TikTok’s parent launched a web scraper that’s gobbling up the world’s online data 25-times faster than OpenAI
TikTok’s parent launched a web scraper that’s gobbling up the world’s online data 25-times faster than OpenAI
The crawler, dubbed Bytespider, is scraping the internet at 3,000 times the rate of other genAI tools like Anthropic.
You're viewing a single thread.
This is fine. I support archiving the Internet.
It kinda drives me crazy how normalized anti-scraping rhetoric is. There is nothing wrong with (rate limited) scraping
The only bots we need to worry about are the ones that POST, not the ones that GET
It’s not fine. They are not archiving the internet.
I had to ban their user agent after very aggressive scraping that would have taken down our servers. Fuck this shitty behaviour.
Isn't there a way to limit requests so that traffic isn't bringing down your servers
I had to block ByteSpider at work because it can't even parse HTML correctly and just hammers the same page and accounts to sometimes 80% of the traffic hitting a customer's site and taking it down.
The big problem with AI scrapers is unlike Google and traditional search engines, they just scrape so aggressively. Even if it's all GETs, they hit years old content that's not cached and use up the majority of the CPU time on the web servers.
Scraping is okay, using up a whole 8 vCPU instance for days to feed AI models is not. They even actively use dozens of IPs to bypass the rate limits too, so theyre basically DDoS'ing whoever they scrape with no fucks given. I've been woken up by the pager way too often due to ByteSpider.
My next step is rewriting all the content with GPT-2 and serving it to bots so their models collapse.
I think a common nginx config is to just redirect malicious bots to some well-cached terrabyte file. I think hetzner hosts one iirc
https://github.com/iamtraction/ZOD
42kB ZIP file which decompresses into 4.5 PB.
wouldn't it be trivial to defend against that with a hash check if the size matches?
though I guess it's possible to create your own that differs
Bytedance ain’t looking to build an archival tool. This is to train gen AI models.
Bullshit. This bot doesn't identify itself as a bot and doesn't rate limit itself to anything that would be an appropriate amount. We were seeing more traffic from this thing that all other crawlers combined.
Not rate limiting is bad. Hate them because of that, not because they're a bot.
Some bots are nice
Even if they were rate limiting they're still just using the bot to train an AI. If it's from a company there's a 99% chance the bot is bad. I'm leaving 1% for whatever the Internet Archive (are they even a company tho?) is doing.
I don't hate all bots, I hate this bot specifically because:
- they intentionally hide that they are a bot to evade our, and everyone else's, methods of restricting which bots we allow and how much activity we allow.
- they do not respect the robots.txt
- the already mentioned lack of rate limiting
this is neither archiving, nor ratelimited, if the AI training purpose and the 25 times faster scraping than a large company did not make it obvious
GET requests can still overload a system.
The type of request is not relevant. It's the cost of the request that's an issue. We have long ago stopped serving html documents that are static and can be cached. Tons of requests can trigger complex searches or computations which are expensive server side. This type of behavior basically ruins the internet and pushes everything into closed gardens and behind logins.
Sounds like you need to fire your sysadmin
It has nothing to do with a sysadmin. It's impossible for a given request to require zero processing power. Therefore there will always be an upper limit to how many get requests can be handled, even if it's a small amount of processing power per request.
For a business it's probably not a big deal, but if it's a self hosted site it quickly can become a problem.
Caches can be configured locally to use near-zero processing power. Or moved to the last mile to use zero processing power (by your hardware)
Near zero isn't zero though. And not everyone is using caching.
Right, thats why I said you should fire your sysadmin if they aren't caching or can't manage to get the cache down to zero load for static content served to simple GET requests
Not every GET request is simple enough to cache, and not everyone is running something big enough to need a sysadmin.