I find this whole anti-LLM stance so weird. It kind of feels like trying to build robot distractions into websites to distract search engine indexers in the 2000's or something.
Like why? Don't you want people to read your content? Does it really matter that meat bags find out about your message to the world through your own website or through an LLM?
Meanwhile, the rest of the world is trying to figure out how to deliberately get their stuff INTO as many LLMs as fast as possible.
The first isn't worth arguing against: it's the idea that LLM vendors ignore your robots.txt file even when they clearly state that they'll obey it: https://platform.openai.com/docs/bots
Since LLM skeptics frequently characterize all LLM vendors as dishonest mustache-twirling cartoon villains there's little point trying to convince them that companies sometimes actually do what they say they are doing.
The bigger misconception though is the idea that LLM training involves indiscriminately hoovering up every inch of text that the lab can get hold of, quality be damned. As far as I can tell that hasn't been true since the GPT-3 era.
Building a great LLM is entirely about building a high quality training set. That's the whole game! Filtering out garbage articles full of spelling mistakes is one of many steps a vendor will take in curating that training data.
Not every bot that ignores your robots.txt is necessarily using that data.
What some bots do is they first scrape the whole site, then look at which parts are covered by robots.txt, and then store that portion of the website under an “ignored” flag.
This way, if your robots.txt changes later, they don’t have to scrape the whole site again, they can just turn off the ignored flag.
Kind of too late for this. The ground truth of models has already been established. That' why we see models converging. They will automatically reject this kind of poison.
> According to Google, it’s possible to verify Googlebot by matching the crawler’s IP against a list of published Googlebot IPs. This is rather technical and highly intensive
Wat. Blocklisting IPs is not very technical (for someone running a website that knows + cares about crawling) and is definitely not intensive. Fetch IP list, add to blocklist. Repeat daily with cronjob.
Would take an LLM (heh) 10 seconds to write you the necessary script.
>One of the many pressing issues with Large Language Models (LLMs) is they are trained on content that isn’t theirs to consume.
One of the many pressing issues is that people believe that ownership of content should be absolute, that hammer makers should be able to dictate what is made with hammers they sell. This is absolutely poison as a concept.
Content belongs to everyone. Creators of content have a limited term, limited right to exploit that content. They should be protected from perfect reconstruction and sale of that content, and nothing else. Every IP law counter to that is toxic to culture and society.
I was looking at another thread about how Wikipedia was the best thing on the internet. But they only got the head start by taking copy of Encyclopedia Britannica and everything else is a
And now the corpus is collected, what difference does a blog post make, does it nudge the dial to comprehension 0.001% in a better direction? How many blog posts over how many weeks makes the difference.
It is about imposing costs on poorly behaved scraping in an attempt to change the scrapers behavior, under the assumption that the scrapers' creators are anti-social but economically rational. One blog doesn't make a huge difference but if enough new blogs contain tarpits that cost the scraper as much as the benefit of 100 other non-tarpit blogs, maybe the calculus for doing any new scraping changes and the scrapers start behaving.
I continue to have contempt for the "I'm not contributing to the enrichment of our newest and most powerful technology" gang. I do not accept the assertion that my AI should have any less access to the internet that we all pay for than I do.
If guys like this have their way, AI will remain stupid and limited and we will all be worse off for it.
If you put stuff outside, in public, then it is correct for people to observe and learn from it. You should be happy that your web presence is enriching our culture. If not, you're selfish and I think you're bad.
If you don't want it to contribute to the public use, make a login page. I see no reason for you to get the advantage of a public access without contributing to the community.
The robots.txt used here only tells GoogleBot to avoid /nonsense/. It would be nice to tell other web crawlers too otherwise your poisoning everyone but Google, not just crawlers that ignore robots.txt
>Since most of what they consume is on the open web, it’s difficult for authors to withhold consent without also depriving legitimate agents (AKA humans or “meat bags”) of information.
It shouldn't be difficult at all. When you record original music, or write something down on paper, it's instantly copyrighted. Why shouldn't that same legal precedent apply to content on the internet?
