly communicate clearly what's *wrong* with asking a LLM "can you simulate what it would be like if you didn't have X in your data set" and just going with the response like it could possibly be what you thought you asked for.
It's not going away.
And right now a lot of people give it more trust and respect than they do other people *because* its a machine.
Nazo
@futurebird @CptSuperlative @emilymbender Summaries aren't reliable either.
There are indeed use-cases. But every single one of them comes with caveats. And, I mean, to be fair, most "quick" methods of doing anything come with caveats. It's just that people forget those caveats.
myrmepropagandist
@david_chisnall @CptSuperlative @emilymbender
But, if I ask it, say, to organize papers into groups based on which hypothesis about eusocial evolution they support. LOL. It tries to give me what a want, on the surface the result had me excited for a moment because it was worded like just what I wanted.
But, some of the papers in the set didn't talk about eusocial evolution, yet they'd been placed in a camp. Some papers were in the wrong camp, and worst? It made up a paper not in the set.
David Chisnall (*Now with 50% more sarcasm!*)
@futurebird @CptSuperlative @emilymbender Summarisation is the one I’d be most nervous about because creating a summary is hard: it requires understanding the content and knowing which parts are relevant in a given context, which is why LLMs tend to be awful at it. They don’t understand the content, which is how you get news summaries that get the subject and object the wrong way around in a murder. They don’t know what is important, which is how you get email summaries that contain a scam message and strip all of the markers that would make it obvious that the message is a scam.
If you’re going to read all of them and are just picking an order, that’s probably fine. The worst that a bad summary can do is make you read them in the wrong order and that’s not really a problem.
myrmepropagandist
@david_chisnall @CptSuperlative @emilymbender
It can summarize scientific papers well in part because they have a clear style and even come with an abstract.
The words and phrases in the abstract of a paper reliably predict the content and main ideas of the paper.
Moreover, even if you remove the abstracts, it has lots of training data of papers with abstracts.
Rich Felker
@futurebird @david_chisnall @CptSuperlative @emilymbender Not a criticism but an observation: despite knowing in an a priori way that it couldn't work, you tried that with some kind of expectation (I think?) that it might work.
This makes me realize that a large part of touching LLMs safely isn't just having your own sound mental model for how they work, but also for how human minds/your own mind work (and might be fooled by them).
Or, one can take the safe path and just never touch LLMs to begin with.
GNU social JP is a social network, courtesy of GNU social JP管理人. It runs on GNU social, version 2.0.2-dev, available under the GNU Affero General Public License.
All GNU social JP content and data are available under the Creative Commons Attribution 3.0 license.