AI assistants as sound change appliers

[Jonlang]

So I've come back to my conlangs after a while off. I've discovered Grok - xAI's AI assistant (specifically Grok 3 - 1 and 2 completely passed me by). Out of curiosity I uploaded a document of sound changes, written in plain language, not the notation used by SCA and the like. And, with only a couple of tweaks needed for Grok to understand the premise, it now handles them perfectly.

Has anyone else had any experience with AI applying sound changes?

[Qwynegold]

I have never tried it, but it's a really good idea. I'll have to try and see at some point if ChatGPT can handle it.

[Raphael]

Isn't that the nazi AI?

[Jonlang]

Isn't that the nazi AI?

No more than the Beetle is a Nazi car or Hugo Boss is a Nazi designer. Or maybe we should stop drinking Fanta because it's a Nazi drink? Or not buy Ford cars because Henry Ford was also a Nazi? IBM provided punch-card "computers" to the Nazis, so we can't use IBM anymore! Same goes for Volkswagen, BMW, Porsche, Mercedes-Benz, and Siemens - all of whom supplied the Nazis.

[malloc]

Isn't that the nazi AI?

Quite. Artificial intelligence is bad enough but adding Elon Musk is even worse. This is evil raised to evil, maybe even tetrated. [This piece of shit phone changed "tetrated" to "treated" without me realizing. Fuck AI]

[Raphael]

Isn't that the nazi AI?

No more than the Beetle is a Nazi car or Hugo Boss is a Nazi designer. Or maybe we should stop drinking Fanta because it's a Nazi drink? Or not buy Ford cars because Henry Ford was also a Nazi? IBM provided punch-card "computers" to the Nazis, so we can't use IBM anymore! Same goes for Volkswagen, BMW, Porsche, Mercedes-Benz, and Siemens - all of whom supplied the Nazis.

All your examples are about when my grandparents were young. In none of your examples does anyone who worked for the company back then still work there. Elon Musk is giving orders to his various companies now, and harming the world now.

And, AI output is more of an "intellectual" thing than car design or even fashion design, so it's more influenced by the ideology of the people behind it than a car is.

[Jonlang]

Isn't that the nazi AI?

No more than the Beetle is a Nazi car or Hugo Boss is a Nazi designer. Or maybe we should stop drinking Fanta because it's a Nazi drink? Or not buy Ford cars because Henry Ford was also a Nazi? IBM provided punch-card "computers" to the Nazis, so we can't use IBM anymore! Same goes for Volkswagen, BMW, Porsche, Mercedes-Benz, and Siemens - all of whom supplied the Nazis.

All your examples are about when my grandparents were young. In none of your examples does anyone who worked for the company back then still work there. Elon Musk is giving orders to his various companies now, and harming the world now.

And, AI output is more of an "intellectual" thing than car design or even fashion design, so it's more influenced by the ideology of the people behind it than a car is.

Okay, you hate Musk. But this isn't the US Politics thread so I'm not entering into this debate. I'm also trying out ChatGPT for the same thing as well, because y'now - choices. No-one is telling you to use Grok, or to even try an AI chatbot for sound changes.

[WeepingElf]

Isn't that the nazi AI?

Please don't use the word "Nazi" this way. If something is objectionable, say "objectionable" or something like that. Don't say "Nazi" unless it is objectionable in the specific way Nazism is objectionable. It does injustice to the victims of the unparalleled atrocities of the actual Nazis.

[Raphael]

Isn't that the nazi AI?

Please don't use the word "Nazi" this way. If something is objectionable, say "objectionable" or something like that. Don't say "Nazi" unless it is objectionable in the specific way Nazism is objectionable. It does injustice to the victims of the unparalleled atrocities of the actual Nazis.

Given that actual "Hitler was great!" types all over the Internet are cheering on Grok, and are cheering on the man driving its creation, and given that it was explicitly designed with the main distinguishing feature of removing the safeguards that are meant to keep other AI projects from (re)producing bigotry, I think it is, well, not objectionable in the specific way in which World War 2 era Nazi war criminals are objectionable, but still objectionable in the specific way in which 21st century sieg-heiling skinheads are objectionable.

[linguistcat]

I'd suggest trying DeepSeek, since it uses less resources to run but I've found it has a hard time with things like sound changes and the like at least when I tried it. But you could still look into it. I mean unless you hate the Chinese but that would be a political issue.

