3 ms·

>And worse, then ChatGPT will digest its own excrement, worsening its own results further I wonder if we'll get a "dead sea effect" with AI, I've seen some stu

>And worse, then ChatGPT will digest its own excrement, worsening its own results further I wonder if we'll get a "dead sea effect" with AI, I've seen some stuff saying they've basically run out of high quality training data and now the training pool will get poisoned by AI generated shit. Basically garbage in, garbage out and these large language models might not be able to improve

CuriouslyC 4y ago
We are nowhere near out of data. We're just out of hyper-relevant modern data. There is probably about 100-200T of old books, newspapers, journal articles, magazines and so forth. For reference, GPT3 was trained on 45T.
- naniwaduni 4y ago
  Just order of magnitude seems pretty close to out of data, actually, if beyond that we're looking at the firehose of low-density bulk-generated modern data.
gptgpp 4y ago
Maybe, but there are something like 700,000 books published on average each year, and almost 2 million scientific journal articles. Let's not even consider newspaper. Of course, some of those books will definitely be AI generated or garbage quality, and we all know many of those journal articles can be worth less than the paper they're printed on. Yet even if we cut it down to 100,000 books and half a million scientific papers, that's a lot of training data each year... And that is just considering print media, there are other ways to get more content too. For example, there is also transcription of video/podcasts/tv-shows/movies, etc. along with descriptions of the scenes for video, which could be used to generate a lot more stuff. With people speaking to their devices and using text-to-speech more often, that's another source too--wouldn't be surprised if some devices just start recording conversations, and transcribing them. Seems like a ton of potential data sources to me, although it will certainly get more difficult to cull AI generated stuff to prevent feedback, I'm sure the tooling will evolve to enable easy AI content detection and exclusion.
- hasmanean 4y ago
  Journal articles (and newspapers) are also plagued by bullshit. Humans are capable of producing intelligently sounding word salads just like chatgpt can.
  - gptgpp 4y ago
    Yep that's why I said let's not even consider newspapers. Many of those have been using AI generated/content-mill/sponsored content for years and years. Also why I acknowledged journal articles can be worth less than the paper they're printed on. Even if you were to select for reputable, high impact journals, those also often experience scandals, retractions, potential data fabrication, etc. But then there are textbooks and technical publications, also being published in the hundreds of thousands globally each year. The fact is that with billions of human beings on the planet, and media increasingly being digitized by default, and AI-content detection, I don't see how we could possibly run out of new content to grow LLMs...
ericd 4y ago
Well, we still haven't really tapped video, which is arguably a much higher source of data on lots of things (especially on how things act in the physical world). And curation will likely help a lot. And it's not like you can assume an indiscriminate crawl of the net is all human generated currently, anyway, let alone accurate. There's always cleaning involved.
Beta-7 4y ago
Considering there are "GPT plagiarism" checkers, I don't think this will become an issue. I wonder at which point an extension will come out that will check a page's text if it was written by a human.
- layer8 4y ago
  Those checkers already have a significant failure rate of false positives/negatives, and that will only get worse as LLMs come closer to human output. Note also that a checker can in principle never outwit a state-of-the-art AI, because the AI can just incorporate and therefore preempt the checker logic.
  - 93po 4y ago
    And even with good checking it's still trivial to have GPT do 95% of the work and then make some stylistic edits to get around detection.
forgetfulness 4y ago
Pre-ChatGPT datasets will become prized commodities, this sort of AI will be trapped in a stasis of pre-2023 pop culture as subsequent AIs will need to use datasets that hadn't yet been contaminated by pervasive ChatGPT spam. It will be like low-background steel, steel that has somehow been isolated from atomic fallout form the mid XX century onwards, and must be used for radiation-sensitive equipment: https://en.wikipedia.org/wiki/Low-background_steel https://en.wikipedia.org/wiki/Low-background_steel Except somehow worse, because it's just steel, this is culture.
- sorokod 4y ago
  Sounds very plausible, along with intensive research on how to decontaminate latter samples.

CuriouslyC 4y ago

We are nowhere near out of data. We're just out of hyper-relevant modern data. There is probably about 100-200T of old books, newspapers, journal articles, magazines and so forth. For reference, GPT3 was trained on 45T.

naniwaduni 4y ago
Just order of magnitude seems pretty close to out of data, actually, if beyond that we're looking at the firehose of low-density bulk-generated modern data.

gptgpp 4y ago

Maybe, but there are something like 700,000 books published on average each year, and almost 2 million scientific journal articles. Let's not even consider newspaper. Of course, some of those books will definitely be AI generated or garbage quality, and we all know many of those journal articles can be worth less than the paper they're printed on. Yet even if we cut it down to 100,000 books and half a million scientific papers, that's a lot of training data each year... And that is just considering print media, there are other ways to get more content too. For example, there is also transcription of video/podcasts/tv-shows/movies, etc. along with descriptions of the scenes for video, which could be used to generate a lot more stuff. With people speaking to their devices and using text-to-speech more often, that's another source too--wouldn't be surprised if some devices just start recording conversations, and transcribing them. Seems like a ton of potential data sources to me, although it will certainly get more difficult to cull AI generated stuff to prevent feedback, I'm sure the tooling will evolve to enable easy AI content detection and exclusion.

hasmanean 4y ago
Journal articles (and newspapers) are also plagued by bullshit. Humans are capable of producing intelligently sounding word salads just like chatgpt can.
- gptgpp 4y ago
  Yep that's why I said let's not even consider newspapers. Many of those have been using AI generated/content-mill/sponsored content for years and years. Also why I acknowledged journal articles can be worth less than the paper they're printed on. Even if you were to select for reputable, high impact journals, those also often experience scandals, retractions, potential data fabrication, etc. But then there are textbooks and technical publications, also being published in the hundreds of thousands globally each year. The fact is that with billions of human beings on the planet, and media increasingly being digitized by default, and AI-content detection, I don't see how we could possibly run out of new content to grow LLMs...

ericd 4y ago

Well, we still haven't really tapped video, which is arguably a much higher source of data on lots of things (especially on how things act in the physical world). And curation will likely help a lot. And it's not like you can assume an indiscriminate crawl of the net is all human generated currently, anyway, let alone accurate. There's always cleaning involved.

Beta-7 4y ago

Considering there are "GPT plagiarism" checkers, I don't think this will become an issue. I wonder at which point an extension will come out that will check a page's text if it was written by a human.

layer8 4y ago
Those checkers already have a significant failure rate of false positives/negatives, and that will only get worse as LLMs come closer to human output. Note also that a checker can in principle never outwit a state-of-the-art AI, because the AI can just incorporate and therefore preempt the checker logic.
- 93po 4y ago
  And even with good checking it's still trivial to have GPT do 95% of the work and then make some stylistic edits to get around detection.

forgetfulness 4y ago

Pre-ChatGPT datasets will become prized commodities, this sort of AI will be trapped in a stasis of pre-2023 pop culture as subsequent AIs will need to use datasets that hadn't yet been contaminated by pervasive ChatGPT spam. It will be like low-background steel, steel that has somehow been isolated from atomic fallout form the mid XX century onwards, and must be used for radiation-sensitive equipment: https://en.wikipedia.org/wiki/Low-background_steel https://en.wikipedia.org/wiki/Low-background_steel Except somehow worse, because it's just steel, this is culture.

sorokod 4y ago
Sounds very plausible, along with intensive research on how to decontaminate latter samples.