Let's chat about AI!

The Something Awful Forums > Discussion > Debate & Discussion > Let's chat about AI!

GABA ghoul: Oct 29, 2011

Seph posted:

This will basically destroy any remaining credibility that videos have on social media. There will be zero cost to create inflammatory, viral propaganda videos to serve whatever purpose you desire.

“A Biden supporter stealing ballots out of a ballot box in a conservative district”

“A black man assaulting a white woman in broad daylight”

“A CNN broadcast showing a person wearing a MAGA hat shooting up a school”

You will have to decide which people and institutions you trust implicitly since nothing will be independently verifiable outside of what you have seen with your own eyes. Anything outside of your trust network will automatically discarded. Our news and media landscape will become even more fragmented and disconnected as these trust networks develop.

This is already happening somewhat since there have been plenty of manipulated or heavily edited videos out there. But the rate was slow enough that it was possible for people to independently verify videos. You could see someone on Twitter debunking a fake Ukrainian war crime, for example. Now with genAI there will be a flood of fake videos making independent verification impossible. The default will be that everything is fake unless someone you trust says it is true.

Sounds like an improvement tbh. A social media where the first post under every video is always "source your video or it didn't happen" may be better than what we have now.

I'm coming here after reading a Twitter post where some random guy posted a video he claims shows Navalny and an MI6 agent plotting a coup against Putin. None of the people in the video look even remotely like Alexei Navalny nor do they discuss anything suspicious, illegal or unusual, yet the replies are 90% ultra gullible people going on about how he deserved to be in jail for this and what a scandal it is that the lamestream media is not reporting it. It would be a huge improvement if more of these people were conditioned to outright dismiss the video, unless someone who can actually Google what Navalny looks like or form a coherent argument why the discussed content is suspicious (aka a journalist, even if it's a right wing one) presents it. It's how the world mostly used to work before the internet. You couldn't just watch some random videos with unknown origins and context.

# ¿ Feb 17, 2024 22:02

Adbot: ADBOT LOVES YOU

# ¿ May 10, 2024 03:29

GABA ghoul: Oct 29, 2011

Main Paineframe posted:

Doesn't really matter what context it's in, a LLM is a LLM. And Siri is definitely not a LLM.

How well a LLM handles different languages depends pretty much exclusively on its training set. And I think you're making the mistake of anthromorphizing it here. LLMs don't work like human thought processes do. They don't really have a sense of "language" in the first place. The only thing they know about a given word is what words it tends to be used along with most often.

They don't know that "ernährungswissenschaftler" is German or that "nutritionist" is English - they just know that "ernährungswissenschaftler" tends to be used along with other words that are way too long and contain way too many consonants. Even though they have no concept of "ernährungswissenschaftler" being a German word, they'll tend to use it with other German words, because that's how it tends to be used in much of their training data.

Unless you use it on an otherwise English sentence along with words like "meaning" or "translation", in which case it'll respond with an English description of what it means, because any instances of "what does ernährungswissenschaftler mean?" in its training data will be followed by an English explanation of its meaning.

Circling back around to your original question, I'll repeat myself: it depends on its training data, because it's basically just repeating back words based on statistical analysis of its training set. They'll tend to use English words with English words and German words with German words, but that's because their training data will most likely use English words with English words and German words with German words. There's no fundamental understanding of language there. If their training data contains a lot of mixing English with German, then they'll be more likely to mix English with German themselves. Simple as that.

That's for responses, anyway. In terms of taking input, if you're typing the words in, then it shouldn't have any issue with multilingual input since its all just words, and the LLM doesn't have a real concept of language. But if you're speaking words, then you're not putting words directly into the LLM. You have to go through voice recognition first, and that's not an LLM. Moreover, it is usually much more complicated. Voice recognition tends to still be language-specific, because it's probably a couple orders of magnitude more complex than just crunching text.

You are making a lot of assumptions about how language processing works in human brains.

Natively bilingual people can effortlessly jump between speaking language A, language B or a random mixture of them in just a single sentence without any feeling that they are speaking different languages. It feels more like "language" is just another classification in the brain that each word and grammar rule has. Like an apple is associated with the category of "fruits" and broccoli with "vegetables", this word/grammar rule is associated with "English" and that with "German". Like I can effortlessly speak only about fruits, I can effortlessly use only English words.

I think it would be really strange if humans had some specific natural mechanism in the brain to separate languages because AFAIK it wasn't really a common scenario in our evolutionary past that you needed to use more than one language.

# ¿ Mar 15, 2024 13:19

GABA ghoul: Oct 29, 2011

Main Paineframe posted:

I'm not quite sure what you read from my post, but it sure as hell doesn't sound like what I wrote. I didn't make any assertions about how language processing works in human brains, nor did I ever claim that there's "some specific natural mechanism in the brain to separate languages". It's a ridiculous claim to make, which is exactly why I didn't make it, nor did I suggest anything even remotely similar to it.

This is what you wrote:

quote:

And I think you're making the mistake of anthromorphizing it here. LLMs don't work like human thought processes do. They don't really have a sense of "language" in the first place. The only thing they know about a given word is what words it tends to be used along with most often.

Humans don't have a "sense of language" either when speaking habitually, as any natively bilingual person will tell you. They freely interchange words and grammar of language A, language B or the pidgin mixture of both, without even being consciously aware of doing it. At no point is there any reasoning about the abstract concept of language A, B or the pidgin involved(unless you make a conscious decision to edit the generated speech in your head to strictly conform to one of them before you say it out aloud).

You assume that the way LLMs generate speech is different from the way humans do, which we just don't know. It all looks suspiciously similar at a first glance though.

# ¿ Mar 15, 2024 19:11

The Something Awful Forums > Discussion > Debate & Discussion > Let's chat about AI!