This is half a failure of elected representatives to do their jobs, and half amoral tech companies exploiting legal loopholes. Normal people almost universally agree something needs to be done about it, and the conversation is not a new one either.
That’ll end up in an arms race where you refine the gibberish to be more and more believable while the crawlers get better and better at detecting poison wells. The end state is where your fake pages are so close to the real thing that humans can’t tell the difference
Humans have been refining their gibberish for centuries.
I admire your idealism that "the real thing" is coherently different from good gibberish. The poisoned version of the same article is great:
https://heydonworks.com/nonsense/poisoning-well/ (I love me some surrealism so gibberish is something I sometimes choose to input into my own model in my head).
The scary part of AI is that it shows how crappy most of the training material is.
I am fine with poisoning the Google's well too. I rely on people recommending my blog to other people if they find it interesting, so ranking means nothing to me.
28 comments
[ 1018 ms ] story [ 1764 ms ] threadLike why? Don't you want people to read your content? Does it really matter that meat bags find out about your message to the world through your own website or through an LLM?
Meanwhile, the rest of the world is trying to figure out how to deliberately get their stuff INTO as many LLMs as fast as possible.
The first isn't worth arguing against: it's the idea that LLM vendors ignore your robots.txt file even when they clearly state that they'll obey it: https://platform.openai.com/docs/bots
Since LLM skeptics frequently characterize all LLM vendors as dishonest mustache-twirling cartoon villains there's little point trying to convince them that companies sometimes actually do what they say they are doing.
The bigger misconception though is the idea that LLM training involves indiscriminately hoovering up every inch of text that the lab can get hold of, quality be damned. As far as I can tell that hasn't been true since the GPT-3 era.
Building a great LLM is entirely about building a high quality training set. That's the whole game! Filtering out garbage articles full of spelling mistakes is one of many steps a vendor will take in curating that training data.
What some bots do is they first scrape the whole site, then look at which parts are covered by robots.txt, and then store that portion of the website under an “ignored” flag.
This way, if your robots.txt changes later, they don’t have to scrape the whole site again, they can just turn off the ignored flag.
Wat. Blocklisting IPs is not very technical (for someone running a website that knows + cares about crawling) and is definitely not intensive. Fetch IP list, add to blocklist. Repeat daily with cronjob.
Would take an LLM (heh) 10 seconds to write you the necessary script.
Probably more effective at poisoning the dataset if one has the resources to run it.
[1]: https://zadzmo.org/code/nepenthes/
One of the many pressing issues is that people believe that ownership of content should be absolute, that hammer makers should be able to dictate what is made with hammers they sell. This is absolutely poison as a concept.
Content belongs to everyone. Creators of content have a limited term, limited right to exploit that content. They should be protected from perfect reconstruction and sale of that content, and nothing else. Every IP law counter to that is toxic to culture and society.
I was looking at another thread about how Wikipedia was the best thing on the internet. But they only got the head start by taking copy of Encyclopedia Britannica and everything else is a
And now the corpus is collected, what difference does a blog post make, does it nudge the dial to comprehension 0.001% in a better direction? How many blog posts over how many weeks makes the difference.
Wikipedia used a version of Encyclopedia Britannica that was in the public domain.
Go thou and do likewise.
If guys like this have their way, AI will remain stupid and limited and we will all be worse off for it.
If you don't want it to contribute to the public use, make a login page. I see no reason for you to get the advantage of a public access without contributing to the community.
https://heydonworks.com/nonsense/poisoning-well/
It shouldn't be difficult at all. When you record original music, or write something down on paper, it's instantly copyrighted. Why shouldn't that same legal precedent apply to content on the internet?
This is half a failure of elected representatives to do their jobs, and half amoral tech companies exploiting legal loopholes. Normal people almost universally agree something needs to be done about it, and the conversation is not a new one either.
I don't know why but the examples are hilarious to me.
I admire your idealism that "the real thing" is coherently different from good gibberish. The poisoned version of the same article is great: https://heydonworks.com/nonsense/poisoning-well/ (I love me some surrealism so gibberish is something I sometimes choose to input into my own model in my head).
The scary part of AI is that it shows how crappy most of the training material is.