[Raphael]

I'd suggest trying DeepSeek, since it uses less resources to run but I've found it has a hard time with things like sound changes and the like at least when I tried it. But you could still look into it. I mean unless you hate the Chinese but that would be a political issue.

Has anyone forked DeepSeek yet? I mean, it's open source, so that should be principally possible.

[rotting bones]

Personally, I would trust the output more if you told the LLM to write a script that applies the sound changes. This would not only improve regularity, it would also let you examine the changes yourself.

It might be easier to just use a SCA tool.

[rotting bones]

Has anyone forked DeepSeek yet? I mean, it's open source, so that should be principally possible.

The training data isn't publicly available IIRC. What they released is the trained model.

[Raphael]

Has anyone forked DeepSeek yet? I mean, it's open source, so that should be principally possible.

The training data isn't publicly available IIRC. What they released is the trained model.

Ah, thank you.

[foxcatdog]

So I've come back to my conlangs after a while off. I've discovered Grok - xAI's AI assistant (specifically Grok 3 - 1 and 2 completely passed me by). Out of curiosity I uploaded a document of sound changes, written in plain language, not the notation used by SCA and the like. And, with only a couple of tweaks needed for Grok to understand the premise, it now handles them perfectly.

Has anyone else had any experience with AI applying sound changes?

Chatgpt is pretty bad at sound changes. It can never remember them or make up consistent derivations.

[bradrn]

Personally, I would trust the output more if you told the LLM to write a script that applies the sound changes. This would not only improve regularity, it would also let you examine the changes yourself.

It might be easier to just use a SCA tool.

Agreed. (I would suggest my own!)

[Jonlang]

So I've come back to my conlangs after a while off. I've discovered Grok - xAI's AI assistant (specifically Grok 3 - 1 and 2 completely passed me by). Out of curiosity I uploaded a document of sound changes, written in plain language, not the notation used by SCA and the like. And, with only a couple of tweaks needed for Grok to understand the premise, it now handles them perfectly.

Has anyone else had any experience with AI applying sound changes?

Chatgpt is pretty bad at sound changes. It can never remember them or make up consistent derivations.

I have realised that Grok gets less good at it with the more changes that you give it. Breaking them up into natural stages (to make it easier for me) helps, and using SCA notation seems to help too, but I'm not particularly good at remembering how to format it. It also seems to help if you use one character per phoneme rather than, say, using th for [θ]. So it seems to favour simplicity but can handle plain English in smaller doses. I've noticed it helps if you give it specific instructions which is easier when you have a rough idea of what the final word will be. So for my Proto-word marasō I can instruct Grok to ignore the rules about i-affection and a-affection, knowing they will not apply (no word-final i or a). Grok correctly gives brās for this (and marso in the sister conlang), though these aren't particularly radical changes. It's not a terrible start; who knows how it may improve in the next five years or so.

[Ares Land]

Personally, I would trust the output more if you told the LLM to write a script that applies the sound changes. This would not only improve regularity, it would also let you examine the changes yourself.

It might be easier to just use a SCA tool.

My thoughts exactly; I'm not sure a LLM is the right tool for the job.
The idea of automating sound change is to have reliable and consistent results, that's not exactly something LLMs are good at.

As for political issues... I'd suggest Mistral https://mistral.ai/ which is reasonably free of, um, ethically dubious connections and is I think as efficient at DeepSeek.
I tried a few historical linguistics questions and did get reasonable outcomes. (Though, generally, I really don't see any real use for LLMs, but that's another question.)

[Travis B.]

Personally, I would trust the output more if you told the LLM to write a script that applies the sound changes. This would not only improve regularity, it would also let you examine the changes yourself.

It might be easier to just use a SCA tool.

Agreed. (I would suggest my own!)

You guys, just follow bradrn's advice and use an SCA already rather than trying to shoehorn LLM's into a role they were not designed for.

[Tsimaah]

Personally, I would trust the output more if you told the LLM to write a script that applies the sound changes. This would not only improve regularity, it would also let you examine the changes yourself.

It might be easier to just use a SCA tool.

Agreed. (I would suggest my own!)

You guys, just follow bradrn's advice and use an SCA already rather than trying to shoehorn LLM's into a role they were not designed for.

I concur, LLM's are bad at arithmetic and anything requiring precision and fastidiousness (such as regular sound change application). Some of my coworkers have even discovered that they are surprisingly unreliable at writing scripts to calculate arithmetic for them,
the LLM always manages to forget some small but important detail or another, which is not discovered until much later.