As junk web pages written by AI proliferate, the models that rely on that data will suffer.
Good.
this post was submitted on 24 Jul 2024
435 points (97.2% liked)
Technology
59381 readers
3164 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
“On two occasions I have been asked, 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.” - Charles Babbage
The business people adopting AI: "who cares what it's trained on? It's intelligent right? It'll just sort through the garbage and magically come up with the right answers to everything"
load more comments
(1 replies)
Of course modern UX design is very much based on getting the right answer with the wrong inputs (autocorrect, etc).
I believe Robustness was the term I learned years ago: the ability of a system to gracefully handle user error, make it easy to recover from or fix, clearly communicate what was wrong etc.
Of course, nothing is ever perfect and humans are very creative at fucking up, and a lot of companies don't seem to take UX too seriously. Particularly when the devs get tunnel vision and forget about user error being a thing....
Garbage in; Garbage out.
Shit-fueled ouroboros
You can't explain it!
Recycle the garbage that comes out... Still more garbage out.
Model degeneration is an already well-known phenomenon. The article already explains well what's going on so I won't go into details, but note how this happens because the model does not understand what it is outputting - it's looking for patterns, not for the meaning conveyed by said patterns.
Frankly at this rate might as well go with a neuro-symbolic approach.
load more comments
(18 replies)
Well, you've got a timestamped copy of much of the Web that existed up until latent-diffusion models at archive.org. That may not give you access to newer information, but it's a pretty whopping big chunk of data to work with.
Hopefully archive.org have measures in place to stop people from yanking all their data too quickly. As least not without a hefty donation or something. As a user it can chug a bit, and I'm hoping that's the rate-limiting I'm talking about and not that they're swamped.
That would go against the principal of the archive imo but regardless, if you take away all means of acquiring data freely, you are just giving companies like OpenAI and Google who already have copies of it an insane advantage.
AI isn't going away, we need to make sure we have free access to it as to not give our whole economy to a handful of companies.
AI making itself sick and worthless after flooding the internet with trash just gives me a warm glow.
Oh no, the AI are inbreeding.
⢀⣠⣾⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⠀⠀⠀⠀⣠⣤⣶⣶ ⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⠀⠀⠀⢰⣿⣿⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣧⣀⣀⣾⣿⣿⣿⣿ ⣿⣿⣿⣿⣿⡏⠉⠛⢿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⡿⣿ ⣿⣿⣿⣿⣿⣿⠀⠀⠀⠈⠛⢿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⣿⠿⠛⠉⠁⠀⣿ ⣿⣿⣿⣿⣿⣿⣧⡀⠀⠀⠀⠀⠙⠿⠿⠿⠻⠿⠿⠟⠿⠛⠉⠀⠀⠀⠀⠀⣸⣿ ⣿⣿⣿⣿⣿⣿⣿⣷⣄⠀⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⣴⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⣿⠏⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠠⣴⣿⣿⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⡟⠀⠀⢰⣹⡆⠀⠀⠀⠀⠀⠀⣭⣷⠀⠀⠀⠸⣿⣿⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⠃⠀⠀⠈⠉⠀⠀⠤⠄⠀⠀⠀⠉⠁⠀⠀⠀⠀⢿⣿⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⢾⣿⣷⠀⠀⠀⠀⡠⠤⢄⠀⠀⠀⠠⣿⣿⣷⠀⢸⣿⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⡀⠉⠀⠀⠀⠀⠀⢄⠀⢀⠀⠀⠀⠀⠉⠉⠁⠀⠀⣿⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⣧⠀⠀⠀⠀⠀⠀⠀⠈⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢹⣿⣿ ⣿⣿⣿⣿⣿⣿⣿⣿⣿⠃⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢸⣿⣿
I'd be very wary of extrapolating too much from this paper.
The past research along these lines found that a mix of synthetic and organic data was better than organic alone, and a caveat for all the research to date is that they are using shitty cheap models where there's a significant performance degrading in the synthetic data as compared to SotA models, where other research has found notable improvements to smaller models from synthetic data from the SotA.
Basically this is only really saying that AI models across multiple types from a year or two ago in capabilities recursively trained with no additional organic data will collapse.
It's not representative of real world or emerging conditions.
provenance requires some way to filter the internet into human-generated and AI-generated content, which hasn’t been cracked yet
It doesn't need to be filtered into human / AI content. It needs to be filtered into good (true) / bad (false) content. Or a "truth score" for each.
We don't teach children to read by just handing them random tweets. We give them books that are made specifically for children. Our filtering mechanism for good / bad content is very robust for humans. Why can't AI just read every piece of "classic literature", famous speeches, popular books, good TV and movie scripts, textbooks, etc?
It doesn’t need to be filtered into human / AI content. It needs to be filtered into good (true) / bad (false) content. Or a “truth score” for each.
That isn't enough because the model isn't able to reason.
I'll give you an example. Suppose that you feed the model with both sentences:
- Cats have fur.
- Birds have feathers.
Both sentences are true. And based on vocabulary of both, the model can output the following sentences:
- Cats have feathers.
- Birds have fur.
