994
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 18 Sep 2026
994 points (98.6% liked)
Technology
88185 readers
3155 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- 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, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
Gonna get real weird soon when they run out of rich new training data and they begin to consume their own shit. When that happens their entire model will collapse and if the bubble hasn't popped already that may be what causes it.
They're already doing it. They call it distilling when they take it from another llm. Pretty sure most content was already stolen in the early days and they now rely on distillation and stealing new content.
So it's like if everyone combined the backwash from their water bottles into a new drink...
Sounds super appealing!
Don't get high off your own supply?
That's my thought. So long as someone inputs anything new to the internet they will be able to scrape it and sell it as their product.
Pretty sure they mostly use the pre-AI internet (that they scraped and kept) and synthetic data currently. Probably trying (and failing so far or we'd have heard) to adapt to using video as training material at the moment, but developments there will likely apply to robotics at some point. Here's hoping the current chuds have crashed and burned before then and that some sanity has taken over from unfettered capitalist oligarchs dreams of computer slavery.
Why don't you thinj they have not adapted to video as training. Is that not what Flock does?
Flock does pattern recognition, a quite old piece of machine learning. Nothing to do with training a large language model or other 'AI' model.
They almost certainly use that to train AI. Big tech is aggregating the data from many places.
Nah, it's in the name, Large Language Models, what gets marketed as 'AI', are trained on text. That's why they're ripping up secondhand books (and copyright law, but that never stopped them) at the moment. No one has really cracked Large Vision Model yet, although there are some primitive versions being attempted in robotics, mostly they translate to text, which is why general purpose robots aren't a thing yet.
"I am sitting in a room"
https://en.wikipedia.org/wiki/I_Am_Sitting_in_a_Room
Why do you think there has been such an emphasis on hacking with LLMs lately (especially by OpenAI). They figured out all those vulnerability databases were an excellent source for training material. In the past they scraped those, but just for general language training. Now they've specifically trained the models on the information within. Some team figured out how to use that data to train a model and have testing scenarios automated so they could write a good reward function. It wasn't that they figured out the models are good at hacking, they ran out of content and found a new source of good data.
With all the books they've been scanning I wonder if the next thing is going to be a writing assistant or editor or something like that. Even though writing good books is an art form and the actual writing down of the words is the easiest part (still not easy tho).
These companies are starving for content and they've not just poisoned but absolutely destroyed the content well that is the internet. Given they were already hitting diminishing returns hard, it doesn't matter too much to them probably. But more compute and storage has also been hitting diminishing returns hard and customers are complaining about the cost. So they are getting a bit desperate on how to improve these things at all.
Or they can just keep a database of actual data on their servers, instead of getting new data every single time for some fucking reason.
model collapse is the endgame. that's the whole point of LLM.
What do you mean? Why would collapse be the endgame?
It's either agi or model collapse. For every AI system actually. If you have a high enough adoption you start to muss original data so if your AI is not self sufficient in time it will collapse, because it will be trained on its own data, which just is an incentive loop
in a manner of speaking - you always end up there. not by design though. models operate via continuous refinement and you can only optimize a model so much until it is a mess and you need to figure out where to roll back. so you either get shit like semantic drift or variance decay and you can whack a mole it to an extent but then you hit the rlhf wall when the model starts gaming its reinforcement framework and the fat lady sings.
Well put. Never thought of it that way.
the only more or less workable way to keep it under control is maintaining a closed loop small-scale environment - kinda like NotebookLM where you upload documents and that's all there is to work with - outside of that it is a mess.