As junk web pages written by AI proliferate, the models that rely on that data will suffer.
Good.
“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"
Not so hard to imagine given that these people have always seen technical systems as magic.
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.
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:
Both sentences are true. And based on vocabulary of both, the model can output the following sentences:
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.
1 - a markov chain only takes previous tokens as input.
2 - It uses a function (in the mathematical sense, so same input results in same output, completely stateless) to generate a set of probabilities for what the next token might be.
3 - The most probable token is picked, else randomness (temperature) is inserted here to choose a different token occasionally.
an llm's internals, the part that's trained is literally the function used in step 2. You could have this function implemented a number of ways, ex you could build a huge table and consult it. Or you could generate it somehow. You could train a big neural network that takes previous tokens as input, and outputs probabilities of tokens as output. You then enumerate its outputs for every possible permutation of inputs and there's your table. This would take too much time and space, so we just run the function on-demand instead. Exact same result. It can be very smart and notice correlations, but ultimately it generates a (virtual) huge static table. This is a completely deterministic process. A trained NN is still a (huge) mathematical function. So the big network that they spend resources training is basically the function used in step 2.
Step 3 is the cause of hallucinations. It's the only nondeterministic part. And it's not part of the llm itself in any way. No matter how smarter the neural network gets, the hallucinations are introduced mainly in step 3. So no, they won't be solving the LLM hallucination problem anytime soon.
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.
Workarounds for those sorts of limitations have been developed, though. Chain-of-thought prompting has been around for a while now, and I recall recently seeing an article about a model that had that built right into it; it had been trained to use tags to enclose invisible chunks of its output that would be hidden from the end user but would be used by the AI to work its way through a problem. So if you asked it whether cats had feathers it might respond "Feathers only grow on birds and dinosaurs. Cats are mammals. No, cats don't have feathers." And you'd only see the latter bit. It was a pretty neat approach to improving LLM reasoning.
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.
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.
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
Water is wet
Is it wet or does it make other things wet?
AI writing, scraped by AI, producing more AI writing...
So not "gray goo" exactly, but "gray slop"?
Woah, that was fast.
Maybe we can use it to train the other AIs to help ourselves.
interdasting
This is a most excellent place for technology news and articles.
