> The actual scary stuff is the dilution of expertise, we contributed for a long time to share our knowledge for internet points (stack overflow, open source projects, etc), and it has been harvested by the AIs already, anyone that pays access to these services for tens of dollars a month can bootstrap really quickly and do what it might had needed years of expertise before.
What scares me more is the opposite of that: information scarcity leading to less accessible intelligence on newer topics.
I’ve completely stopped posting on Reddit since the API changes, and I was extremely prolific before[1] because I genuinely love writing about random things that interest me. I know I’m not the only one: anecdotally, the overall quality of content on Reddit has nosedived since the change and while there doesn’t seem to be a drop in traffic or activity, data seems to indicate that the vast majority of activity these days is disposable meme content[2]. This seems to be because they’re attempting desperately to stick recommendation algorithms everywhere they can, which are heavily weighted toward disposable content since people view more of it. So even if there were just as many long discussion posts like before, they’re not getting surfaced nearly as often. And discussion quality when it does happen has noticeably dipped as well: the Severance subreddit has regularly gotten posts and comments where people question things that have already been fully explained in the series itself (not like subtext kind of things, like “a character looked at the camera and blatantly said that in the episode you’re talking about having just watched” things). Those would have been heavily downvoted years ago, now they’re the norm.
But if LLMs learn from the in-depth posting that used to be prominent across the Internet, and that kind of in-depth posting is no longer present, a new problem presents itself. If, let’s say, a new framework releases tomorrow and becomes the next big thing, where is ChatGPT going to learn how that framework works? Most new products and platforms seem to centralize their discussion on Discord, and that’s not being fed into any LLMs that I’m aware of. Reddit post quality has nosedived. Stack Overflow keeps trying to replace different parts of its Q&A system with weird variants of AI because “it’s what visitors expect to see these days.” So we’re left with whatever documentation is available on the open Internet, and a few mediocre-quality forum posts and Reddit threads.
An LLM might be able to pull together some meaning out of that data combined with the existing data it has. But what about the framework after that? And the language after that? There’s less and less information available each time.
“Model collapse” doesn’t seem to have panned out: as long as you have external human raters, you can use AI-generated information in training. (IIRC the original model collapse discussions were the result of AI attempting to rate AI generated content and then feed right back in; that obviously didn’t work since the rater models aren’t typically any better than the generator models.) But what if the “data wells” dry up eventually? They can kick the can down the road for a while with existing data (for example LLMs can relate the quirks of new languages to the quirks of existing languages, or text to image models can learn about characters from newer media by using what it already knows about how similar characters look as a baseline), but eventually quality will start to deteriorate without new high-quality data inputs.
What are they gonna do then when all the discussion boards where that data would originate are either gone or optimized into algorithmic metric farms like all the other social media sites?
[1]: https://old.reddit.com/user/Nathan2055
[2]: I can’t find it now, but there was an analysis about six months ago that showed that since the change a significant majority of the most popular posts in a given month seem to originate from /r/MadeMeSmile. Prior to the API change, this was spread over an enormous number of subreddits (albeit with a significant presence by the “defaults” just due to comparative subscriber counts). While I think the subreddit distribution has gotten better since then, it’s still mostly passive meme posts that hit the site-wide top pages since the switchover, which is indicative of broader trends.