Mechanical Turk workers are using AI to automate being human • Hacker News Board

Re: Mechanical Turk workers are using AI to automate being human

#51

Post by the8472 » Fri, Jul 07, 2023, 2:04 PM UTC

boesboes wrote: post #10

bmmayer1 wrote: post #4

This is going to cause problems as many Mechanical Turk tasks are being used to train AI models...

And even fine-tuning using AI causes the models to quickly degrade I've seen here previously. Perhaps we are already at peak of these LLMs now and the next gen won't come because all training data will be poisened with AI generate crap. That would be kinda funny to me tbh

That seems unlikely to me. Models are multi-modal now, so they'll likely train on video and audio content too. Once that's tapped out they can learn better world models by not passively ingesting data but interacting with it, e.g. via simulated environments or capturing human-bot interactions that are flagged as interesting and incorporating those in the next run.

Re: Mechanical Turk workers are using AI to automate being human

#52

Post by lacrimacida » Fri, Jul 07, 2023, 2:09 PM UTC

jwie wrote: post #28

“AI” eating itself is the best outcome for everyone. It’s a mass energy balance. The inputs equal the outputs; garbage in, garbage out.

It’s not “AI”, it’s just LLMs. When/if AI becomes a reality there will be no need for mechanical turks

Re: Mechanical Turk workers are using AI to automate being human

#53

Post by Salgat » Fri, Jul 07, 2023, 2:11 PM UTC

bmmayer1 wrote: post #4

This is going to cause problems as many Mechanical Turk tasks are being used to train AI models...

Yes and no. You can utilize ML to do 90% of the work in preparing an input, then have a human verify to ensure it's correct. It's similar to how ML models are starting to consume ML generated artwork, which people think is bad, but in reality those ML generated artworks are the ones humans found good enough to publish online, so they still retain value for training (as being appealing to humans).

Re: Mechanical Turk workers are using AI to automate being human

#54

Post by kortex » Fri, Jul 07, 2023, 2:14 PM UTC

lacrimacida wrote: post #52

jwie wrote: post #28

“AI” eating itself is the best outcome for everyone. It’s a mass energy balance. The inputs equal the outputs; garbage in, garbage out.

It’s not “AI”, it’s just LLMs. When/if AI becomes a reality there will be no need for mechanical turks

It is AI; artificial intelligence has been around since the 60s. It's not AGI, and no one (credible) is claiming LLMs are AGI.

Re: Mechanical Turk workers are using AI to automate being human

#55

Post by edgyquant » Fri, Jul 07, 2023, 2:21 PM UTC

mahathu wrote: post #50

nomilk wrote: post #35

Earlier quoted context omitted.

> Mapmakers deliberately put slight mistakes in their maps so they can tell when someone copies them. If another map has the same mistake, that's very convincing evidence. http://www.paulgraham.com/say.html

> What scares me is that there are moral fashions too. They're just as arbitrary, and just as invisible to most people. But they're much more dangerous. Fashion is mistaken for good design; moral fashion is mistaken for good. What dog whistling drivel. That dude's trying so hard to use fancy words and walls of texts to defend his racist views without stating them outright. This is coming from a person who previously…

“That dude,” is the owner of this website and not everything is bad or racist just because it goes against your view of the world. “Dog whistle,” give me a break.

Re: Mechanical Turk workers are using AI to automate being human

#56

Post by cwillu » Fri, Jul 07, 2023, 2:27 PM UTC

sillysaurusx wrote: post #12

boesboes wrote: post #10

Earlier quoted context omitted.

And even fine-tuning using AI causes the models to quickly degrade I've seen here previously. Perhaps we are already at peak of these LLMs now and the next gen won't come because all training data will be poisened with AI generate crap. That would be kinda funny to me tbh

The invention of nukes caused a similar problem. Scientists had to harvest metal from pre-ww2 sunken battleships, since that’s the only metal not contaminated with some measurable amount of radiation. I forget the details, but it’s very likely pre-chatgpt training data will be prioritized, with exceptions for news and current events (which largely doesn’t matter whether it’s AI generated anyway).

https://en.wikipedia.org/wiki/Low-background_steel

Note also: “Since the end of atmospheric nuclear testing, background radiation has decreased to very near natural levels,[3] making special low-background steel no longer necessary for most radiation-sensitive applications, as brand-new steel now has a low enough radioactive signature that it can generally be used in such applications.”

Re: Mechanical Turk workers are using AI to automate being human

#57

Post by ilyt » Fri, Jul 07, 2023, 2:45 PM UTC

jwie wrote: post #28

“AI” eating itself is the best outcome for everyone. It’s a mass energy balance. The inputs equal the outputs; garbage in, garbage out.

I mean yeah, while wasting a lot of energy and brainpower...

Kinda like crypto but on smaller scale.

Re: Mechanical Turk workers are using AI to automate being human

#58

Post by CharlesW » Fri, Jul 07, 2023, 2:46 PM UTC

TZubiri wrote: post #27

sillysaurusx wrote: post #12

Earlier quoted context omitted.

The invention of nukes caused a similar problem. Scientists had to harvest metal from pre-ww2 sunken battleships, since that’s the only metal not contaminated with some measurable amount of radiation. I forget the details, but it’s very likely pre-chatgpt training data will be prioritized, with exceptions for news and current events (which largely doesn’t matter whether it’s AI generated anyway).

Also happens with wikipedia. Later edits with web references are much more prone to citogenesis, while an early edit with a web reference is probably clear, even if the link is dead. Even printed references can be based on wikipedia now

> …citogenesis…

https://en.wikipedia.org/wiki/Circular_reporting#Circular_re...

Re: Mechanical Turk workers are using AI to automate being human

#59

Post by BoxOfRain » Fri, Jul 07, 2023, 2:52 PM UTC

sillysaurusx wrote: post #12

boesboes wrote: post #10

Earlier quoted context omitted.

And even fine-tuning using AI causes the models to quickly degrade I've seen here previously. Perhaps we are already at peak of these LLMs now and the next gen won't come because all training data will be poisened with AI generate crap. That would be kinda funny to me tbh

The invention of nukes caused a similar problem. Scientists had to harvest metal from pre-ww2 sunken battleships, since that’s the only metal not contaminated with some measurable amount of radiation. I forget the details, but it’s very likely pre-chatgpt training data will be prioritized, with exceptions for news and current events (which largely doesn’t matter whether it’s AI generated anyway).

Apparently a lot of it comes from what remains of the German fleet that scuttled itself at Scapa Flow rather than be divided among the allies. Most of the ships were salvaged between the world wars but some remain.

Re: Mechanical Turk workers are using AI to automate being human

#60

Post by m3kw9 » Fri, Jul 07, 2023, 2:52 PM UTC

jwie wrote: post #28

“AI” eating itself is the best outcome for everyone. It’s a mass energy balance. The inputs equal the outputs; garbage in, garbage out.

Why would you not root for AI? Its potential to improve humanity is huge even if there are potential bad sides. Some People bunching crypto with AI on the usefulness scale is quite ridiculous