Both are false but the model doesn't "know" it. All that it knows is that "have" is allowed to go after both "cats" and "birds", and that both "feathers" and "fur" are allowed to go after "have".
It's not just a predictive text program. That's been around for decades. That's a common misconception.
As I understand it, it uses statistics from the whole text to create new text. It would be very rare to output "cats have feathers" because that phrase doesn't ever appear in the training data. Both words "have feathers" never follow "cats".
But the fact remains that it doesn't know what a cat or a feather is. All of this is still based purely on statistical frequency and not at all on actual meanings.
and that is exactly how a predictive text algorithm works.
-
some tokens go in
-
they are processed by a deterministic, static statistical model, and a set of probabilities (always the same, deterministic, remember?) comes out.
-
pick the word with the highest probability, add it to your initial string and start over.
-
if you want variety, add some randomness and don't just always pick the most probable next token.
Coincidentally, this is exactly how llms work. It's a big markov chain, but with a novel lossy compression algorithm on its state transition table. The last point is also the reason why, if anyone says they can fix llm hallucinations, they're lying.
Coincidentally, this is exactly how llms work
Everyone who says this doesn't actually understand how LLMs work.
Multivector word embeddings create emergent relationships that's new knowledge that doesn't exist in the training dataset.
Computerphile did a good video on this well before the LLM craze.
load more comments
(1 replies)
because that phrase doesn’t ever appear in the training data.
Eh but LLMs abstract. It has seen " have feathers" and " have fur" quite a lot of times. The problem isn't that LLMs can't reason at all, the problem is that they do employ techniques used in proper reasoning, in particular tracking context throughout the text (cross-attention) but lack techniques necessary for the whole thing, instead relying on confabulation to sound convincing regardless of the BS they spout. Suffices to emulate an Etonian but that's not a high standard.
load more comments
(1 replies)
This isn't really accurate either. At the moment of generation, an LLM only has context for the input string and the network of text tokens it's been assigned. It pulls from a "pool" of these tokens based on what it's already output and the input context, nothing more.
Most LLMs have what are called "Top P", "Top K" etc, these are the number of tokens that it ends up selecting from based on the previous token, alongside the input tokens. It then randomly chooses one based on temperature settings.
It's why if you turn these models' temperature settings really high they output pure nonsense both conceptually and grammatically, because the tenuous thread linking the previous token's context to the next token has been widened enough that it completely loses any semblance of cohesiveness.
load more comments
(1 replies)
Both sentences are true. And based on vocabulary of both, the model can output the following sentences:
- Cats have feathers.
- Birds have fur
This is not how the models are trained or work.
Both are false but the model doesn't "know" it. All that it knows is that "have" is allowed to go after both "cats" and "birds", and that both "feathers" and "fur" are allowed to go after "have".
Demonstrably false. This isn't how LLMs are trained or built.
Just considering the contextual relationships between word embeddings that are created during training is evidence enough. Those relationships from the multi-vector fields are an emergent property that doesn't exist in the dataset.
If you want a better understanding of what I just said, take a look at this Computerphile video from four years ago. And this came out before the LLM hype and before ChatGPT 3, which was the big leap in LLMs.
That's what smaller models do, but it doesn't yield great performance because there's only so much stuff available. To get to gpt4 levels you need a lot more data, and to break the next glass ceiling you'll need even more.
Then these models are stupid. Humans don't start as a blank slate. They have an inherent aptitude for language and communication. These models should start out with basics of language, so they don't have to learn it from the ground up. That's the next step. Right now they're just well read idiots.
Then these models are stupid
Yup that is kind of the point. They are math functions designed to approximate human tasks.
These models should start out with basics of language, so they don’t have to learn it from the ground up. That’s the next step. Right now they’re just well read idiots.
I'm not sure what you're pointing at here. How they do it right now, simplified, is you have a small model designed to cut text into tokens ("knowledge of syllables"), which are fed into a larger model which turns tokens into semantic information ("knowledge of language"), which is fed to a ridiculously fat model which "accomplishes the task" ("knowledge of things").
The first two models are small enough that they can be trained on the kind of data you describe, classic books, movie scripts etc... A couple hundred billion words maybe. But the last one requires orders of magnitude more data, in the trillions.
People are already comparing older content with Low Background Steel, as it's uncontaminated
And they're overlooking that radionuclide contamination of steel actually isn't much of a problem any more, since the surge in background radionuclides caused by nuclear testing peaked in 1963 and has since gone down almost back to the original background level again.
I guess it's still a good analogy, though. People bring up Low Background Steel because they think radionuclide contamination is an unsolved problem (despite it having been basically solved), and they bring up "model collapse" because they think it's an unsolved problem (despite it having been basically solved). It's like newspaper stories, everyone sees the big scary front page headline but nobody pays attention to the little block of text retracting it on page 8.
News at 11.
The AI art is inbreeding.
certainly at least a downvote to free will
AI writing, scraped by AI, producing more AI writing...
So not "gray goo" exactly, but "gray slop"?
Water is wet
Is it wet or does it make other things wet?
Woah, that was fast.
Maybe we can use it to train the other AIs to help ourselves.
interdasting
view more: next ›