@plexus

Extruded language food product

@plexus

I agree with this, but it makes me appreciate LLMs.

@plexus Reminds me of reading LTI, Lingua Tertii Imperii, by Victor Klemperer, on how the NSDAP changed German language and devalued it for their goals (including constant use of invented buzzwords)

https://en.wikipedia.org/wiki/LTI_%E2%80%93_Lingua_Tertii_Imperii

Not sure how well the book works in English translation

@plexus
I hate the AI goo. Not because it's AI, but because it's goo.

I also know that this is what calligraphers said about the typewriter.
And it's true, a beautiful calligraphy looks much better than even the most meticulous type set in Helvetica. It's also much less accessible.

Here's hoping the new tools help us reach further in the end, and that the layers of art stick around for additional enjoyment!

@plexus AI beat chess in the '90s. No one cared, chess spectators are all still most interested in the human players, even 30 years later. These people weren't interested in well played chess. They were interested in the human choices of the players.

Similarly, people don't value language. They value human expression.

@plexus Thoughts to counter your points:

1. Plenty of posts composed by real people leave the same sad taste of unoriginality. I find it sad to come to Reddit every day and read the same useless threads that repeat in style and meaning. Limited repeated vocabulary of online users hurts sometimes.

2. Try local LLMs which vary in style. They are capable of much more varied content than one might imagine initially. With correct prompt, it's able to replicate your post, for example, word-for-word.

@plexus I remember reading a good post (I think by @mekkaokereke ) about how language is used by groups to protect against people trying to infiltrate them without a shared background.

It wouldn't surprise me if the same thing happened to defend against LLMs - you won't accept, say, a job application if there aren't at least a couple of uses of 0-day slang or memes. Language evolves fast, it's one of the amazing things about it.

@plexus Language is a tool. It’s a versatile tool, but a tool nonetheless. Not every word needs to be skillfully crafted and sweated out for weeks before it’s committed to the page.

I’m in no position to tell you how to feel about LLMs. Your position is certainly valid and I share it _in some contexts_. LLM is unlikely to write your favorite novel. Not this year, anyway. But it’s probably OK for all the bulshit elsewhere. It can write a decent marketing copy. Results are not stellar but you probably wouldn’t be able to tell the difference from something written by a person. And like it or not there’s a lot of it out there and demand doesn’t seem to shrink any time soon.

Since most of us are geeks around here let me suggest an analogy. Back in the day we used to craft software in machine code. That was too hard so we came up with assembly which was pretty close to machine code but easier to work with. And what we could do with assembly! Read source code of PacMan, or Zork, or Impulse Tracker. Those are works of art.

But then heathens came up with compilers. How those messed up our machine code! All those automatic optimizations… unrolled loops, and constant elimination, and reordering, and speculative execution, and undefined behavior exploits… The blasphemy! The garbage those compilers spit out is unreadable. Will a compiler ever come up with a fast inverse square root on its own?!

Well, it might since it’s us who’s in charge of teaching that optimization to the compilers.

Machine code/language is a tool. It’s primarily used to achieve some goal that is mostly unrelated to the tool itself. It is OK to automate production if the output quality doesn’t suffer for the purpose. Like compilers, LLMs are fine for some use cases.

We still have people who write ingenious assembly and beautiful poetry and we always will.

Let me give you another analogy. In games there’s a notion of skill floor. Effectively, it’s the basis where a player can play the game. We have tutorials in games to ensure everyone has the basic skills in order to not get stuck. There’s also skill ceiling. It’s how many unessential skills you can get before becoming the best player you can be. Games want this interval to be as wide as possible. This would mean that many people can start playing but also they won’t get bored fast as there’s new skills to learn and we tend to be excited about new things.

LLMs are effectively widening this gap with language. Ceiling remains high. But floor gets lower. People who couldn’t write marketing copy now can guide an LLM to produce something acceptable. People who are just starting to learn English and are unsure of their skill now can write an OK blog post on the topic they care about. More people can use language effectively with LLMs. At the same time the value of good prose/poetry remains the same.

I see why you might think the language is devalued, though. Providers of LLMs promise limitless ability to do anything with their products. In that light results are certainly disappointing. But if we look at LLMs and see them for what they actually are we might get some use of them.

@plexus
We have an unasked for 'feature' in Atlassian where every time you highlight something, even a single word, you get a popup across the screen asking if you would like this reformatted to sound 'business like' or 'casual' or summat...
Goes well with the two dozen AI macros you have to scroll past asking to do translations into pretty much any language...

At our last account meeting I asked them to turn the effing feature off and don't come back till they offered Pirate Speech...

@plexus
The term for awful AI art gaining traction is "Slop"

https://elblog.pl/2024/06/12/understanding-slop-the-emergence-of-low-quality-ai-content-online/

We need a similar term for the drivel generated as text. Spew?

Why should I bother to read something that nobody could be bothered to write?

@plexus

Just struck me -- you need AI to FAIL.

@plexus Well said!

In a sense I think LLMs bring this full circle with machines now being able to autogenerate content to feed other machines (and, oddly, LLMs) - This all makes both Search engines and also the LLMs themselves so useless that I wouldn't be surprised if a few of the organisation methods we invented at the beginning of the web (Webrings, Human curated Directories) will make a comeback just to find meaningful, human content"

@plexus I worry a bit that LLM can lead to a stagnation of language.

Blog posts, press articles,... etc. from 20 years ago sound a bit different than todays texts.
Press articles from 40 years ago sounds imho already a bit outdated.
Going back 60 years it becomes already quite noticable old fashioned with differences in the lexicon.

But if from now on a significant part of the press corpus becomes LLM generated, feeding only on itself, how can a language change over time anymore?

@plexus openAI and its ilk are the enemy

and not even a GPL-like open source model will help, because they'll just slurp those weights into their training process if it's good enough - we're way past caring about any kind of licensing

it's giant laundering machines, washing off the creative origin of anything it has touched

Edit: the only good thing about the GPTs is that they're fundamentally a dead-end

@plexus I wonder, whether ancient people who valued storytelling, memmorised poems with thousands of verses and worshiped Ogmios or other gods of eloquence, would be as easily fooled by the glitter and effortlessness as we are.

I honestly don't know.