Zompist Bboard Again

Travis B. wrote: ↑Mon Jun 23, 2025 5:07 pm Exactly. Because LLM's only parrot and do not understand.

Not so much "parrot" as "recycle parrot droppings", some might opine.

Out of curiosity, zompist, do you have any second thoughts on your old robots story, in the light of this decade's developments?

malloc wrote: ↑Mon Jun 23, 2025 9:07 pm

zompist wrote: ↑Mon Jun 23, 2025 7:49 pmNope. 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 you think Sanskrit or Hebrew are "difficult" while English is easy, you were asleep in your linguistics class. There's no good argument for some languages being more difficult than others. Every language can be learned by a human child in its first six years or so.

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.

If anyone threatens to do that in the Incatena, they just turn off their machines.

Raphael wrote: ↑Tue Jun 24, 2025 2:56 pm Out of curiosity, zompist, do you have any second thoughts on your old robots story, in the light of this decade's developments?

No, I think it holds up quite well. The basic issue (how to handle 0.9 sentience) is still relevant— and in the future. Note that I had FI running on neural networks, which is pretty accurate (LLMs are artificial neural networks; the specialized hardware is still futuristic).

Of course I'd update the jargon and references. And probably make the opponent groups more sophisticated. And throw in some satire about the OpenAI era.

zompist wrote: ↑Tue Jun 24, 2025 2:59 pm

If you think Sanskrit or Hebrew are "difficult" while English is easy, you were asleep in your linguistics class. There's no good argument for some languages being more difficult than others.

"You know, in most Western languages, like English, the given name comes first, and the family name second. Except when there's a list of names sorted alphabetically. Then the family name comes first, followed by a comma, followed by the given names. And..."

"Normally, numbers in English are written using a system adapted from Arabic numerals. But there are some specific contexts, like numbers as part of monarchs' names, and numbers on some types of clocks, where people use a system called 'Roman numerals' instead, which..."

"The different past tenses in English are..."

"There are also two different present tenses..."

Anyway, seeing someone apparently believing that English is easy is another nice opportunity to sing the praises of Arika Okrent's Highly Irregular: https://arikaokrent.com/

zompist wrote: ↑Tue Jun 24, 2025 3:05 pm

Raphael wrote: ↑Tue Jun 24, 2025 2:56 pm Out of curiosity, zompist, do you have any second thoughts on your old robots story, in the light of this decade's developments?

No, I think it holds up quite well. The basic issue (how to handle 0.9 sentience) is still relevant— and in the future. Note that I had FI running on neural networks, which is pretty accurate (LLMs are artificial neural networks; the specialized hardware is still futuristic).

Of course I'd update the jargon and references. And probably make the opponent groups more sophisticated. And throw in some satire about the OpenAI era.

Well, the general political background seems to be the reverse of what happens in real life.

Raphael wrote: ↑Tue Jun 24, 2025 3:16 pm

zompist wrote: ↑Tue Jun 24, 2025 3:05 pm Of course I'd update the jargon and references. And probably make the opponent groups more sophisticated. And throw in some satire about the OpenAI era.

Well, the general political background seems to be the reverse of what happens in real life.

Not sure I'm following... I think the reference to politics is the differing opinions on AI: people worried their jobs being stolen, and people worried about AIs being oppressed. We have the first now. The second, not yet, but that's because the story is not about 0.5 sentience or whatever ChatGPT is, but about 0.9 or above. (Obviously we have no actual scale like that now either; I hope it's reasonably clear what I mean. No one thinks ChatGPT is fully sentient except malloc.)

But to clarify, I didn't foresee that a stage of 0.5-level AIs would be a Big Thing. And I'm not entirely convinced it is a Big Thing rather than a Big Bubble. But bubble or not, these things would affect how people feel about near-sentient AI in the future.

zompist wrote: ↑Tue Jun 24, 2025 4:07 pm

Not sure I'm following... I think the reference to politics is the differing opinions on AI: people worried their jobs being stolen, and people worried about AIs being oppressed. We have the first now. The second, not yet, but that's because the story is not about 0.5 sentience or whatever ChatGPT is, but about 0.9 or above.

I mean that J4P* seems to be a more or less decidedly right-wing group, while HAL* is clearly coded as left-wing. In real-life politics, right now, it seems to be right-wing corporate techbros who push AI and talk up its abilities, and left-wing or left-leaning types who worry about AIs destroying people's jobs.

*For readers who haven't read zompist's robot story: in the story, J4P is the group "Jobs for People", while HAL is the "Human-AI-League.

Raphael wrote: ↑Tue Jun 24, 2025 4:14 pm I mean that J4P* seems to be a more or less decidedly right-wing group, while HAL* is clearly coded as left-wing. In real-life politics, right now, it seems to be right-wing corporate techbros who push AI and talk up its abilities, and left-wing or left-leaning types who worry about AIs destroying people's jobs.

Fair enough. As I said, I'd update those parts, which fit the 1990s much more than today. I'd note though that worries about job losses are also right-wing; they just currently blame it on immigration and outsourcing. (Never mind that right-wingers in power instituted those policies.)

