this post was submitted on 21 Feb 2024
289 points (95.0% liked)
Technology
61871 readers
2746 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 other!
- 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
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
The development of LLMs is possibly becoming self defeating, because the training data is being filled not just with human garbage, but also AI garbage from previous, cruder LLMs.
We may well end up with a machine learning equivalent of Kessler syndrome, with our pool of available knowledge eventually becoming too full of junk to progress.
I mean, surely the solution to that would be to use curated/vetted training data? Or at the very least, data from before LLMs became commonplace?
Yes but that only works if we can differentiate that data on a pretty big scale. The only way I can see it working at scale is by having meta data to declare if something is AI generated or not. But then we're relying on self reporting so a lot of people have to get on board with it and bad actors can poison the data anyway. Another way could be to hire humans to chatter about specific things you want to train it on which could guarantee better data but be quite expensive. Only training on data from before LLMs will turn it into an old people pretty quickly and it will be noticable when it doesn't know pop culture or modern slang.
Pretty sure this is why they keep training it on books, movies, etc. - it's already intended to make sense, so it doesn't need curated.