AIs gunning for our precious freelancers

[Travis B.]

and was not the mainstream opinion on art before the mid-20th century. Before that, artists were expected to realistically portray the world, and before that, to glorify God or the ruling classes. Art was not really a niche for social misfits as it is today; artists were expected to be disciplined and obedient just like any profession.

I think there was already a bit of an art-as-a-niche-for-social-misfits thing going on in the time from the mid-19th to the mid-20th century, though less than today.

It might depend on the specific art, too. I think poetry was a niche-for-social-misfits thing long before the visual arts, for instance.

Agreed -- take painting, for instance, and the kind of painters that showed up in the mid-19th century after the advent of the photograph obsoleted trying to make perfectly faithful paintings of actual (typically very wealthy) people.

[Raphael]

Agreed -- take painting, for instance, and the kind of painters that showed up in the mid-19th century after the advent of the photograph obsoleted trying to make perfectly faithful paintings of actual (typically very wealthy) people.

More generally, while one particular form of abstract art - abstract paintings on canvass - might have been an early 20th century thing, other forms of abstract art were much older. I mean, there have been abstract ornaments on pottery and in textiles for thousands of years.

[Travis B.]

Agreed -- take painting, for instance, and the kind of painters that showed up in the mid-19th century after the advent of the photograph obsoleted trying to make perfectly faithful paintings of actual (typically very wealthy) people.

More generally, while one particular form of abstract art - abstract paintings on canvass - might have been an early 20th century thing, other forms of abstract art were much older. I mean, there have been abstract ornaments on pottery and in textiles for thousands of years.

I'm thinking of people like Vincent van Gogh and Edvard Munch (to be exact they did most of their work in the later 19th century).

[Raphael]

I'm thinking of people like Vincent van Gogh and Edvard Munch (to be exact they did most of their work in the later 19th century).

Well, modern art in general arguably started with Goya's Black Paintings.

[Travis B.]

I'm thinking of people like Vincent van Gogh and Edvard Munch (to be exact they did most of their work in the later 19th century).

Well, modern art in general arguably started with Goya's Black Paintings.

I agree that Goya was probably the start of modern art myself.

[malloc]

So any ideas on what features would make a conlang unusually difficult for an LLM to learn?

[keenir]

So any ideas on what features would make a conlang unusually difficult for an LLM to learn?

irregularity'd be my hunch

[zompist]

So any ideas on what features would make a conlang unusually difficult for an LLM to learn?

irregularity'd be my hunch

Not at all— LLMs do fine on (say) French. You have to avoid anthropomorphizing: LLMs don't "learn" a language at all, not like a human does. They produce language-shaped text, based on their absorbing reams and reams of similar stuff.

They don't do well with things not amply represented in their training set. That means the only real defense a language has, so to speak, is obscurity. If it doesn't have a lot of texts, the LLM gets lost.

As a test I asked ChatGPT to translate a short text into Esperanto. It did fine— I can read Esperanto well enough to find grammatical errors, and I found none. Admittedly Esperanto is highly regular, but ChatGPT had no problem with the accusative and with local morphological agreement.

I also asked it to write a sentence in Verdurian, and it failed spectacularly— it wasn't even close. Obviously there are enough Esperanto texts in the corpus for its methods to work, not enough Verdurian.

[Raphael]

So the worst case scenario, from an LLM's point of view, would be if a conlang consists only of a written grammar, with no sample texts except for brief sample snippets in the grammar? It could get a "feel" (for lack of a better word) for how to translate into a language from processing a lot of texts in the language, but could not learn and apply formal codified written rules for a language?

[Travis B.]

So the worst case scenario, from an LLM's point of view, would be if a conlang consists only of a written grammar, with no sample texts except for brief sample snippets in the grammar? It could get a "feel" (for lack of a better word) for how to translate into a language from processing a lot of texts in the language, but could not learn and apply formal codified written rules for a language?

Exactly. Because LLM's only parrot and do not understand.

[zompist]

So the worst case scenario, from an LLM's point of view, would be if a conlang consists only of a written grammar, with no sample texts except for brief sample snippets in the grammar? It could get a "feel" (for lack of a better word) for how to translate into a language from processing a lot of texts in the language, but could not learn and apply formal codified written rules for a language?

Hey, there's more Verdurian on the web than that. But yeah, it's mostly volume of input. Rotting bones reports that it doesn't do well with Bengali, which is surely pretty well represented.

Can it follow rules? Well, yes and no. LLMs produce language-shaped output. So they can discourse about grammatical rules (see my upcoming blog post!)-- because grammatical discussions are part of their corpus. That doesn't mean that they can actually use those rules.

Example: I just asked ChatGPT to transform a text according to simple rules, to wit:
1. Replace every noun with a different noun.
2. For verbs, remove all the vowels.
3. Delete any prepositions.

I don’t buy that a life in computronium has any real advantage to compensate for the lack of bodies. Singularity theorists seem to have a priggish distaste for the biological... a long trope in science fiction, one that C.S. Lewis had a great time parodying back in That Hideous Strength, where his villains had a cringing disgust for the messiness, the fluidness, the grossness of organisms. But that taste for dead, totalitarian order is way past its sell-by date.

I don’t by that a llm in mtrcn has any real drwbrdg to cmpnst for the lack of hammocks. Snglrty thrsts seem to have a scnwthr dstst for the spndx... a long swrd in lgtrmn, one that C.S. Lewis had a great time vrdrllng back in That Hideous Strength, where his dndrmns had a gnshn dsgst for the mncls, the blntss, the grbnss of stfflns. But that pssn for dead, totalitarian bdrng is way past its wndshld.

