Reddit is actually extremely good for AI. It’s a vast trove of examples of people talking to each other.
When it comes to factual data then there are better sources, sure, but factual data has never been the key deficiency of AI. We’ve long had search engines for that kind of thing. What AIs had trouble with was human interaction, which is what Reddit and Facebook are all about. These datasets train the AI to be able to communicate.
If the Fediverse was larger we’d be a significant source of AI training material too. Would be surprised if it’s not being collected already.
The “glue on pizza” thing wasn’t a result of the AI’s training, the AI was working fine. It was the search result that gave it a goofy answer to summarize.
The problem here is that it seems people don’t really understand what goes into training an LLM or how the training data is used.
So Meta’s AIs will mainly reflect non-EU cultural values.
EU cultural values include resisting against corporations doing whatever they want with our data. Let’s see meta try to reflect those.
So you want Meta’s AI to have values that don’t include resisting against corporations doing whatever they want with your data?
This is a seriously double-edged sword here. The training data of these AIs is what gives these AIs their capabilities and biases.
Training AI is not some noble endeavor that must be done no matter what. It’s a commercial grab that needs to balance utility with consumer rights.
Anyway, no matter from which parts of the world it’s trained, we’re talking about 2024 Facebook content. We’ve seen what Reddit does to an AI.
Can’t wait for meta’s cultured AI to share its wisdom with us.
Reddit is actually extremely good for AI. It’s a vast trove of examples of people talking to each other.
When it comes to factual data then there are better sources, sure, but factual data has never been the key deficiency of AI. We’ve long had search engines for that kind of thing. What AIs had trouble with was human interaction, which is what Reddit and Facebook are all about. These datasets train the AI to be able to communicate.
If the Fediverse was larger we’d be a significant source of AI training material too. Would be surprised if it’s not being collected already.
I think they were referencing glue on pizza ring and stuf .like that
The “glue on pizza” thing wasn’t a result of the AI’s training, the AI was working fine. It was the search result that gave it a goofy answer to summarize.
The problem here is that it seems people don’t really understand what goes into training an LLM or how the training data is used.
Mate, it takes the combined discussions of the internet
I would not expect any sort of values
I would expect plenty of deeply-held values:
Rule 34
Shitposts
CSAM
Disingenuous partisan mis/disinformation
Worst hot takes imaginable