3 ms·

A big part of AI is weaponising our biases, it's not the first time, won't be the last. Feed it shit and it'll give you shit, too bad the training set (interne

A big part of AI is weaponising our biases, it's not the first time, won't be the last. Feed it shit and it'll give you shit, too bad the training set (internet) is already and open air dump. We'll be in a very bad downward spiral really soon when AI trains on AI generated content

ostenning 4y ago
I don't really know how AI works. Is it possible for AI to train itself on its own generated content? Wouldn't this be like some kind of lossy recursive reproduction, like an audio tape loop continually degrading itself?
- askiiart 4y ago
  Yes, it can, and it could be degrading on itself. Let's use an example of AI art. A lot of AI art is pretty terrible. But if it was trained on data containing AI art that was 1. the description of an image 2. and the image itself, and 3. the image turned out how exactly how the user wanted, then this could, in some cases, actually produce decent training data, particularly with weird, super specific data. But again, it would have to be exactly what was wanted from the description (which could be the original prompt), otherwise it would be bad training data. Really it's just a matter of how well the photo matches its description. Disclaimer: I don't know that much about AI, take this with a big grain of salt. Maybe some other people can chime in here.
- snowpid 4y ago
  If the data is sparse, some researches used AI generated content to train another network. E.G. https://arxiv.org/pdf/1802.06739.pdf https://arxiv.org/pdf/1802.06739.pdf In some cases like here it also has privacy issues.
absqueued 4y ago
AI of our era so far is working solely based on a trained model and haven't evolved beyond the "data" behind training. All the material fed to train the AI machine leads to this mess. I heard Sam (OpenAI founder) talk about current state of AI, we are far from AI making its own concious decisions.

ostenning 4y ago

I don't really know how AI works. Is it possible for AI to train itself on its own generated content? Wouldn't this be like some kind of lossy recursive reproduction, like an audio tape loop continually degrading itself?

askiiart 4y ago
Yes, it can, and it could be degrading on itself. Let's use an example of AI art. A lot of AI art is pretty terrible. But if it was trained on data containing AI art that was 1. the description of an image 2. and the image itself, and 3. the image turned out how exactly how the user wanted, then this could, in some cases, actually produce decent training data, particularly with weird, super specific data. But again, it would have to be exactly what was wanted from the description (which could be the original prompt), otherwise it would be bad training data. Really it's just a matter of how well the photo matches its description. Disclaimer: I don't know that much about AI, take this with a big grain of salt. Maybe some other people can chime in here.
snowpid 4y ago
If the data is sparse, some researches used AI generated content to train another network. E.G. https://arxiv.org/pdf/1802.06739.pdf https://arxiv.org/pdf/1802.06739.pdf In some cases like here it also has privacy issues.

absqueued 4y ago

AI of our era so far is working solely based on a trained model and haven't evolved beyond the "data" behind training. All the material fed to train the AI machine leads to this mess. I heard Sam (OpenAI founder) talk about current state of AI, we are far from AI making its own concious decisions.