this post was submitted on 23 Jun 2024
344 points (97.8% liked)
Technology
59398 readers
2959 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
The game plan is to scrape, store and utilise as much data as possible regardless of conventions, best practice, license agreements etc until specifically regulated to stop.
At that point, a few early companies will have used vast swathes of data that any newly established company is banned from also using
And they will be “unable” to purge it.
Hoping the EU drops GDPR 2 requiring them to delete the entire model if it infringes or something.
Expecting the US to meaningfully regulate US companies is like expecting.........
You know what, even including physical impossibilities, I'm struggling to think of anything less likely
I've yet to understand how the hell they get away with "I don't know how it works". Either figure out how it works or stop using it, shithead. It's software not magic beans.
There's lots of complicated fields out there, none of them get a pass for "I don't know how my drugs work" or "I don't know how my rockets work". That's absolutely ridiculous.
It’s just how machine learning has been since ever.
We only know the model’s behavior by testing, hence we only know more or less the behavior in relation to the amount of testing that was done. But the model internals has always been a black box of numbers that individually mean nothing and if tracked which neurons fire here and there it’ll appear just random, because it probably is.
Remember the machine learning models aren’t carefully designed, they’re just brute-force trained for a long time and have the numbers adjusted again and again whenever the results look closer or further away from the desired output.
If the models are random then we shouldn't be trusting them to do anything, let alone serious applications. If any other type of software told us that it's based on partially random results we'd say "get that shit out of here, I want my software to work first time, every time".
"Statistically good enough" works for some applications but not for others. If a LLM finds a formula that has an 80% chance to be the cure for cancer or a new magical fuel or some amazing new material that's cool, we're not going to look the gift horse in the mouth.
But using LLM to polute the web with advertising texts that are barely inteligible, and using it as a pretext to break copyright in the process, who does that help? So far the only readily available commercial application for LLMs has been to spit out semi-nonsense so that a bunch of bottom-crawling parasitic industries can be enabled to keep on pinching pennies and shitting up everything they touch.
Which, ironically, it will help them to hit bottom all the faster, so in a strange way it's a positive return, but the problem is they're going to take down a lot of useful things with them.
That's not the reason we shouldn't be using them for anything other than generating lorem ipsum style text or dialogue for non quest critical NPCs in games.
The reason is that, paraphrasing Neil Gaiman, LLMs don't generate information, they generate information shaped sentences.
Specifically, an LLM takes a sequence of characters (not a word or text; LLMs have no concept of words, or text, or anything else for that matter; they're just an application of statistics on large volumes of sequences of characters; no meaning or intelligence involved, artificial or not)... as I was saying, an LLM takes a sequence of characters, pushes it through its model, and outputs the sequence of characters most likely to follow it in the texts its model has been trained on (or rather, the most likely after discarding the ones its creators have labelled as politically incorrect).
That's all they do, and they'll excellent at it (or would be if it weren't for the aforementioned filters), but that'll never give you a cure for cancer unless there already was one in their training data.
They take texts written by humans, shred them, and give you their badly put back together dessicated corpses, drained of any and all meaning or information, but looking very convincingly (until you fact check them) like actually meaningful or informative texts.
That is what makes them dangerous. That and the fact that the bastards selling them are marketing them for the jobs they're least capable of doing, that is, providing reliable information.
(And that's while they can still be trained on meaningful and informative texts written by humans — inasmuch as anything found on reddit, facebook, or xitter can be considered to be meaningful or informative —, but given that a higher and higher percentage of the text on the internet is being generated by LLMs soon enough it'll be impossible to train new models on anything but 99% LLM generated garbage, at which point the whole bubble will implode, as anyone who's wasted time, paper, and toner playing with a photocopier or anyone familiar with the phrase “garbage in, garbage out” will already have realised... which is probably why the LLM peddlers are ignoring robots.txt and copyright laws in a desperate effort to scrape whatever's left of the bottom of the barrel.)
That is besides the point. A random number generator is more or less random but it still has applications.
The problem is not them being random, it's hiding that they are being random so they can be used for applications where randomness is not a feature.
They are not random, that's the point. They're entirely deterministic and very precise, and they aren't hiding anything; they will give you the most likely (not blacklisted) sequence of characters to follow your input according to their model. What they won't give you is information, except by accident.
If they were random (hidden or not) they'd be harmless, no one would trust them any more than one of those eight ball toys, or your average horoscope.
The issue is that they're very not random, so much that there's no way to know if what they are saying bears any accidental semblance to the truth without fact checking... and that very soon they'll have replaced any feasible way to fact check them, since all the supposed "facts" we'll have access to will have been generated by LLMs train on LLM generated garbage.
Uh, we don't really know how our drugs work (especially the older ones). We have a vague understanding of their mechanisms, but we really don't know how they work. We don't even have a clear idea of what the structures of most drugs look like, and how they interact with their binding sites.
Luckily, we don't actually have to know how they work, to know that they work. Instead we use clinical trials and real world evidence to support their use.
(Fun fact: there's actually a branch of drug development called phenotypic drug discovery which actually does away with the understanding of the mechanisms altogether. )
I’m in the US so yeah…. Even if the current of future GDPR requires deletion I guarantee it’ll still be used in the US. I have no faith that any US company will follow rules like that. Any fines are just looked at as the cost of doing business.
Or they'll "purge" it and somehow the canaries will end up in the model anyway
It's like weapons testing. You only move to ban testing after you've developed it yourself.
Whatever happened to those "nightshade" images that poison the model?
You mean that work that took open source software, closed sourced it and refused to release the source code and the poisoning only worked against one specific open source model (stable diffusion)? I don't think that's going to come riding to anyone's rescue.
They only kinda work but more importantly they need mass adoption to actually poison training data. Most people aren't going to add another step to their posts so probably the only way to mass adopt it is to have platforms automatically poison uploaded images. I wonder if reposts on a platform like that would start to have noticable artifacts in the images like jpeg but different
Same approach as all the other 'disruptive' new companies that ignore industry standards, rules, and laws.
I'd say they are pushing for regulations behind the scene because they know it gives them an instant monopoly.
They are already pass the door, they can afford to shut it behind them to own the room. Having to send checks to websites like Reddit and Getty in the future is a small price to pay.