Just to be sure I'm not being unfair, I searched for writeups of LLM uses. Here's a representative example of the genre: https://www.techopedia.com/12-practical-large-language-model-llm-applications

They mention 12 use cases. Some are done better with simpler, cheaper models or non-ML techniques (1, 4, 6, 8, 12). Some are wildly speculative (2, 9, 10).

But that leaves 4 items that I want look at carefully: content creation, customer support, sales automation, and writing code. Those are at least superficially plausible places where LLMs and large image models could have practical uses.

2/

First, though an important philosophical point: LLMs are fancy autocomplete. You give them a set of words, they'll predict the next word based on the enormous corpuses they've been trained on. This can give them the appearance of sentience. People will talk as if they "understand" things. They don't. It's the million-monkeys-with-typewriters thing, but the monkeys have seen enough English text that the next word is statistically consistent with the previous words.

Humans are subject to to pareidolia, and we really like to anthropomorphize things. It's not just thunder, it's a guy with a name and a look and a personality and a whole family. It's not just a bit of winter where we celebrate with our dearest; it's a fat guy in a red suit with specific facial hair. So although the text can feel human, we'll have to work hard to think of ChatGPT as a bit of unfeeling machinery, not our plastic pal who's fun to be with.

3/

Ok, so first, content creation. That seems positive, right? Wrong! The best way I've seen of explaining this: "Why should I take the time to read something nobody took the time to write?"

I think this one is a huge net societal negative. The people out there who want instant "content" are almost entirely not readers. They're people who want something to run ads against. They're people who want the credit for writing without doing the work. They're people who want to sell you something without understanding the something or whether or not it might be good for you. In short, they're people with various levels of contempt for their readers.

4/

@spocko @williampietri @Sevoris

🤘🏻🤖 aha glad you liked that one!

@clive @spocko @williampietri it has stuck around with me! TBH I think it has gained in relevance as further evaluation of language models has revealed how their statistic view of language pulls ideas and writing towards a washy common denominator and then the NYT lawsuit came along and demonstrated just how deep the overfitting goes...

Language models to me now look like "content machines" with all the absolute issues thereunto...

@Sevoris @spocko @williampietri

It really seems to be the apotheosis of the content fetish, doesn't it?