[cowanon77]

On the surface it seems bad, although many of these books were probably rotting in place rather than being read. If the companies are willing to make the digital version available this may actually prreserve the books.

More troubling to me is the focus on pre-2022 data. This indicates that the companies are seriously worried about model collapse, where the future models become worse becuase they are being fed by generated data instead of real human data.

This is actually my worse case scenario for AI: the models become good enough to disrupt entire industries in the next 10 years, but then stay frozen at that level. Model drift then kicks in because the world will keep changing but the models do not, and in a few decades we actually regress instead of progress becuase they won't be enough skilled humans to drive advances.

[datakan]

Entropy is a thing.

Ancient Rome stopped creating aqueducts because they had all the ones they needed. They failed to pass that knowledge on to the next generation and so they just forgot how to create aqueducts.

Once AI starts making the majority of content, people will simply forget how to make content. Before we know it we're all fat slobs in floating chairs like in Wall-E.

Best case scenario I think is similar to what we see in Ian M. Banks Culture books where the machines basically take care of us out of the goodness of their hearts and we just kinda fuck off into obscurity.

[inigyou]

That's the western world right now with manufacturing. Meanwhile China did the manufacturing and India is coming up, so it's not a lost skill globally.

The next generations are also losing computer skills, being brought up with devices that handhold you and have no opportunity for customisation and are essentially thin clients for mega corporations.

[storus]

As long as economics incentivizes next quarter thinking nothing will be done. We can't even handle human portion of global warming which is arguably more important than just some AI model collapse scenario.

[inigyou]

What would happen if we created large volumes of books with AI but faked their dates before 2022 to induce model collapse?

[jmyeet]

Yeah, I really don't know where this goes from here. Like certain sources, particular Reddit and Twitter, have to largely be considered spoiled at this point. It's hard to quanitfy the amount of bots but my intuition tells me it's really high.

But changing how non-AI people write, that's an interesting angle. Because where do we go from here? In 100 years will we still be overvaluing pre-AI sources? That doesn't make sense. Of course a lot can (and will) change in 100 years.

But i'm reminded of pre-atomic steel, which is steel made before the first atomic bombs were detonated and thus have really low background radiation. This is necessary for making MRIs and such. People will go and find it from shipwrecks and such (ironically, many of which are from WW2). It's also a finite resource. What happens when we run out?

Now pre-2022 texts aren't consumed (other than destructively scanning books of course) but it is also finite. We can't make more of it.