this post was submitted on 19 Oct 2023
540 points (96.6% liked)
Technology
59179 readers
2130 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
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
Nah, nah to all of it. LLM is a parlor trick and not a very good one. If we are ever able to make a general artificial intelligence, that's an entirely different story. But text prediction on steroids doesn't move the needle.
The best ones can literally write pretty good code, and explain any concept on the Internet to you that you ask them to. If you don't understand a specific thing about their explanation, they can add onto their explanation, and they can respond in the style you want (explain as if I'm ten, explain as if I'm an undergrad, etc).
I use it literally every day for work in a somewhat niche field. I don't really agree that it's a "parlor trick".
LLMs are awful for facts, because they don't understand what facts are. You should never rely on them if you require factual correctness.
They are OK for text summation, formatting and just making shit up. For summation a human with experience still produces nicer output, because they understand the content and don't just look at words. As for making shit up you will get the statistically most likely output, so it's usually trite and boring. I think the progress is amazing, but there are still so many problems to be solved.
Right now I use them for boiler plate stuff, like writing a text with some parameters and then I polish it. For code I find them quite useless, because with an IDE I can write boiler plate just as fast as when I polish the prompts until the LLM delivers useful stuff. And with the IDE I don't get references to methods or entire libraries that just don't exist.
It's actually great for dnd to produce NPC dialogue or names on the fly. We also tried using it to calculate area of effect spells, ie "how many average sized humans in armor with swords could fit in a circle with a diameter of 30ft." We were rolling with it before someone pointed out that it didn't calculate the area of a circle correctly, however it got the rest more or less accurate. So we don't use it for that anymore, and it's funny how what often appears to be the simplest component of a question is the thing it most often gets wrong.
People are also kind of shit at facts. There are so many facts, and many of them aren’t practical for every person who needs to assess a fact’s accuracy to do so. But it isn’t structurally impossible to mimic how humans learn how to gauge truthfulness, we just have to be prepared for the idea that it will be bound by the limitations of language, as well as the risk inherent in trusting data that it has not independently verified.
I use LLMs for having things explained to me, too.. but if you want to know how much salt to pour in that soup, try asking it about something niche and complicated you already know the answer to.
They can be useful in figuring out the correct terminology so that you can find the answer on your own, or for pointing some very very obvious mistakes in your understandings (but it will still miss most of them).
Please don't use those things as answer machines.
I'm going to use those things as answer machines and you can't stop me.
Jokes aside, I always validate what chatbots tell me, not even just important things. I use GPT-4 for work and 90% of the time it can show me how to use very specific functions in complex ways, but yesterday (for the first time in awhile) it made up a function that didn't exist. To its credit, I said, "Are you sure about [function]?" and it said, "I'm sorry, I got confused. That function doesn't exist. However, look into X, Y, Z for further resources" and I did and they were the correct things to look into.
If you press it the same way again ("are you sure the function doesn't exist?"), there is a high chance it will "rectify" its rectification.
It's just regurgitating info that people already know.
In a small convenient package, which speeds up the process.
No they can't. Your phrasing is misleading. It's a Chinese Room test output and nothing more. I had an Encarta CD that could do rudimentary version of this in 1995. That was more impressive, tbh.
If you're really comparing LLM's to your Encarta cd from 1995 and saying the Encarta CD was the superior experience...
I'm afraid there's not much left for us to discuss.... Our views are too far apart.
Sam Altman (Creator of the freakish retina scanning based Worldcoin) would agree, it seems. The current path for LLMs and GPT seems to be in something of a bind, because to seriously improve upon what it currently does it needs to do something different, not more of the same. And figuring out something different could be very hard. https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/
At least that's what I understand of it.
He's not saying "AI is done, there's nothing else to do, we've hit the limit", he's saying "bigger models don't necessarily yield better results like we had initially anticipated"
Sam recently went before congress and advocated for limiting model sizes as a means of regulation, because, at the time, he believed bigger would generally always mean better outputs. What we're seeing now is that if a model is too large it will have trouble producing truthful output, which is super important to us humans.
And honestly, I don't think anyone should be shocked by this. Our own human brains have different sections that control different aspects of our lives. Why would an AI brain be different?
Future of AI is definitely going towards Manager/Agent model. It allows for an AI to handle all the tasks without keeping it to one model or method. We’re already seeing this with ChatGPT using Mathematica for math questions. Soon we can see art AI using different models and methods based on text input.
In humans, abstract thinking developed hand in hand with language. So despite their limitations, I think that at least early AGI will include an LLM in some way.
I've been having a lot of vague thoughts about the unconscious bits of our brains and body, in regards to LLMs. The parts of our brains/neurons that started evolving back in simple animals as basically super primitive ways to process visual/audio/whatever input.
Our brains do a LOT of signal processing and filtering that never reaches conscious thought, that we can't even reach with our conscious thought if we tried, but which is necessary for our squishy body-things to take in input from our environment and turn it into something useful instead of drowning in a screeching eye-searing tangled mess of chaotic sensory input all the time.
