pull down to refresh
I suppose you might say that even their outrageous statements are mid and I would probably agreeRight! I'd suggest that to be caused by that everything still goes through some form of dampening by median() over (everything_ever_written) - or at least within the "expert" in a MoE model. That takes the edge off.
It took me more time than I would like to admit to write this response. Many minutes were spent with my fingers hovering half way through a sentence because I had to think about what I was writing.
That's my world, also because I guess I suck as a writer. Every post I write goes like that, including this one.
I keep using the phrase "labor theory of value for writing" even though I know it's wrong. But there is something to this idea that you can't get to a thoughtful conclusion without actually spending time thinking -- and slop is too fast.
I fully agree with that, which is why I have been emphasizing LLM use for half-fabricates or intermediate work product, never end product, ever since the output became better (about a year and a bit ago.) This takes away the "consciousness" dream/nightmare (or what I believe to be a fable, pun intended) and replaces it with the assigned role of tool. After all, if I wanted to read bot output, I don't need anyone to middleman it, I'd just ask the bot directly - even if you have more skills in prompting then me: the interesting part then becomes how you build your prompts, not the resulting output from the bot.
This goes for articles or books, but also for code, for images, for business process documentation... for everything basically. The tool is there, it's largely a commodity and anyone can use it. So you don't need to publish it and expect to get something for nothing.
Personal example: Since this morning I did 30+ analyses on software. 2 android apps, their dependencies and also a bunch of libraries SN (or mostly 3p components of SN) depends on, and each of them got sent through an LLM for pre-analysis. I read (almost) all the slop coming out. It's very detailed because I let the LLMs do what I would do: go over every changed line of code, trace it end-to-end and form an opinion on correctness and risks, check for oddities like opaque blobs or postinstall scripts, network and filesystem calls, and so on. I then read what the LLM says and flag up what needs to be further analyzed, which I do manually. I then document my findings next to that of the LLM (also when it was wrong - happens sometimes, but less often now that I've spent almost a year tuning this process and while newer model versions got finetuned a lot more) and I can make decisions.
No one ever sees any of this slop. I don't need to read it a second time either, unless maybe I made a mistake somewhere and I need to go over my steps. I guess my usage of LLMs in this is 100% as a research tool; not as an assistant, "agent", replacement or extension of myself, or fake simulated entity that doesn't exist, because it is a computer program, not real.
I personally like it that way. Also I hate reading slop and I do it non-stop. Which is why, when I'm away from the bots for a while, I really hate slop.
I agree with this. But it gets messy because I am still offput by a human with a surprising or possibly interesting idea using a wad of slop to present it. I don't think that I need to marvel at every turn of phrase (although there are some writers who almost achieve this), yet most slop leaves me feeling frustrated.
There's no doubt that
midis boring, but AI is still somewhat boring if prompted into outrageous statements. (I suppose you might say that even their outrageous statements aremidand I would probably agree). I am willing to go out on the limb to say that should the geniuses develop AI that surprises us, I still think it may be slop.It took me more time than I would like to admit to write this response. Many minutes were spent with my fingers hovering half way through a sentence because I had to think about what I was writing. It could almost be described as grinding it out -- writing is work. There's something about slop that feels like fake work. Like a badly put-together building where human writing can be a well-crafted structure that stands for a thousand years.
I keep using the phrase "labor theory of value for writing" even though I know it's wrong. But there is something to this idea that you can't get to a thoughtful conclusion without actually spending time thinking -- and slop is too fast.