574
this post was submitted on 26 Jul 2023
574 points (99.0% liked)
Technology
34788 readers
344 users here now
This is the official technology community of Lemmy.ml for all news related to creation and use of technology, and to facilitate civil, meaningful discussion around it.
Ask in DM before posting product reviews or ads. All such posts otherwise are subject to removal.
Rules:
1: All Lemmy rules apply
2: Do not post low effort posts
3: NEVER post naziped*gore stuff
4: Always post article URLs or their archived version URLs as sources, NOT screenshots. Help the blind users.
5: personal rants of Big Tech CEOs like Elon Musk are unwelcome (does not include posts about their companies affecting wide range of people)
6: no advertisement posts unless verified as legitimate and non-exploitative/non-consumerist
7: crypto related posts, unless essential, are disallowed
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
The thing is, the LLM doesn't actually know anything, and lies about it.
So you go to How Stuff Works now, and you get bullshit lies instead of real information, you'll also get nonsense that looks like language at first glance, but is gibberish pretending to be an article. Because sometimes the language model changes topics midway through and doesn't correct, because it can't correct. It doesn't actually know what it's saying.
See, these language models are pre-trained, that the P in chatGPT. They just regurgitate the training data, but put together in ways that sort of look like more of the same training data.
There are some hard coded filters and responses, but other than that, nope, just a spew of garbage out from the random garbage in.
And yet, all sorts of people think this shit is ready to take over writing duties for everyone, saving money and winning court cases.
Yeah, this is why I can't really take anyone seriously when they say it'll take over the world. It's certainly cool, but it's always going to be limited in usefulness.
Some areas I can see it being really useful are:
That's about it.
LLM generated scams are going to such problem. Quality isn't even a problem there as they specifically go for people with poor awareness of these scams, and having a bot that responds with reasonable dialogue will make it that much easier for people to buy into it.
AI tools can be very powerful, but they usually need to be tailored to a specific use case by competent people.
With LLMs it seems to be the opposite, where people not competent for ML are applying it for the broadest of use cases. Just that it looks so good they are easily fooled and lack the understanding to realize the limits.
But there is a very important Usecase too:
Writing stuff that is only read and evaluated by similiar AI tools. It makes sense to write cover letters with ChatGPT because they are demanded but never read by a human on the other side of the job application. Since the weights and stuff behind it serm to be similiar, writing it with ChatGPT helps to pass the automatic analysis.
Rationally that is complete nonsense, but you basically need an AI tool to jump through the hoops made by an AI tool applied by stupid people who need to make themselves look smart.
It isnt going to take over, its being put in control by idiots.
Here is an alternative Piped link(s): https://piped.video/watch?v=oqSYljRYDEM
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I'm open-source, check me out at GitHub.
Absolutely. Creating new documentation will always be a human sport.
It could be AI sport when we actually have an general purpose AI. That based on people working on llm and gpt, would take between 6 years and never happening.
It's not easy to create a super ai who's realistically smarter than humans in every aspect.
Just like the mutant Olympics that we have today.
I mean I would say maybe "regurgitating their training data" is putting it a bit too simple. But it's true, we're currently at the point where the AI can mimic real text. But that's it - no one tells it not to lie rn, the programmatic goal of the AI is to get indistinguishable from real text with no bearing on the truthfulness of the information whatsoever.
Basically we train our AIs to pretend to know, not to know. And sometimes it's good at pretending, sometimes it isn't.
The "right" way to handle what the CEOs are doing would be to let go of a chunk of the staff, then let the rest write their articles with the help of chatgpt. But most CEOs are a bit too gullible when it comes to the abilities of AI.
I've graded papers from students who obviously used chatGPT to write them. They were a pass at best. Zero critical synthesis of ideas and application of them to the topic. I'm sure chatGPT has its uses but people really overhype its writing ability. There's more to writing than putting words in the right places.
Literally predictive text but for whole articles.
It doesn't know the limits of it's knowledge or indeed know anything. It just "knows" what an answer smells like. It even "knows" what excuses are supposed to look like when you call it out.
This is a very good write up about how ChatGPT works.
Just like your average human journalist. If you ever read an article from not specialist journal on a topic you are familiar with - you know. This seems actually where LLM are very similar to how human brain works - if we don't know something, we come up with some bullshit.
Even medium human writers can comprehend their work as a whole, though. There is a cohesiveness even to the bullshit. The LLM is just putting words down that match the prompt. It's rng driven, readable Lorum Ipsum.
