Fuck AI

AI ingesting AI slop and falling apart is not dissimilar to boomers ingesting rightwing slop and falling apart.

If I had the money and a computer able to handle the amount of stuff I'd be throwing at it with a local model, I would have a giant website full of AI generated nonsense purely for the purpose of letting AI gobble it up to help the AI incest problem.

Imagine if a whole metric ton of "websites" did this. The thieving AI companies would either have to start blocking all of these sites or deal with an issue they don't wanna because they're too stingy and will probably just have their AI try ( and fail ) to fix the problem.

Human society does the same thing.

We're ok when we talk about what we saw.

Less so when we talk about what somebody else saw.

Crazier and crazier when we talk about what somebody said about what somebody said about what somebody saw. Which is arguably the internet.

Ffs, neural networks and LLMs have their place and can be useful, but setting up datacentres that snort up the entire internet indiscriminately to create a glorified chatbot that spews data that may or may not be correct is insane.

I predicted this. It is similar to a photocopy of a photocopy that eventually ends up a mess of garbage noise.

I've been predicting this for a while now and people kept telling me I was wrong. Prepare for dot com burst two, electric boogaloo.

The silver lining of AI slop filling the WWW

fill up your free cloud services with ai generated info. i mean thousand text file. like "how to make homemade butterfly". all of them will scrap by ai.

Good. Eat yourself you technological prion disease.

Another problem I've realized today, is the proliferation of data that was originally hallucinated by AI.

I was discussing an issue on a software with a coworker and he asked an AI for help configure around it. He then sent me "apparently we can try changing this setting to this value". I told him to first validate if that setting really existed because AI tends to make up things like that when it's what you would want to hear and running a test would take us 20~30 minutes.

He found some discussions about that setting not working as people expected. "ok at least it exists then" and we tried it. It didn't work. I later cloned the source of that software and checked, the setting didn't exist - ever.

There is a solution to this. Make a **perfect ** AI detecting tool. The only way I can think of is through adding a tag to every bit of AI-generated data,

Though it could easily be removed from text, I guess.And no, training AI to recognize AI will never work. Also every model would have to join this, or it won't work.

Related XKCD

Good. Poison the AI well. Rot this shit to the ground.

Aww boo hoo, did someone generate a degenerative feedback loop? Yeah? Did someone make a big ol' oppsy whoopsy that's gunna accelerate in hallucinations and slop as it collapses in on itself? How's the coded version of a microphone whine going to go, you silly buttholes?

Garbage in, garbage out.

Who could have possibly predicted that?

Oh no! I HOPE us Taxpayers can Bail Out these AI Companies when they go Under! AFTER ALL we CUT my Child's LIFESAVING MEDICATION so I KNOW we have the Funds to Help these Poor Billionaire CEOS!

So, reading this article, it's not about model collapse, but about RAG - letting the AI model google the question essentially. The problem is, the first 10 pages of google search results are all low effort adfarming slop sites, because of course it is, which is making the answers from the AI worse, as these slop sites often have incorrect or otherwise unproofed articles, which biases the AI to fork out the wrong answer.

I'm sure the major AI services will try and fix this with some slop site detection routines.

It's very tempting to have schadenfreude about this failure but also disgusting that so much has been invested in it that should have been put to better use.

It's just another example of a system whose narrow definition of success is taking human and environmental value and using it to extract more. It's not aimed at solving worthwhile problems or making things better, which is why people are becoming more miserable and the planet is getting wrecked.

You could say that it's the system we live in which is the AI, feeding on itself and becoming more sick.

Cue Price is Right failure trombone.

The great news is that these ponze schemes will either collapse or spend the next decade trying to fix it by creating algorithms which detect AI content so as to filter it out.

makes me think about the human centipede