this post was submitted on 22 Feb 2024
806 points (98.1% liked)
Technology
59389 readers
3236 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
Ai actually has huge problems with this. If you feed ai generated data into models, then the new training falls apart extremely quickly. There does not appear to be any good solution for this, the equivalent of ai inbreeding.
This is the primary reason why most ai data isn't trained on anything past 2021. The internet is just too full of ai generated data.
Pay intelligent humans to train AI.
Like, have grad students talk to it in their area of expertise.
But that's expensive, so capitalist companies will always take the cheaper/shittier routes.
So it's not there's no solution, there's just no profitable solution. Which is why innovation should never solely be in the hands of people whose only concern is profits
OR they could just scrape info from the "aska____" subreddits and hope and pray it's all good. Plus that is like 1/100th the work.
The racism, homophobia and conspiracy levels of AI are going to rise significantly scraping Reddit.
Even that would be a huge improvement.
Just have a human decide what subs it uses, but they'll just turn it losse on the whole website
That reminds me, any AI trained on exclusively Reddit data is going to use lose vs. loose incorrectly. I don't know why but I spotted that so often there.
Its a loose-lose situation
And the "would of" thing
Ooh ooh and "tow the line"
Haha. Grad students expensive. God bless.
And unlike with images where it might be possible to embed a watermark to filter out, it's much harder to pinpoint whether text is AI generated or not, especially if you have bots masquerading as users.
This is why LLMs have no future. No matter how much the technology improves, they can never have training data past 2021, which becomes more and more of a problem as time goes on.
You can have AIs that detect other AIs' content and can make a decision on whether to incorporate that info or not.
Fun fact. You can't. Ais are surprisingly bad at distinguishing ai generated things from real things.
What is this then?
https://copyleaks.com/ai-content-detector
Just because a tool exists doesn't mean it's particularly good at what it's supposed to do.
Yeah, just tested the website using bots from reddit's SubSimulatorGPT2. Only got 1/13 correct.
can you really trust them in this assessment?
Doesn't look like we'll have much of a choice. They're not going back into the bag.
We definitely need some good AI content filters. Fight fire with fire. They seem to be good at this kind of thing (pattern recognition), way better than any procedural programmed system.
last time i've checked ais are pretty bad at recognizing ai-generated content
anyway there's xkcd about it https://xkcd.com/810/