LLMs strike me as that sort of low-level input processing, the pattern-recognition and filtering. I think true generalized AI would have to be built on pieces like this--probably a lot of them. Ways to pluck patterns out of complex but repeated input. Like, this stuff definitely isn't self-aware, but could eventually end up as some sort of processing library for something else far down the line.
Now might be a good time to pick up Peter Watts' sci-fi book Blindsight. He doesn't exactly write about AI in it, but he does write about a creature that responds to input but isn't exactly conscious like you or I.
This is what I meant.
I just got the EPUB, thanks. Looking forward to reading it.
FYI Blindsight is free for audible members.
Parlor trick is a perfect description.
People don't get that these things aren't anymore intelligent than their smartphones predicting the next word. The main difference is instead of a couple words it has thousands to choose from.
Half of the trick is how it uses the prompt to decided what words to start with.
That is not how it works. Your smartphone has all the dictionary available, same as LLM. It is simply something very different. People super confidently discussing about AI on lemmy are the real hallucinating parrots
There is an inverse relationship between the intelligence of a person and their amazement at what these large language models can produce.
People who aren't amazed at what LLMs produces have no clue how complicated it is to generate plausible language in the first place. Dunning–Kruger and all that.
The ability to generate plausible language was a lack of compute power. The actual programs running the LLM is not complicate.
The model that is produced is complex.
Its training required compute power that was not previously available but the math/code behind these systems is not complex. They are resource intensive. There is a difference that a layperson often cannot comprehend.
So what is it now?
Are LLMs more intelligent than your smartphone, or do they need a lot more computer power to produce the same thing as your smartphone?
I heard the same for people who downvote on lemmy when notified about being an exemplification of the dunning Kruger effect
Have you ever even bothered to play around with any of the LLMs or are you just parroting what you heard in badly written articles?
The fact that the LLM predicts the next word does in no way shape or form limits its intelligence. That's after all the same thing you do while writing your post.
These idiotic claims about AI not being intelligent really make me questions if humans are.
Yes. I've used them. I have used it beyond the point of it hallucinating.
I am also a software engineer and have deeper understanding of how these systems work than your average user.
The software community tends to approach these things with more caution than the general population. The media overblows the capabilities of these systems.
A more concrete example is autonomous vehicles which were promised for decades and even now with a form of them on the road, they are still closer to remote controlled vehicles than the intelligent self contained systems we have been promised.
The difference between predictive text on a smart phone and predictive text of an LLM is my smart phone is predicting what I am likely to type next based on things i have typed in the past, while the LLM is predicting what comes next based on a larger body of work from source pulled from all across the internet. The LLM is then tuned by humans. This tuning step is under reported.
The LLM is unable to determine the truth of its own output. I would argue that is a key to claiming intelligence but determining what intelligence means is itself a philosophical question up for debate.
Yeah exactly and a great way to see this is by asking it to produce two viewpoints about the same subject, a negative and positive review of something you're familiar with is perfect. It produces this hilarious "critic" type jargon but you can tell it doesn't actually understand. Coincidentally, it's drawing from a lot of text where the original human author(s) might not understand either and are merely themselves re-producing a jargon-heavy text for an assignment by their employer or academic institution. If AI can so accurately replicate some academic paper that probably didn't need to be written for anything other than to meet publishing standards for tenured professors, then that's really a reflection on the source material. Since LLM can only create something based on existing input, almost all the criticisms of it, are criticisms that can apply to it's source material.
It's not really "intelligent" though, as in it's not thinking about what it's doing. What AI will do very well is reproduce jargon, and if it's jargon that we associate with intelligence then it appears intelligent. Academic papers for instance it can do a very convincing job because that format is so repetitive and jargon heavy.
You can do an experiment by asking it to produce a positive review of something niche and academic you're familiar with, then ask it to produce a negative review of the same subject. It will produce convincing dialogue for either scenario, but it does not know which is more true/accurate, and it will come across as a student writing about something they didn't do the reading for.
The "question if humans are [intelligent]" is the more relevant thing here. We're constantly expected to communicate with thoughtlessly reproduced jargon, and many of us can do this very well in a way that gives the impression of intelligent thought. The fact AI can do this, and that people are concerned about how intelligent it appears, is more a reflection on how derivative our notions of intelligence can be in these settings.
The fact that you believe an LLM is "intelligent" tells me you have no clue how they work and your comments on the matter can be ignored.
Oh look, another parrot.
Still waiting for any of you to actually define "intelligent" in some way that ChatGPT fails at or are you just going to pull the old boring "human exceptionalism"-card out of the hat?
LLM's are like an interface to allow computers to talk to humans.
They are a necessary step in order to create general AI, because a general AI that can't generate text wouldn't be able to convey what they learned.
Nah, nah to your understanding of LLM's
No it's not true intelligence. Yes, it makes humans much faster at their work
It has really sped up my work, especially when coding in unfamiliar languages.
It's silly to compare it to a parlor trick or text prediction.