Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

People paid to train AI are outsourcing their work to AI

technologyreview.com

Post reply on HN
Print view View original

221–230 of 233 posts

Re: People paid to train AI are outsourcing their work to AI

#221

Post by jahsome »

olalonde wrote: post #126
jahsome wrote: post #51

Earlier quoted context omitted.

This is funnier than it has any right to be, mostly given the over-abundant "AI will fix everything" narrative currently dominating the hn discourse. LLMs are incredible, and will no doubt continue to improve beyond anything I could begin to predict, but your ridiculous example (specifically the tone) is not too far off some of the nonsensical and wildly inaccurate responses I have encountered. I find rhe unfailing c…

> given the over-abundant "AI will fix everything" narrative currently dominating the hn discourse. It seems we are not reading the same HN.

True, it's died down a bit in recent weeks. But it's still taken crypto's place of the mind share.

Re: People paid to train AI are outsourcing their work to AI

#223

Post by chickenpotpie »

lisasays wrote: post #207
chickenpotpie wrote: post #161

This article is complete bunk. The researchers used a "chatgpt detector" which as we've seen over and over in academia, do not work. This study is completely unfounded. God I'm choking on the irony of an article about the dangers of using AI to train AI based on a study that used AI to detect AI

Per the article, the didn't just use the static detector: They also extracted the workers’ keystrokes in a bid to work out whether they’d copied and pasted their answers, an indicator that they’d generated their responses elsewhere. So while I don't yet know if the article is bunk -- I do know that your hot take is bunk.

Or they typed up their responses in a different text editor and copy and pasted from that?

Re: People paid to train AI are outsourcing their work to AI

#225

Post by lisasays »

vminvsky wrote: post #220
lisasays wrote: post #207

Earlier quoted context omitted.

Per the article, the didn't just use the static detector: They also extracted the workers’ keystrokes in a bid to work out whether they’d copied and pasted their answers, an indicator that they’d generated their responses elsewhere. So while I don't yet know if the article is bunk -- I do know that your hot take is bunk.

They never used a static detector.

"Static" in this case refers its being used in isolation (not to any particular kind of detector).

I could have said "they didn't just use the detector all by itself", I suppose.

Re: People paid to train AI are outsourcing their work to AI

#226

Post by lisasays »

iinnPP wrote: post #201
vminvsky wrote: post #177

Earlier quoted context omitted.

according to the paper they get 98% accuracy. another recent paper came out saying it's always possible to discriminate between real and synthetic text [1]. i think the core problem is with the generalist classifiers (gptzero, openai detector, etc). ex. openai's classifier has an accuracy of around 25% on it's own text. however, when you train a bespoke classifier (like the authors did), you can get really good resul…

The moment a detector is taken seriously is the moment it will be trivially beaten by another AI designed to beat the detector.

Is it now?

Adversarial training isn't infinitely scalable either, has its limitations also.

Also - the moment that companies start training models to resist detectors, they expose themselves to regulation. Won't stop dark AI models running on some website somewhere, but it can be very effectively applied to companies running at Google or OpenAI scale.

Re: People paid to train AI are outsourcing their work to AI

#227

Post by sidewndr46 »

pavlov wrote: post #167
sidewndr46 wrote: post #116

Earlier quoted context omitted.

But that wasn't a scare. It's a real thing that made people ill with no cure.

I didn’t mean to imply it wasn’t a real thing. (I’m not a native English speaker so I’m probably missing a nuance of “scare” beyond what appears to be its literal meaning of the general public being afraid.)

"scare" in English is often used to describe a public over-reaction to a perceived threat. In the US we had the "red scare" that was an over reaction to communists in the US.

A similar concept is the "satanic panic", the idea that all our children were being abducted by satanic cults.

Re: People paid to train AI are outsourcing their work to AI

#228

Post by astrange »

avgcorrection wrote: post #175
astrange wrote: post #172

Earlier quoted context omitted.

This is a lump of labor fallacy. Automation has the opposite effect of what you just said.

Yours is a constant trajectory fallacy: things will continue to develop as they have forever. (Easy to dismiss things once you give it a name.) I’m merely taking the LLM disruption hype at face value.

Not my problem if people hyping it are also wrong.

Also mine's real and you just made that one up…

Re: People paid to train AI are outsourcing their work to AI

#229

Post by avgcorrection »

astrange wrote: post #228
avgcorrection wrote: post #175

Earlier quoted context omitted.

Yours is a constant trajectory fallacy: things will continue to develop as they have forever. (Easy to dismiss things once you give it a name.) I’m merely taking the LLM disruption hype at face value.

Not my problem if people hyping it are also wrong. Also mine's real and you just made that one up…

Anything that questions the maxim of eternal growth is a fallacy according to mainstream Economics.

Re: People paid to train AI are outsourcing their work to AI

#230

Post by wizzwizz4 »

startupsfail wrote: post #178
wizzwizz4 wrote: post #88

Earlier quoted context omitted.

I assumed OpenAI had added that themselves.

They’ve added incorrect answers?

No; those were there already. I assumed they'd added the "does not double down" behaviour. Bing Chat (also an OpenAI model, but presumably with a different set of fine-tuning / filters to ChatGPT – possibly the bare OpenAI API) doubles down in situations like this: https://nitter.dark.fail/_akhaliq/status/1672267392280571905

GPT models can't tell the difference between truth and fiction. All you can choose by fine-tuning is their threshold for "admitting" mistakes.

Post reply on HN