69
this post was submitted on 22 Jun 2024
69 points (84.2% liked)
General Discussion
11880 readers
6 users here now
Welcome to Lemmy.World General!
This is a community for general discussion where you can get your bearings in the fediverse. Discuss topics & ask questions that don't seem to fit in any other community, or don't have an active community yet.
πͺ About Lemmy World
π§ Finding Communities
Feel free to ask here or over in: [email protected]!
Also keep an eye on:
For more involved tools to find communities to join: check out Lemmyverse and Feddit Lemmy Community Browser!
π¬ Additional Discussion Focused Communities:
- [email protected] - Note this is for more serious discussions.
- [email protected] - The opposite of the above, for more laidback chat!
- [email protected] - Into video games? Here's a place to discuss them!
- [email protected] - Watched a movie and wanna talk to others about it? Here's a place to do so!
Rules
Remember, Lemmy World rules also apply here.
0. See: Rules for Users.
- No bigotry: including racism, sexism, homophobia, transphobia, or xenophobia.
- Be respectful. Everyone should feel welcome here.
- Be thoughtful and helpful: even with βsillyβ questions. The world wonβt be made better by dismissive comments to others on Lemmy.
- Link posts should include some context/opinion in the body text when the title is unaltered, or be titled to encourage discussion.
- Posts concerning other instances' activity/decisions are better suited to [email protected] or [email protected] communities.
- No Ads/Spamming.
- No NSFW content.
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
LLM peddlers have turned the term AI into a synonym for LLM, despite LLMs being as far from anything resembling true AI as eliza was.
Which is tragic, because funding and research into this evident dead end is syphoning away any time or money that could be spent on actual AI research, and once the LLM bubble bursts it'll poison the public's opinion on AI (they won't know or care that LLMs have nothing to do with AI) and prevent any investments or research on real AI for decades.
I don't personally think there's necessarily anything wrong callin LLM an AI. After all; autocorrect could be considered an AI aswell. The main issue is that people speak of AI when they actually mean AGI or ASI. That's why they have unrealistic expectation from things like GPT4.
Wether the LLM route is a dead end I have no idea. It might but it might not. Because LLMs sometimes makes mistakes that are so obvious to us humans it seems to make many people dismiss it as fundamentally flawed but I don't think we're giving it the credit it deserves. Humans make mistakes aswell but if there was a person with an equivalent capability to source expert level knowledge on such an insanely wide range of topics it would seem insane to describe them anything else but a genious.
There's nothing resembling intelligence, general or not, in any autocorrect implementation so far, including LLMs.
LLMs don't make mistakes. If you think they do, you're completely misunderstanding what LLMs are, how they work, and what they do (probably because of the aforementioned misinformation by LLM peddlers trying to equate them to intelligence, artificial or not).
LLMs simply give you the most statistically likely word to follow a given text. Then they do it again, adding the word they generated in the previous cycle to the text. That's all they do, they're excellent at it, and they don't make mistakes, the word they output will be the most statistically likely, regardless of whether it makes sense or not (though attempts by their peddlers to keep them politically correct might cause them to discard the first several most likely words, leaving them able to only output a significantly unlikely β but hopefully politically correct β one, which might seem like a mistake to the user).
You seem to be assuming that LLMs are trained on knowledge. They're not. They're trained on text. They have no idea what the text means (they don't even have anything to have ideas with), and they don't care (nor have more ability to care than a desk lamp).
They have a model of what words (meaning sequences of characters, not concepts with any actual meaning) may come after certain others, they push the input sequence of meaningless characters through that model, and out comes the most statistically likely meaningless sequence of characters to follow said text. That's all.
Paraphrasing Neil Gaiman, βLLMs don't produce information. They produce information shaped sentences.β
They produce the dessicated corpses of the texts they were fed, shredded and put back together, drained of any actual information but indistinguishable enough from texts containing actual information to give the illusion of also containing it.
They're great as an alternative to lorem ipsum, or possibly as speech generators for non quest critical NPCs in games, but they're extremely dangerous for anything else, especially the uses LLM peddlers are peddling them for.
Maybe. Maybe not. Intelligence and consciousness may also just be an emergent feature of information processing. LLMs process information. Whose to say that there's zero chance of the lights coming on one day. I personally tend to be distrustful of people who speak as if they have absolute knowledge.
AI can be extremely dangerous in either case. LLMs are no different from that perspective.
The key point is that LLMs don't process information, as we see it. The knowledge they have is predigested, and embedded into the text they were trained on.
Don't get me wrong, they are a big step towards a true AI, but they cannot do some things that seem fundamental to intelligence. The best analogy is that they are a lobotomized speech centre. They can put on a veneer of being intelligent and self aware, but it's a veneer.
I personally suspect they will be a critical component to a future AI, but are a dead end path on their own.
No, they don't. They merely tell you which sequence of characters comes most often in their training set after the sequence of characters you gave them. That's all. No processing going on, no information being generated or retrieved other than statistical trivia about their training set.
General AI could be dangerous because it could be smarter than us while having interests, objectives, and morals that could clash with our own, causing it to antagonise us.
That's obviously impossible for LLMs, which have as much intelligence, interests, objectives, or morals as your average paperweight.
LLMs are dangerous because they're good enough at sounding like they know what they're saying that you people actually believe them to be intelligent (and the fact that the bastards selling them are using their apparent intelligence as their main selling point obviously doesn't help either), and they can be convincing enough that when they randomly tell you to get a bleach and ammonia enema to help with that headache you might actually believe them since by that point there'll be no way left to check your facts. Which, hey, fair enough, natural selection and all that... but at some point one of you is going to fart that chlorine gas in my general vicinity, and that isn't so good.