3 ms·

My point is that if crawlers have to worry about poison that may make them start to respect robots.txt or something. It's a bit like a "Beware of Dog" sign.

by jordanb 6mo ago

lxgr 6mo ago
How would that become a strong, stable signal, if both highly valuable and highly slopified content will use robots.txt?
- jordanb 6mo ago
  For clarification poisoning and slop are different concepts. Slop is the output of AI. Poisoning is making your content (that may otherwise be good content) fuck up in the internals of an LLM. Classic example is the nightshade attack on image generators. One could imagine an open source project that doesn't want to be ingested by an LLM. They could try to put that in the license but of course the license won't be obeyed. Alternately, if they could alter the code such that the OSS project itself remains high quality, but if you try to train a coding LLM on it the LLM will output code full of SQL injection exploits (for instance) or maybe just bogus uncompilable stuff, then the LLM authors will suddenly have a reason to start respecting your license and excluding the code from their index.
  - lxgr 6mo ago
    To me, slop is anything that makes sense at the surface level but falls apart upon closer examination, since there's "nobody home, semantically". So why would this not be poisoning with (in this case human-generated) slop?
    - dspillett 6mo ago
      I may be picking nits on nits here, but… If slop indicates content with no reasoning, then deliberate slop isn't slop as it is generated with both reason and purpose known to, and understood by, the creator. Though if someone deliberately uses a generative model to create slop that line of reasoning might eat itself…
      - lxgr 6mo ago
        Maybe we should extend "death of the author" to non-human author entities as well :)
dspillett 6mo ago
Unfortunately the use of the sign often highlights what the scrapers want most, so if they pay attention to it (rather than just completely ignoring it as most do now) it will be to specifically follow where told not to. The scrapers ideally want content that is original. Often content that is also new is more highly prized, but not as much as you might think⁰. This will only become more of a driver as the amount of LLM generated content that is out there to be mixed in increases, in order to limit the Habsburg problem they won't want too much regurgitated content in the training data. Bad content from before LLM scraping became a resource problem¹ is highly unlikely to be marked in robots.txt, the same for content newly generated-by-an-LLM. People attempting to fend off scrapers and other bots with robots.txt entries are likely protecting the sort of content the scrapers actively want - original output that they've put some time into or code in a repo they don't want scraped (as scraping a repo is incredibly inefficient and resource heavy from the PoV of the repo owner). I strongly suspect that the amount of desirable content behind robots.txt “blocks” is far too valuable to ignore despite the amount of poison content traps, or just things otherwise not worth the time scouring through, that might also be there. A “beware of the dog” sign is of no protection when the reader actively wants to see the doggies! -------- [0] if scraping for training an LLM you don't want just new content, but you would prefer as much of your input data as possible to be as few steps as possible from original [1] and a copying concern, though I'll avoid that discussion as it can get quite thorny and whichever side or fence you are on in that matter the resource consumption is objectively a problem all the same.

My point is that if crawlers have to worry about poison that may make them start to respect robots.txt or something. It's a bit like a "Beware of Dog" sign.

by jordanb 6mo ago

My point is that if crawlers have to worry about poison that may make them start to respect robots.txt or something. It's a bit like a "Beware of Dog" sign.

lxgr 6mo ago

How would that become a strong, stable signal, if both highly valuable and highly slopified content will use robots.txt?

jordanb 6mo ago
For clarification poisoning and slop are different concepts. Slop is the output of AI. Poisoning is making your content (that may otherwise be good content) fuck up in the internals of an LLM. Classic example is the nightshade attack on image generators. One could imagine an open source project that doesn't want to be ingested by an LLM. They could try to put that in the license but of course the license won't be obeyed. Alternately, if they could alter the code such that the OSS project itself remains high quality, but if you try to train a coding LLM on it the LLM will output code full of SQL injection exploits (for instance) or maybe just bogus uncompilable stuff, then the LLM authors will suddenly have a reason to start respecting your license and excluding the code from their index.
- lxgr 6mo ago
  To me, slop is anything that makes sense at the surface level but falls apart upon closer examination, since there's "nobody home, semantically". So why would this not be poisoning with (in this case human-generated) slop?
  - dspillett 6mo ago
    I may be picking nits on nits here, but… If slop indicates content with no reasoning, then deliberate slop isn't slop as it is generated with both reason and purpose known to, and understood by, the creator. Though if someone deliberately uses a generative model to create slop that line of reasoning might eat itself…
    - lxgr 6mo ago
      Maybe we should extend "death of the author" to non-human author entities as well :)

dspillett 6mo ago

Unfortunately the use of the sign often highlights what the scrapers want most, so if they pay attention to it (rather than just completely ignoring it as most do now) it will be to specifically follow where told not to. The scrapers ideally want content that is original. Often content that is also new is more highly prized, but not as much as you might think⁰. This will only become more of a driver as the amount of LLM generated content that is out there to be mixed in increases, in order to limit the Habsburg problem they won't want too much regurgitated content in the training data. Bad content from before LLM scraping became a resource problem¹ is highly unlikely to be marked in robots.txt, the same for content newly generated-by-an-LLM. People attempting to fend off scrapers and other bots with robots.txt entries are likely protecting the sort of content the scrapers actively want - original output that they've put some time into or code in a repo they don't want scraped (as scraping a repo is incredibly inefficient and resource heavy from the PoV of the repo owner). I strongly suspect that the amount of desirable content behind robots.txt “blocks” is far too valuable to ignore despite the amount of poison content traps, or just things otherwise not worth the time scouring through, that might also be there. A “beware of the dog” sign is of no protection when the reader actively wants to see the doggies! -------- [0] if scraping for training an LLM you don't want just new content, but you would prefer as much of your input data as possible to be as few steps as possible from original [1] and a copying concern, though I'll avoid that discussion as it can get quite thorny and whichever side or fence you are on in that matter the resource consumption is objectively a problem all the same.