That no one has actually solved the underlying problem at all, and the generation of the example LLM has no bearing on the nature of the fundamental problem.
You are totally misunderstanding my argument then. As I said, garbage in garbage out. Your article is just an example of that. It’s pretty obvious that if you train an LLM on bad data, you will get bad output.
What I’m saying is that the AI labs are handling this not by fixing the “garbage out” part, but by minimizing the “garbage in” part.
The fact that all you could come up with was research (not an actual example of poisoning a real training set) from 2025 kind of proves that this isn’t some kind of widespread, unsolvable problem like you seem to be claiming.
I literally just grabbed a random link. I’ve seen dozens of real life examples of poisoning.
The poisoning issue makes it so that no one can use the internet for training anymore, because more and more internet content is poisoned as a side effect - or poisoned intentionally. And .001% of poisoned data is enough to screw things up if included in the training data.
It’s also one reason why Google search results have been getting so much worse - it’s hard to not find a SEO page with subtly (or not so subtly) wrong AI slop on almost every topic you can imagine. Most folks won’t recognize it, but that’s what is going on if you know what to look for.
One other way of putting it is the ouroborus problem - more and more internet content is AI generated, because of people trying to game the system, and they are making it is indistinguishable from real content as possible to get by the AI detection algorithms.
Anyone trying to train on it just ends up eating the shit from another LLM, which poisons it.
Another name for it is ‘model collapse’, which also doesn’t have a known solution yet.
You’ve seen actual model poisoning? Or have you seen a model return the wrong answer due to what it saw in a search result? Or were they hallucinations perhaps? How do you know it’s due to poisoned training data?
And do you even realize how much data 0.001% of the training data for a frontier models is? They’re trained on 10s of trillions of tokens, meaning you’d need hundreds of millions of tokens of poisoned data.
Some of these problems you mention could become real barriers to models improvements, though there are plenty of countermeasures, such as by focusing on high quality data sources like I mentioned before.
We’ve already probably gotten as much as we’re ever going to get from simply scraping more and more unstructured text from the web as a way to improve model performance.
The type of training being done now is around tool use and solving specific types of problems better, which is the type of training data you simply don’t find lying around on the web.
You’re expecting me to know your job? Give me a break.
I’m wondering the same thing. You keep talking of some grand poisoning problem but can’t point to any specific public information except an article saying that it’s possible. As if that was ever in doubt.
