So, there has been some big news going around lately about LLMs. I won’t go into a lot of detail about what they are, you can find that in various places, in escalating levels of detail. And that’s not even including the very latest news from just this week (sentence written on 2023-03-16).
I thought it might be worthwhile to me to write down some of my thoughts at the moment. Not to you or anything, I mean it will be worthwhile to me so I can sort out what I think. You can even read it, if you want.
I want to be clear here: I did not use any LLM in the writing of this in anyway. These are my thoughts in my own poor wordings, and lacking in good editing style.
LLMs are a tool. Some have said their like having an intern. I think that is doing a disservice to interns, and they should be angry and offended at that. Yes, it’s true that both interns and LLMs can be unreliable. That’s true of everything and everybody to some degree.
We’re all just having our first impression of this right now. Something new has arrived on the scene. Well, new isn’t quite the right word, but newly available and potentially obviously possibly useful? Maybe even a little bit (a lotta bit?) threatening.
Certainly seems to be evolving fast. I think the first widespread leading edge of it was probably the Github Copilot thing, um, last year, I think it was. That was something.
We had our midjourney/stable diffusion fun last year too. They still haven’t worked all the kinks out of that one, but there has been some very fun stuff out of that. It’s kind of in the really effing creepy part of the uncanny valley now, were things look right until you notice just how wrong they are, also this.
Then along came GPT-3, and that has been fun if actually really quite the confident fabulist (aka bullshitter). And you don’t have to try very hard. These tools, which I don’t mean to make out to be sort of equivalent in anyway, but from an outside perspective, they sort of do seem to form a technological cohort.
I guess GPT-4 is blowing the socks off of people right now.
Those threads are
I’m going to mostly focus on the text based LLMs, hence the title. But probably some of these things are related.
The extraction has already begun here, and I would say it is moving rapidly.
What are the costs?
Well, I’m writing text right now. A lot of people write text. Should we do that anymore? I already don’t like it that much when Google Docs suggests a different word choice or some grammar improvement. Sometimes it’s better, and I’ll go for it. Other times it just feels like it’s flattening my own voice out of the document I am writing.
I have a friend that is a professor. And I was talking to them the other day, back in those early days of ChatGPT3. I think it was mid-February 2023 (about a month ago). And they said that ChatGPT3 was a great boon to them, so helpful for different writing tasks, such as composing possible tweets on topics for their daily twittering, or composing letters or recommendation for students that needed them, tweaked and edited by hand also, naturally. I don’t know just how hypothetical these statements were, necessarily. But they worried me. Worried me on several fronts.
Isn’t a slippery slope from taking the output of these models as suggestions and editing or rewriting them yourself, maybe even in your own words, to just copy/pasta’ing that text verbatim because, oh I don’t know, you’re in a hurry that day there is a grant proposal due and the kids going to be late and I think that about covers what I wanted to day. And after that first time, it get’s a little easier the second time once the seal has come off.
Another angle here is that maybe if you need a force multiplier here to get all this text written you should probably just question if all that verbosity is really necessary. If this text generator is generating all that text, why not just send the prompt and leave it to the reader to stick it in the LLM if they want more to read. They’re probably going to ask the LLM to summarize that long letter you sent them anyway.
I hear Google is working on building that stuff right into Gmail and Docs already, but I’m not convinced. If they were so far ahead with AlphaGo or whatever, why didn’t they add this already? I know the arguments about protecting their search business or ad business or whatever. Are they stupid? Maybe all that money made them stupid. It sure seems to have made Meta stupid.
What the hell is going on here? The future is getting weird, maybe I should turn pro.
PS: I didn’t really edit this after I wrote it. Sorry about that.
If you want to comment, do that on Mastodon:
To Reply: Email me your initial thoughts
Posted: in Notes.
Other categories: garo.lol.
Back references: none.
Tags that connect: [[AI]] No Government in the Cloud, More thoughts on LLMs.
Tags only on this post: none.