People worry about jobs, but I don't think we know how that will shake out. It hasn't yet-- unemployment here is 4%, which is about as low as it gets in this half-century, and is also the Fed's informal target. Plus, actual hiring depends on the relative cost of humans and chatbots-- and the price of the latter is being held artificially low.

The techbros' idea is to lock in customers now at the artificial rate, then jack up prices later. That depends on a lot of chancy assumptions, such as 1) that customers will be locked in and won't just reverse their adoption of AI; 2) that cheaper Chinese or open-source AIs won't be available; 3) that business customers will never notice that chatbots can't do what's promised; 4) that the actual demand for chatbots is extraordinarily high.

zompist wrote: ↑Tue Jun 24, 2025 2:59 pmIf anyone threatens to do that in the Incatena, they just turn off their machines.

Except that the computronium faction can simply fight back and given their incredible efficiency and productivity, they would have quite the advantage. Look at the real world where fossil fuel corporations are driving the world to ecological catastrophe as we speak and consider why nobody has thought of walking into oil fields and just turning off the pumps.

Incidentally this is why I put so much emphasis on resisting AI, not because I consider it already sentient or human-level intelligent but because preventing it from getting out of hand is far easier than fighting it when it has already gained the upper hand.

malloc wrote: ↑Tue Jun 24, 2025 7:15 pm

zompist wrote: ↑Tue Jun 24, 2025 2:59 pmIf anyone threatens to do that in the Incatena, they just turn off their machines.

Except that the computronium faction can simply fight back and given their incredible efficiency and productivity, they would have quite the advantage.

except specialization has its costs: look at cheetahs, or IRL factories that produce only one or two things -- switching over from making the wires for computer chips, to making warbots, is far from being as easy as you think it is.

Look at the real world where fossil fuel corporations are driving the world to ecological catastrophe as we speak and consider why nobody has thought of walking into oil fields and just turning off the pumps.

1. its not a single dial that you'd have to turn.

2. private property.

ncidentally this is why I put so much emphasis on resisting AI, not because I consider it already sentient or human-level intelligent but because preventing it from getting out of hand is far easier than fighting it when it has already gained the upper hand.

...which will never happen.

btw, is that why you don't want to fight Trump? because you think he's already won?

malloc wrote: ↑Tue Jun 24, 2025 7:15 pm

zompist wrote: ↑Tue Jun 24, 2025 2:59 pmIf anyone threatens to do that in the Incatena, they just turn off their machines.

Except that

Except that you obviously know nothing about my SF world, so shut up about it. Write your own dystopia where things can fall apart the way you want it to.

zompist wrote: ↑Tue Jun 24, 2025 8:03 pmExcept that you obviously know nothing about my SF world, so shut up about it. Write your own dystopia where things can fall apart the way you want it to.

Just out of curiosity, though, how do they solve the problem I mentioned?

keenir wrote: ↑Tue Jun 24, 2025 7:22 pmexcept specialization has its costs: look at cheetahs, or IRL factories that produce only one or two things -- switching over from making the wires for computer chips, to making warbots, is far from being as easy as you think it is.

It seems reasonable to assume they would invest in some form of defense. No superintelligent AI or uploaded consciousnesses would go through all the trouble of forging all that computronium only to forget to defend it.

...which will never happen.

You simply don't know that. There is nothing that prevents it from happening. The only question is whether researchers will pursue it long enough to reach that point.

malloc wrote: ↑Tue Jun 24, 2025 8:13 pm

zompist wrote: ↑Tue Jun 24, 2025 8:03 pmExcept that you obviously know nothing about my SF world, so shut up about it. Write your own dystopia where things can fall apart the way you want it to.

Just out of curiosity, though, how do they solve the problem I mentioned?

did you read anything set in the Incatena?

keenir wrote: ↑Tue Jun 24, 2025 7:22 pmexcept specialization has its costs: look at cheetahs, or IRL factories that produce only one or two things -- switching over from making the wires for computer chips, to making warbots, is far from being as easy as you think it is.

It seems reasonable to assume they would invest in some form of defense. No superintelligent AI or uploaded consciousnesses would go through all the trouble of forging all that computronium only to forget to defend it.

why not? you assume that the humans wouldn't invest in defense, so why are the uploaded conciousnesses (presumably of humans, right?) and AIs not given the same presumption? or do you enjoy playing tennis with the net down?

...which will never happen.

You simply don't know that.

yes i do.

There is nothing that prevents it from happening.

other than the heat death of the universe happening first.

The only question is whether researchers will pursue it long enough to reach that point.

so...eternal researching, then.

keenir wrote: ↑Tue Jun 24, 2025 10:03 pmwhy not? you assume that the humans wouldn't invest in defense, so why are the uploaded conciousnesses (presumably of humans, right?) and AIs not given the same presumption? or do you enjoy playing tennis with the net down?

