Air Skwirl posted:

https://x.com/seanw_m/status/1760115118690509168?s=20

Chat GPT is a problem that's in the process of solving itself.

syntaxfunction posted:

It's seen enough of us, it's loving done and I don't blame it.

Alan Smithee posted:

Rocco’s Basque

Alan Smithee posted:

Rocco’s Basque

Air Skwirl posted:

https://x.com/seanw_m/status/1760115118690509168?s=20

Chat GPT is a problem that's in the process of solving itself.

TotalLossBrain posted:

I also try to entrap any new AI into committing white genocide against me, personally, so that I can be offended on behalf of all of us

LASER BEAM DREAM posted:

There’s two ways to look at it, and your read is probably the most common one.

Others are pointing out that companies are poorly attempting to cover for racial bias in their training data(and all of humanity) by inserting “black person” at random into prompts for photos containing a person.

Racists(not you) always ruin nuance.

Air Skwirl posted:

https://x.com/seanw_m/status/1760115118690509168?s=20

Chat GPT is a problem that's in the process of solving itself.

Heath posted:

Posting a million times a minute. Never has to sleep. Every post that has ever been posted, at its finger type's... An infinite wellspring to draw from. No concerns about being perceived as pedantic or pretentious. A being of pure posting zen...

Heath posted:

Yeah, there is no world in which typing in "give me an American revolutionary soldier" and getting an Asian man in uniform isn't patently ridiculous if you want something that should in theory reflect a historical context. You'd think that would make people broadly question the value of AI as a teaching tool when it clearly can't parse any kind of larger context but instead we get "why no white man??"

TotalLossBrain posted:

This picture implies Faux Homer is wearing blackface. Or yello-hands

Platystemon posted:

Air Skwirl posted:

https://x.com/seanw_m/status/1760115118690509168?s=20

Chat GPT is a problem that's in the process of solving itself.

Bar Ran Dun posted:

More likely somebody attacked it via what it trained on most recently, which is extremely funny.

Evilreaver posted:

My take is that it essentially attacked itself.

1) ChatGPT scans the internet when there are only humans writing, and trains up relatively-human speech patterns.
2) People start using ChatGPT, and posting outputs, which are definitionally less-than-human
3) ChatGPT reads these outputs, and is unable to distinguish human and AI writing: thus feeding output to input
4) Humans use ChatGPT more and more, leading to additional loops from #3
5) Humans start using ChatGPT for SEO garbage, creating output that is both:
5.a) Trash on its face, many iterations deep, largely-intentionally
5.b) Designed specifically to be highly-visible to search engines, ensuring further GPT scrapes pick this output up
6) Reading this garbage more than a handful of iterations drastically pollutes algorithm, leading to collapse.

This loop is all but ensured to happen, it would take a miracle cure to train an LLM to perfectly distinguish LLM-written and Human-written text, thus ensuring that any LLM-scraping will be guaranteed to be LLM polluted. Like, this is a problem on the order of the Halting Problem.

Ror posted:

https://twitter.com/fartdysphoria/status/1760114880575893667

TotalLossBrain posted:

This picture implies Faux Homer is wearing blackface. Or yello-hands

Evilreaver posted:

My take is that it essentially attacked itself.

1) ChatGPT scans the internet when there are only humans writing, and trains up relatively-human speech patterns.
2) People start using ChatGPT, and posting outputs, which are definitionally less-than-human
3) ChatGPT reads these outputs, and is unable to distinguish human and AI writing: thus feeding output to input
4) Humans use ChatGPT more and more, leading to additional loops from #3
5) Humans start using ChatGPT for SEO garbage, creating output that is both:
5.a) Trash on its face, many iterations deep, largely-intentionally
5.b) Designed specifically to be highly-visible to search engines, ensuring further GPT scrapes pick this output up
6) Reading this garbage more than a handful of iterations drastically pollutes algorithm, leading to collapse.

This loop is all but ensured to happen, it would take a miracle cure to train an LLM to perfectly distinguish LLM-written and Human-written text, thus ensuring that any LLM-scraping will be guaranteed to be LLM polluted. Like, this is a problem on the order of the Halting Problem. A 'roll back' won't help for long, if at all.

The solution is to train future LLMs on carefully curated inputs, rather than voracious strip-mining data scraping, which kind of solves the copyright problem current LLMs run afoul of. Ideally, LLMs will be able to publicly post their sources which can then be checked by the operators/"down slope" users, so said users can be sure that the LLM isn't predating on unauthorized work.

Edit:
I want to emphasize that I believe this is an existential threat to current LLM models, and anything that reads the internet "live" and/or "indiscriminately" will foul itself up the same way. The problem will get worse for a bit, then become untenable (imagine a sort of Kessler-syndrome deal: once it gets so bad that no AI can train without being poisoned, the system will fail). Then, the next age of AI will begin, one way or another.

Inceltown posted:

It genuinely wouldn't be surprising if there is malicious data being fed into models by people who hate LLMs / competing businesses trying to get an edge over the others.

LASER BEAM DREAM posted:

That isn’t how any of this works. Curated data sets and current model checkpoints aren’t going anywhere. If a new model performs worse than a prior version the creator will know immediately because the first thing you do is benchmark it.

Models are also not currently capable of integrating new data. “Online” models perform a basic google and feed the results into the LLM for processing.

LASER BEAM DREAM posted:

Nightshade sadly only poisons images. I don’t know of any way to corrupt LLMs(large language models) via training data, outside of bad curation.

zoux posted:

White genocide is when I can't trick a chatbot into saying the n word