Did it follow the rules? Not really, but it didn't entirely not follow them. It was able to recognize nouns and replace most of them. It applies the "verb" rule rather indiscriminately, including to words it 'thinks' are nouns. It missed most of the prepositions, though in a different paragraph it deleted "the". Yet I'll guarantee you that if you ask "what is a verb" it will give (regurgitate?) an acceptable definition.

I'd say this is unusually bad performance for ChatGPT, probably because by accident I asked it the sort of question it doesn't have reams of answers to.

[rotting bones]

Rotting bones reports that it doesn't do well with Bengali, which is surely pretty well represented.

A lot of Bengali text on the Internet exists as graphics inside pdfs. Most of the text that's out there still uses Latin characters with a Bengali font. Unicode has been adopted recently, but the old media hasn't been updated much. When I text in Bengali, I use Latin characters with sound approximations. My relatives do the same thing. We don't even bother to use fonts. I have only seen journalists use Unicode online.

[zompist]

Rotting bones reports that it doesn't do well with Bengali, which is surely pretty well represented.

A lot of Bengali text on the Internet exists as graphics inside pdfs. Most of the text that's out there still uses Latin characters with a Bengali font. Unicode has been adopted recently, but the old media hasn't been updated much. When I text in Bengali, I use Latin characters with sound approximations. My relatives do the same thing. We don't even bother to use fonts. I have only seen journalists use Unicode online.

How does Google Translate do with Bengali? They have the same basic problem. (I expect better simply because GT will be optimized for translation and ChatGPT isn't.)

[malloc]

So you don't think there are any grammatical or lexical features that would make a language harder than usual for LLMs to learn?

I don’t buy that a life in computronium has any real advantage to compensate for the lack of bodies. Singularity theorists seem to have a priggish distaste for the biological... a long trope in science fiction, one that C.S. Lewis had a great time parodying back in That Hideous Strength, where his villains had a cringing disgust for the messiness, the fluidness, the grossness of organisms. But that taste for dead, totalitarian order is way past its sell-by date.

Incidentally, where is this from? It sounds interesting and perhaps germane to this discussion.

[zompist]

So you don't think there are any grammatical or lexical features that would make a language harder than usual for LLMs to learn?

Nope. But if they have trouble with Verdurian because of a lack of examples, it'll be all the more so with a newly created language.

I don’t buy that a life in computronium has any real advantage to compensate for the lack of bodies. Singularity theorists seem to have a priggish distaste for the biological... a long trope in science fiction, one that C.S. Lewis had a great time parodying back in That Hideous Strength, where his villains had a cringing disgust for the messiness, the fluidness, the grossness of organisms. But that taste for dead, totalitarian order is way past its sell-by date.

Incidentally, where is this from? It sounds interesting and perhaps germane to this discussion.

It's from my meta-analysis of why the Incatena is written the way it is.

[malloc]

Nope. But if they have trouble with Verdurian because of a lack of examples, it'll be all the more so with a newly created language.

So LLMs have no problem learning complicated sandhi rules, Semitic-style consonantal roots, and free word order? That sounds pretty intelligent to me.

It's from my meta-analysis of why the Incatena is written the way it is.

I see then. For what it's worth, I think the real advantage of computronium and why it would presumably dominate in the future is its sheer efficiency. Giving up sex and food would suck but dispensing with all those corporeal pleasures would allow you to focus on growth for the sake of growth, maximal productivity, and so forth. While flesh-based people loaf around eating and shagging, computronium-based entities are busy gobbling up natural resources to build more computronium and weapons to fight anyone who objects. Compare how capitalism, with its emphasis on maximizing efficiency and reinvesting profits (not to mention colonialism and genocide), overcame feudalism and indigenous societies.

[rotting bones]

How does Google Translate do with Bengali? They have the same basic problem. (I expect better simply because GT will be optimized for translation and ChatGPT isn't.)

When it was first launched, the translations were unreadable. Google Translate has been accepting corrections from users for years now. At this point, the results are no longer unreadable, only unreliable. IME The Bengali translations are consistently worse than Tagalog.

When I have free time, I should investigate how it does for the other non-English languages I read, Japanese and Russian.

[rotting bones]

So LLMs have no problem learning complicated sandhi rules, Semitic-style consonantal roots, and free word order? That sounds pretty intelligent to me.

A basic LLM is not doing abstract calculations. It's only trying to predict the most likely next token. It's surprising how well this works, but as you can expect, it's not a reliable method to get consistent answers.

[Torco]

Nope. But if they have trouble with Verdurian because of a lack of examples, it'll be all the more so with a newly created language.

So LLMs have no problem learning complicated sandhi rules, Semitic-style consonantal roots, and free word order? That sounds pretty intelligent to me.

sigh... at this point you just trolling. i'm getting flashbacks about The South.

[keenir]

Nope. But if they have trouble with Verdurian because of a lack of examples, it'll be all the more so with a newly created language.

So LLMs have no problem learning complicated sandhi rules, Semitic-style consonantal roots, and free word order? That sounds pretty intelligent to me.

if those are the directions it is programmed with, then yes it will have an easy time with those. but it will only perform the things it was programmed with - it will never innovate those into agglutinative TAM-rich statements.

you're basically still being surprised that a typewriter will always return to the same point on a page that you've placed the left alignment at.