Hacker News Reader

Hacker News

Comment by jollofricepeas | Hacker News Reader

Parent Full thread

jollofricepeas

I find it funny that…

Literally no one is talking about how all of the LLMs are training on Reddit data via the API.

So far…

- PaLM

- GPT3

- Falcon

Yes, Reddit is closing on its IPO because it’s not a non-profit.

But that doesn’t mean they should be forced to provide a low cost service to OpenAI, Microsoft, Google and the entire investor community backing LLM startups.

Source:

https://huggingface.co/blog/falcon

View on HN

Perhaps because killing third-party apps is completely unrelated to LLMs? While they’re both related to the API lockdown, the core problem is not LLM-related. It’s Reddit’s own short-sighted incompetence.

They can EASILY provide high prices to LLMs and low prices to app devs. They just won’t. It’s in their best interests to remove third parties and force everyone on the inferior, ad-ridden official app. They could even make premium-based account tokens to allow them to use third party apps, but that’s too long-term oriented. They have to pump’n’dump NOW.

While it seems that some people feel reddit is "owed" something when its data is used by machine learning companies, the reality is that reddit themselves claims that posts are owned by users and so really they are not in any position to be owed anything.

They can fully dictate the terms of their API sure, but unless they want to drop off all search indexes there is nothing they can do about people scraping the data the old fashioned way.

I disagree entirely.

There’s a real non-negligible cost to Reddit hosting the content.

The bottom line is there’s solutions here.

- High volume API users pay up and subsidize the cost of everyone else

- All Reddit users pay a monthly 9.99 subscription and the API stays the same.

- A not-for-profit let’s say Internet Archive takes ownership and begs the Reddit community for donations (ie. Wikipedia)

If it was that expensive to host the API, they wouldn't have done it for over a decade. Reddit's popularity was at least partially built on the labor of 3rd party devs building mobile apps, moderation tools, useful bots, all types of stuff.

The bottom line is they'll happily pull the rug from under those 3rd party devs if it means they can pump their numbers for their IPO.

Multiple LLMs scraping all of reddit is a far different deal than how the APIs were used in the past.

Small cost x moderate usage = moderate cost Small cost x overthetop usage = overthetop cost

What they should do is really hammer the LLMs.

They could just block them if they want. Can't be that hard.

I think this is more about the IPO and inflating figures.

- Everyone will just run web scrapers increasing the load on reddit's servers even more

An API actually REDUCES load and lets you manage it.

Your solution isn’t a solution.

Please answer the question:

How is Reddit supposed to make money with increased load and demand for its content?

The framing in the question is wrong. Reddit gets ~80 comments per second on average. An API to get comments with ids starting after n would let you reasonably scrape all new comments with 1 request per second (or one every 10 seconds if the API returned up to 1000 results, which would be entirely reasonable). They get ~10 posts per second, so add another 1 request for that. The load is trivial. Storing the content is also trivial; the entire history of reddit fits very comfortably on a single consumer-tier SSD.

The infrastructure they provide is easily replicable at almost no cost (really it's just the bandwidth that costs any money at all). The community curation and moderation is done by volunteers. The content is all from the users. Reddit Inc is providing almost none of the value, and is just benefiting from network effects. People go there because people go there.

Reddit hosts images and videos

Only recently. For most of its history of only stored text.

It was 7 years ago.

Yes, and those are features most Reddit users could care less about and or can be linked easy enough.

The real meat and potatoes is plain old text with Reddit style markup.

Your question isn't the question you should be asking.

Please answer the question:

How is reddit supposed to prevent scraping when it is legal to do so, and they have incentive to appear in search engines? Considering that scraping is legal and WILL happen, why not opt to reduce the load by offering an API?

If consuming their content by scraping was easy, nobody would be complaining about the APIs being taken away... it's very easy to make web scraping impractical if they want. Don't be deluded.

The context of the thread is machine learning companies. You're delusional if you think they don't have a ton of web scrapers running 24/7 already.

This content is text, it is extremely cheap to host. The site itself is not particularly resilient, so there is not a lot of overhead there. 10 euros a month is an absolutely ridiculous price for what probably doesn't even cost them 20 cents per user, per month.

Isn't this mostly closing the barn door after the horse left? The AI companies already have their data. It may grow stale, but the bulk of the corpus has already been ingested. Either that or they do old-fashioned scraping.

Depends on the model but that HuggingFace blog post on Falcon was just released in the last couple of days where they mention which models are using Reddit.

Reddit content is actively being pillaged by LLMs.

The web is open, the site is available to anyone with a computer and internet connection. LLMs aren't pillaging anything, they're using it as intended. By this definition Google has pillaged the internet far longer and made a lot more money.

It's not "Reddit content" it's content posted on Reddit. It's an important distinction as hosting the content is the reason the business exists at all.

Reddit is able to license its content in a way that prevents companies from using it while leaving the API free.

In fact, making the API paid does nothing for that goal. Scrapping has been long established as legal so if their license doesn’t change, nothing about LLMs using their content will either

> Reddit is able to license its content in a way that prevents companies from using it while leaving the API free.

And there's the scam laid bare.

It's *NOT* Reddit's content. Sure, by posting you're granting them a whole slew of rights. But you dont transfer ownership to them.

And when your content creators leave, blammo, you have no more new content. And I'm already seeing some subreddits drop dramatically in terms of comments and posts. Tech ones, naturally. Turns out Lemmy is actually pretty cool for a fediverse version of reddit.

We're already seeing the start of death of Imgur as "reddit image host", oh wait "imgur social network", oh wait "no porn Imgur". We all know how this works out. Same game, just a bit ahead of reddit.

It's quite key to the legal strategy of all the LLMs that they do not need a license for their input data. So if they want reddit data they'll just scrape it (which is also legally protected). And if they do need a license then reddit does not have one to sell to the AI companies, at least for most of the text on their site. I don't believe it's a reasonable strategy for any platform (except maybe marketplaces like shutterstock) to attempt to charge money for data for machine learning.

I find it funny also. This fact seemed to have its moment in the sun for about 15 minutes a few months ago when everyone was first talking about ChatGPT and then everyone forgot?

Reddit data? It's entirely user generated data, merely hosted by reddit, regardless of what the terms and conditions say.

This isn’t how the world works. We live in a class based society. Reddit could just have LLMs pay more and focus on moderating correct API payment usage.

Any one can scrape too. The bigger one is, the more likely they have extensive scraping tools.

Reply on news.ycombinator.com