I have no idea why the makers of LLM crawlers think it’s a good idea to ignore bot rules. The rules are there for a reason and the reasons are often more complex than “well, we just don’t want you to do that”. They’re usually more like “why would you even do that?”
Ultimately you have to trust what the site owners say. The reason why, say, your favourite search engine returns the relevant Wikipedia pages and not bazillion random old page revisions from ages ago is that Wikipedia said “please crawl the most recent versions using canonical page names, and do not follow the links to the technical pages (including history)”. Again: Why would anyone index those?
Because it takes work to obey the rules, and you get less data for it. The theoretical competitor could get more ignoring those and get some vague advantage for it.
I’d not be surprised if the crawlers they used were bare-basic utilities set up to just grab everything without worrying about rules and the like.
This is getting ridiculous. Can someone please ban AI? Or at least regulate it somehow?
Imagine how much power is wasted on this unfortunate necessity.
Now imagine how much power will be wasted circumventing it.
Fucking clown world we live in
On on hand, yes. On the other…imagine frustration of management of companies making and selling AI services. This is such a sweet thing to imagine.
I just want to keep using uncensored AI that answers my questions. Why is this a good thing?
Because it only harms bots that ignore the “no crawl” directive, so your AI remains uncensored.
Because it’s not AI, it’s LLMs, and all LLMs do is guess what word most likely comes next in a sentence. That’s why they are terrible at answering questions and do things like suggest adding glue to the cheese on your pizza because somewhere in the training data some idiot said that.
The training data for LLMs come from the internet, and the internet is full of idiots.
From the article it seems like they don’t generate a new labyrinth for every single time: Rather than creating this content on-demand (which could impact performance), we implemented a pre-generation pipeline that sanitizes the content to prevent any XSS vulnerabilities, and stores it in R2 for faster retrieval."
Surprised at the level of negativity here. Having had my sites repeatedly DDOSed offline by Claudebot and others scraping the same damned thing over and over again, thousands of times a second, I welcome any measures to help.
thousands of times a second
Modify your Nginx (or whatever web server you use) config to rate limit requests to dynamic pages, and cache them. For Nginx, you’d use either fastcgi_cache or proxy_cache depending on how the site is configured. Even if the pages change a lot, a cache with a short TTL (say 1 minute) can still help reduce load quite a bit while not letting them get too outdated.
Static content (and cached content) shouldn’t cause issues even if requested thousands of times per second. Following best practices like pre-compressing content using gzip, Brotli, and zstd helps a lot, too :)
Of course, this advice is just for “unintentional” DDoS attacks, not intentionally malicious ones. Those are often much larger and need different protection - often some protection on the network or load balancer before it even hits the server.
Already done, along with a bunch of other stuff including cloudflare WAF and rate limiting rules.
I am still annoyed that it took me over a day’ of my life to finally (so far) restrict these things. And several other days to offload the problem to Cloudflare pages for sites that I previous self hosted but my rural link couldn’t support.
this advice is just for “unintentional” DDoS attacks, not intentionally malicious ones.
And I don’t think these high volume AI scrapes are unintentional DDOS attacks. I consider them entirely intentional. Not deliberrately malicious, but negligent to the point of criminality. (Especially in requesting the same pages again so frequently, and all of them ignoring robots.txt)
I guess this is what the first iteration of the Blackwall looks like.