If the results were still edited afterwards, there may be some merit to the output, but any company going full LLM isn't looking for quality. They want to use it to churn out endless content that they simply can't get from even a team of humans. More than could be edited even if they kept editors on staff.
Sure, but a lot of humans are rather bad writers.
That is true for 24h news cycle of online media, regardless LLM.
Bad writing is still a step above rng junk, imo.
Yes, that was my point. Setting up your company to put out more content than can possibly be processed by humans is a glaring sign of their values - ie quantity far above quality.
I'v read writing worse than GTP. I had to help someone write an essay - and I just wrote it for him in the end, because he absolutely lacked the skills to write a long meaningful text. At at the same time - genius of a percussionist.
Do you think that person was signing up for jobs writing for blogs or content farms?
Have you read some low quality journalism? The whole yellow press can be replaced with GTP and no one would ever see a difference.
Ok, so do you wanna talk about your terrible writing partner in school? Or "yellow press"? Or maybe the topic of the article, which isn't journalism in the slightest? Or how about my point, which was, again, that even bad writers have context, as opposed to an LLM which is just filling in the arbitrary patterns it's programmed to delineate. Readability is not what I'm talking about.
That's how you get the room
That's how you get the room
Dude, what's with aggression? We just having a conversation that floats along. I'm talking about general LLMs capabilities to write text - which are in my opinion comparable to human writing, since again - a lot of people lack the same things LLMs generated texts are lacking. And I had some examples. No idea what made you so upset.
You brought up several different, unrelated topics and pretty much ignored anything I said to disprove something I never claimed. That is frustrating to deal with.
Except you are the one who responded to me. And if there is a point you made I overlooked - I will gladly answer it. I also didn't disprove anything - just voiced my opinion. I'm not interested in a debate club and winning arguments, just sharing opinions and trying to understand others.
The top comment is about how LLMs don't comprehend what they're writing, and your first comment (as I read it) was about how LLMs work how human brains do. My point was that they don't and why, not about how good or bad humans or machines are at writing, which is what you kept bringing up, hence the frustration.
My first comment is, that there are enough humans out there that don't really comprehend what they are writing and often also make shit up as they go. I was not talking about the underlying mechanism, which is rather speculative since we have little idea how complex functions of the brain - like text generation, work. Just making a humorous light hearted comparison.
Our conversation is a nice illustration how, maybe we as humans aren't as good at understanding text - as we might think. (Again - that is a light hearted comment and not some profound complex observation).
To be clear, I'm not talking about underlying mechanisms, either, but the approach to the task. A human writer, even one bad at writing and not understanding the topic, will approach the writing with a goal and write to that goal and topic. They can even research if they so choose, but even if they are just making things up, there is intent and context there.
An LLM doesn't have any of that. It literally just generates words that match certain patterns, with no actual purpose or goal. It may have been programmed with a goal in mind, but it doesn't have one of its own. It can't reason, it can't research, it can't make decisions. I think that is an important distinction that people who are just saying "Who cares? It's all bad writing anyways" are missing.
You never made an experience of having to writer for a topic you genuinely don't care about, where you just string along words, vaguely related to the topic to make specific word count? I'm not arguing that all human writing is like this - people are definitely capable of writing text with purpose and context, at least some. But that is not all human writing.
And exactly that was my point, that humans often do the same. Not all the time. But it definitely happens, especially in professional writing where you maybe have to write about a topic you don't understand or care about.
And again, there are tons of people out there that can't do this things either. It's like a very intelligent chimpanzee is smarter than a very dumb human. So are LLMs better at generating text than quite a lot of humans.
Then you're missing the distinction as well.
I think you are just in a bubble where you don't have much contact with dumb people. I would have thought the same during my university time, but chosing not academia or science for that matter opened a whole new wold of ignorance for me. By the way - maybe our brain is just a Markov Model prediction machine and consciousness ist just an illusion, would not surprise me much if we are not that different from LLMs. But that is even more speculative.
So modern journalists were redundant all along?
But yeah, the quality of what is passing as journalism now is often ridiculous. But the only way to combat this is by having editors that are knowledgable about topics. But it seemed editors were the first people laid off, when internet articles became a thing.
24 hours news cycle of online media creates junk journalism on new level. Good journalism needs time and can't spit out news articles every minute of the day. Editors won't help, because it's just not possible to do good journalism on that scale. But jeh - in general with AI, the jobs will shift more to editing. Which will be extremely soul-draining, going though tons of AI generated bullshit