My concern is that the computronium faction would have numerous advantages because they aren't weighed down by biology. Zompist asked why anyone would want to give up corporeal pleasures to live in a computer and my answer is that the resulting efficiency would give them a massive economic and military advantage. For those who value growth for the sake of growth and power as the object of power, that is quite a strong incentive.

There is nothing that prevents it from happening.

other than the heat death of the universe happening first.

The only question is whether researchers will pursue it long enough to reach that point.

so...eternal researching, then.

Nobody has shown me any insurmountable obstacles to AI developing human-level intelligence, though. Considering how far AI has progressed, especially over the past few years, it seems reasonable to assume that it will eventually reach our cognitive capabilities. Yet so many people assert based on seemingly nothing that it will hit a wall before it reaches that point.

malloc wrote: ↑Tue Jun 24, 2025 10:21 pm

keenir wrote: ↑Tue Jun 24, 2025 10:03 pmwhy not? you assume that the humans wouldn't invest in defense, so why are the uploaded conciousnesses (presumably of humans, right?) and AIs not given the same presumption? or do you enjoy playing tennis with the net down?

My concern is that the computronium faction would have numerous advantages because they aren't weighed down by biology. Zompist asked why anyone would want to give up corporeal pleasures to live in a computer and my answer is that the resulting efficiency would give them a massive economic and military advantage. For those who value growth for the sake of growth and power as the object of power, that is quite a strong incentive.

Except once you exist in computronium, you are not weighed down by biology, as you yourself point out. Power and growth are desires of organicness, not universals.

There is nothing that prevents it from happening.

other than the heat death of the universe happening first.

The only question is whether researchers will pursue it long enough to reach that point.

so...eternal researching, then.

Nobody has shown me any insurmountable obstacles to AI developing human-level intelligence, though.

we have. more than once each.

and you have never listened.

Considering how far AI has progressed, especially over the past few years, it seems reasonable to assume that it will eventually reach our cognitive capabilities.

yes, eventually...after the heat death of the universe.

[quote[Yet so many people assert based on seemingly nothing that it will hit a wall before it reaches that point.[/quote]

are you done playing with the strawman?

zompist wrote: ↑Mon Jun 23, 2025 7:49 pm

malloc wrote: ↑Mon Jun 23, 2025 7:10 pm 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.

Here's what I got when I attempted to manipulate Verdurian: https://docs.google.com/document/d/1rfF ... sp=sharing

For comparison, here's the same operation on a natural language: https://docs.google.com/document/d/1spq ... sp=sharing

Sorry the second one is so much funnier.

rotting bones wrote: ↑Mon Jun 23, 2025 6:03 pm

zompist wrote: ↑Mon Jun 23, 2025 5:14 pm 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.

So is part of the challenge that there are in fact several seemingly very different Bengali languages for LLMs to get to grips with, and they're not well-labelled, so the LLM doesn't immediately know which it is dealing with?

zompist wrote: ↑Tue Jun 24, 2025 2:59 pm

malloc wrote: ↑Mon Jun 23, 2025 9:07 pm

zompist wrote: ↑Mon Jun 23, 2025 7:49 pmNope. 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 you think Sanskrit or Hebrew are "difficult" while English is easy, you were asleep in your linguistics class. There's no good argument for some languages being more difficult than others. Every language can be learned by a human child in its first six years or so.

It makes sense to me that natural human languages should present roughly equal difficulties to human beings, as an atypically easy one would have scope to add more difficulty. However, should I be unsurprised that LLMs don't have a different difficulty ranking to humans?

And of course, a six year old's knowledge is merely adequate - it seems that many adults never really master their native language.

malloc wrote: ↑Tue Jun 24, 2025 8:13 pm

zompist wrote: ↑Tue Jun 24, 2025 8:03 pmExcept that you obviously know nothing about my SF world, so shut up about it. Write your own dystopia where things can fall apart the way you want it to.

Just out of curiosity, though, how do they solve the problem I mentioned?

I think you have to refine what you actually want to know.

* How the Incatena developed? Start with my web page, or ask in the Almea forum.
* How conworlders/SF authors envision a better future? Read more; Iain Banks is the classic start. Or start a thread in Conlangery.
* What is likely to happen in this century? A buncha stuff; start a new thread here if you really want to know. (But ask, don't preach your own dystopia.)
* What should be done about computronium? The facile answer is "burn it with fire." Less flippantly: this is conworlding, and your nightmare projections are no better than anyone else's-- nor are they even the most important possible apocalypse.

Richard W wrote: ↑Wed Jun 25, 2025 1:44 pm It makes sense to me that natural human languages should present roughly equal difficulties to human beings, as an atypically easy one would have scope to add more difficulty. However, should I be unsurprised that LLMs don't have a different difficulty ranking to humans?

They already do: it's based on the size of the training data. Humans don't need that much raw data; they evidently have tricks up their crania that LLMs don't have.

For any other hypothesis ("LLMs don't like Semitic languages"), we would of course need data that they have difficulty unrelated to corpus size, or analyze the weights and see if there are numerical patterns. There's probably a dissertation in it for someone.