Machine translators have made it easier than ever to create error-plagued Wikipedia articles in obscure languages. What happens when AI models get trained on junk pages?
Spreading knowledge via machine translation where there are no human translators available, had to be better than not translating
Have you not read my entire comment…?
One of the Greenlandic Wiki articles “claimed Canada had only 41 inhabitants”. What use is a text like that? In what world is learning that Canada has 41 inhabitants better than going to the English version of the article and translating it yourself?
Perhaps part of the solution is machine readable citations
The contents of the citations are already used for training, as long as they’re publicly available. That’s not the problem. The problem is that LLMs do not understand context well, they are not, well, intelligent.
The “Chinese Room” thought experiment explains it best, I think: imagine you’re in a room with writing utensils and a manual. Every now and again a letter falls in to the room through a slit in the wall. Your task is to take the letter and use the manual to write a response. If you see such and such shape, you’re supposed to write this and that shape on the reply paper, etc. Once you’re done, you throw the letter out through the slit. This goes back and forth.
To the person on the other side of the wall it seems like they’re having a conversation with someone fluent in Chinese whereas you’re just painting shapes based on what the manual tells you.
LLMs don’t understand the prompts - they generate responses based on the probability of certain characters or words or sentences being next to each other when the prompt contains certain characters, words, and sentences. That’s all there is.
There was a famous botched experiment where scientists where training an AI model to detect tumours. It got really accurate on the training data so they tested it on new cases gathered more recently. It gave a 100% certainty of a tumour being present if the photograph analysed had a yellow ruler on it, because most photos of tumours in the training data had that ruler for scale.
But even then you have huge gaps on one side with untrustworthy humans (like comedy) and on the other side with machine generated facts such as from a database
“Machine generated facts” are not facts, they’re just hallucinations and falsehoods. It is 100% better to NOT have them at all and have to resort to the English wiki, than have them and learn bullshit.
Especially because, again, the contents of the Wikipedia are absolutely being used for training further LLM models. The more errors there are, the worse the models become eventually leading to a collapse of truth. We are already seeing this with whole “research” publications being generated, including “source” material invented on the spot, proving bogus results.
Have you not read my entire comment…?
One of the Greenlandic Wiki articles “claimed Canada had only 41 inhabitants”. What use is a text like that? In what world is learning that Canada has 41 inhabitants better than going to the English version of the article and translating it yourself?
The contents of the citations are already used for training, as long as they’re publicly available. That’s not the problem. The problem is that LLMs do not understand context well, they are not, well, intelligent.
The “Chinese Room” thought experiment explains it best, I think: imagine you’re in a room with writing utensils and a manual. Every now and again a letter falls in to the room through a slit in the wall. Your task is to take the letter and use the manual to write a response. If you see such and such shape, you’re supposed to write this and that shape on the reply paper, etc. Once you’re done, you throw the letter out through the slit. This goes back and forth.
To the person on the other side of the wall it seems like they’re having a conversation with someone fluent in Chinese whereas you’re just painting shapes based on what the manual tells you.
LLMs don’t understand the prompts - they generate responses based on the probability of certain characters or words or sentences being next to each other when the prompt contains certain characters, words, and sentences. That’s all there is.
There was a famous botched experiment where scientists where training an AI model to detect tumours. It got really accurate on the training data so they tested it on new cases gathered more recently. It gave a 100% certainty of a tumour being present if the photograph analysed had a yellow ruler on it, because most photos of tumours in the training data had that ruler for scale.
“Machine generated facts” are not facts, they’re just hallucinations and falsehoods. It is 100% better to NOT have them at all and have to resort to the English wiki, than have them and learn bullshit.
Especially because, again, the contents of the Wikipedia are absolutely being used for training further LLM models. The more errors there are, the worse the models become eventually leading to a collapse of truth. We are already seeing this with whole “research” publications being generated, including “source” material invented on the spot, proving bogus results.