Reading through this entire thread, I suspect that somehow generative AI actually became a political issue. Polarized politics is like a vortex sucking all kinds of unrelated things in.
In case that doesn't get my comment completely buried, I will go ahead and say honestly that even though "AI slop" and paywalled content is a problem, I don't think that generative AI in itself is a negative at all. And I also think that part of this person's reaction is that LLMs have made previous NLP techniques, such a those based on simple usage counts etc., largely irrelevant.
What was/is wordfreq used for, and can those tasks not actually be done more effectively with a cutting edge language model of some sort these days? Maybe even a really small one for some things.
> As soon as generative AI has the capability to take someone's job, that is political.
What is political is people enshrining themselves in chokepoints and demanding a toll for passing through or getting anything done. That is what you do when you make a certain job politically 'untakable'.
People who espouse that the 'personal is political' risk making the definition of politics so broad that it is useless.
Everything involving any kind of coordination, cooperation, competition, and/ot communication between two or more people involves politics by its very nature. LLMs are communication tools. You can't divorce politics from their use when one person is generating text for another person to read.
The simplest example that comes to mind of something frequency analysis might be useful for would be if you had simple ciphertext where you knew that the characters probably 1:1 mapped, but you didn't know anything about how.
It could also be useful for guessing whether someone might have been trying to do some kind of steganographic or additional encoding in their work, by telling you how abnormal compared to how many people write it is that someone happened to choose a very unusual construction in their work, or whether it's unlikely that two people chose the same unusual construction by coincidence or plagiarism.
You might also find statistical models interesting for things like noticing patterns in people for whom English or others are not their first language, and when they choose different constructions more often than speakers for whom it was their first language.
I'm not saying you can't use an LLM to do some or all of these, but they also have something of a scalar attached to them of how unusual the conclusion is - e.g. "I have never seen this construction of words in 50 million lines of text" versus "Yes, that's natural.", which can be useful for trying to inform how close to the noise floor the answer is, even ignoring the prospect of hallucinations.
Yes, it's become extremely politicized and its very tiresome. Tech in general, to be frank. Pray that your field of interest never gets covered in the NYT.
In case that doesn't get my comment completely buried, I will go ahead and say honestly that even though "AI slop" and paywalled content is a problem, I don't think that generative AI in itself is a negative at all. And I also think that part of this person's reaction is that LLMs have made previous NLP techniques, such a those based on simple usage counts etc., largely irrelevant.
What was/is wordfreq used for, and can those tasks not actually be done more effectively with a cutting edge language model of some sort these days? Maybe even a really small one for some things.