this post was submitted on 07 Mar 2024
574 points (97.7% liked)
Technology
59605 readers
3156 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
If this is true, it shifts the problem from "not having it" to "not knowing which version should be used" (to train the LLM).
They could feed it the unedited versions and call it a day, but a lot of times people edit their content to correct it or add further info, specially for "meatier" content (like tutorials). So there's still some value on the edits, and I believe that Google will be at least tempted to use them.
If that's correct, editing it with nonsense will lower the value of edited comments for the sake of LLM training. It should have an impact, just not as big as if they kept no version system.
I know from experience (I'm a former Reddit janny) that moderators can't see earlier versions of the content, only the last one. The admins might though.
The one from TD, right?
Wouldn't be hard to scan a user and say:
It would.
First you'd need to notice the problem. Does Google even realise that some people want to edit their Reddit content to boycott LLM training?
Let's say that Google did it. Then it'd need to come up with a good (generalisable, low amount of false positives, low amount of false negatives) set of rules to sort those out. And while coming up with "random" rules is easy, good ones take testing, trial and error, and time.
But let's say that Google still does it. Now it's retrieving and processing a lot more info from the database than just the content and its context, but also account age, when the piece of content was submitted, when it was edited.
So doing it still increases the costs associated with the corpus, making it less desirable.
Huh? Reddit has all of this plus changes in their own DBs. Google has nothing to do with this, it's pre handover.
I'm highlighting that having the data is not enough, if you don't find a good way to use the data to sort the trash out. Google will need to do it, not Reddit; Reddit is only handing the data over.
Is this clear now? If you're still struggling to understand it, refer to the context provided by the comment chain, including your own comments.
I'm saying reddit will not ship a trashed deliverable. Guaranteed.
Reddit will have already preprocessed for this type of data damage. This is basic data engineering and trivial to do to find events in the data and understanding timeseries of events.
Google will be receiving data that is uncorrupted, because they'll get data properly versioned to before the damaging event.
If a high edit event happens on March 7th, they'll ship march 7th - 1d. Guaranteed.
Edit to be clear: you're ignoring/not accepting the practice of noting high volume of edits per user as an event, and using that timestamped event as a signal of data validity.
Nobody said anything about the database being trashed. What I'm saying is that the database is expected to have data unfit for LLM training, that Google will need to sort out, and Reddit won't do it for Google.
Do you know it, or are you assuming it?
If you know it, source it.
If you're assuming, stop wasting my time with shit that you make up and your "huuuuh?" babble.
I know it because I've worked in corporate data engineering and large data migrations and it would be abnormal to do anything else. there's a full review of test data, a scope of work, an acceptance period, etc.
You think reddit doesn't know about these utilities? You think Google doesn't?
You need to chill out and acknowledge how an industry works. I'm sure you are convinced but your idea of things isn't how the industry works.
I don't need to explain to you that the sky is blue. And I shouldn't need to explain to you that Google isn't going to accept a damaged product, and that reddit can or can't do some basic querying and timeseries manipulations.
Edit like you literally asked for a textbook.
In other words: "I dun have sauce, I'm assooming, but chruuuust me lol"
At this rate it's safe to simply ignore your comments as noise. I'm not wasting further time with you.
Seems like people are voting your comment as noise but whatever.
You are trying to prove something normal ISNT happening. I'm describing normal industry behavior.
Seems like you need to prove an abnormal sitch is occuring.
Edit it's like your asking for proof that they'll build stairs with a hand rail
It sounds like what's needed here is a version of this tool that makes the edits slowly, at random intervals, over a period of time. And perhaps has the ability to randomize the text in each edit so that they're all unusable garbage, but different unusable garbage (like the suggestion of taking ChatGPT output at really high temp that someone else made). Maybe it also only edits something like 25% of your total comment pool, and perhaps makes unnoticeably minor edits (add a space, remove a comma) to a whole bunch of other comments. Basically masking the poison by hiding it in a lot of noise?
Now you're talkin .
Intra comment edit threshold would be fun to explore
Honestly, parsing through version history is actually something an LLM could handle. It might even make more sense of it than without. For example, if someone replies to a comment and then the parent is edited to say something different. No one will have to waste their time filtering anything.
They could use an LLM to parse through the version history of all those posts/comments, to use it to train another LLM with it. It sounds like a bad (and expensive, processing time-wise) idea, but it could be done.
EDIT: thinking further on this, it's actually fairly doable. It's generally a bad idea to feed the output of an LLM into another, but in this case you're simply using it to pick one among multiple versions of a post/comment made by a human being.
It's still worth to scorch the earth though, so other human users don't bother with the platform.