[indigochill]

> You no longer have to hire a copywriter, as many SEO spammers used to do.

I used to do SEO copywriting in high school and yeah, ChatGPT's output is pretty much at the level of what I was producing (primarily, use certain keywords, secondarily, write a surface-level informative article tangential to what you want to sell to the customer).

> At some point it may become impossible to separate the wheat from the chaff.

I think over time there could be a weird eddy-like effect to AI intelligence. Today you can ask ChatGPT a Stack Overflow-style and get a Stack Overflow-style response instantly (complete with taking a bit of a gamble on whether it's true and accurate). Hooray for increased productivity?

But then, looking forward years in time, people start leaning more heavily on that and stop posting to Stack Overflow and the well of information for AI to train on starts to dry up, instead becoming a loop of sometimes-correct goop. Maybe that becomes a problem as technology evolves? Or maybe they train on technical documentation at that point?

[zeruch]

I think you are generally correct in where things will likely go (sometimes correct goop) but the problem I think will be far more existential; when people start to feel like they are in a perpetual uncanny valley of noise, what DO they actually do next? I don't think we have even the remotest grasp of what that might look like and how it will impact us.

[kbrkbr]

That is an interesting thought. Maybe the problem is not the ai generated useless noise, but that it is so easy and cheap to publish it.

One possible future is going back to a medium with higher cost of publication. Books. Handchiseled stone tablets. Offering information costs something.

[BetaDeltaAlpha]

This was the original use case of bitcoin's Proof of Work system. Initally it was to impose a (nominal) fee on senders of email by mail clients.

If you didn't submit a proof of work of N or greater difficulty the email would be thrown out.

[Swizec]

> One possible future is going back to a medium with higher cost of publication. Books

Honestly I’ve switched to books and papers a few years ago and it has been fantastic. 2 hours of reading a half decent book or paper outweighs a week of reading the best blogposts, twitter threads, or YouTube videos.

[sanroot99]

I generally go to cited papers in Wikipedia articles

[tjr]

Do you have a favorite source for papers?

[Swizec]

HN surfaces a lot of good ones. Sometimes friends recommend stuff. Or I search for things I'm interested in.

Then once you have a hook into your topic, it usually cites 30+ other papers that may be worth reading. You will never run out.

[latexr]

> One possible future is going back to a medium with higher cost of publication. Books.

The grifters are all over that already. No AI necessary to generate and publish drivel.

See “Contrepreneurs: The Mikkelsen Twins”¹ by Dan Olson² for an informative and entertaining documentary on the matter.

¹ https://www.youtube.com/watch?v=biYciU1uiUw

² A.k.a Folding Ideas. A.k.a. the creator of “Line Goes Up – The Problem With NFTs”.

[penjelly]

fun thought: its more reliable to store information on stone tablets over very long time periods of time then it is harddrives or other modern data storage devices

[jt2190]

I think we have plenty of examples of published “noise”, probably just not on the same scale. (“Noise” is subjective of course: I don’t watch reality television but others do, for example.) For the most part, I just ignore “noise”, so I suspect that the entire World Wide Web will eventually be considered “noise” by many. Instead it seems like it will be necessary to deploy AI to retrieve information as it will be necessary to programmatically evaluate the received content to filter out anything that you’ve trained it to consider “noise”.

[zeruch]

"(“Noise” is subjective of course: I don’t watch reality television but others do, for example.)"

This brings up a good sub-topic. "Noise" as I mean it is where it's something you cannot definitely validate the veracity of in short order, or you do and it's useless.

The trash TV thing is a great example: if you are watching Beavis & Butthead because you know its trash and you need to zone out, that's a conscious, active decision, and you are in effect, 'in on the joke'...if you can't discern that it's satire and find yourself relating to the characters, you might be part of the problem :)

[Tiktaalik]

Spending less time on the internet in general or perhaps hyper strict closed off walled garden social networks for humans only.

[soupfordummies]

...Leading to the interesting thought-experiment/SF-story-concept of, "how do you prove you're human to a computer?"

[spaceman_2020]

Its already becoming hard to tell the wheat from the chaff.

AI generated images used to look AI generated. Midjourney v6 and well tuned sdxl models look almost real. For marketing imagery, Midjourney v6 can easily replicate images from top creative houses now.

[ren_engineer]

>But then, looking forward years in time, people start leaning more heavily on that and stop posting to Stack Overflow and the well of information for AI to train on starts to dry up

for coding tasks I'd imagine it could be trained on the actual source code of the libraries or languages and determine proper answers for most questions. AI companies have seen success using "synthetic" data, but who knows how much it can scale and improve

[WalterBright]

I've rarely found stackoverflow to give useful answers. If I am looking for how to do something with Linux programming, I'll get a dozen answers, half of which are only partial answers, the other half don't work.

[nuancebydefault]

Weird, I've found hundreds of useful SO answers that worked for me.

I've also learned a lot from chatting with Bing AI. The caveat is there that you all the time have in the back of your mind that the answer might be wrong. It helps to keep asking more detailed questions and check whether the set of answers keep on making sense as a whole. That way of using it has helped me a lot. See it as getting info from a very smart friend who sometimes had too much to drink.

[mort96]

And to be fair, I've rarely found ChatGPT to give useful answers ... so I guess it produces perfect StackOverflow-like answers?

[WalterBright]

When I've used chatgpt and bard to write example code, it's always generated a complete example, not half of one.

Of course, I carefully frame the query so that's what I get.

However, when I asked stackoverflow, google, and bard a question about how to do something with github, all I received were wrong answers. I finally had to throw in the towel and ask people. I think it was the third person I asked who gave an answer that worked.

google itself has an annoying habit of answering a fundamentally different question than what I type in.

[mort96]

I often also get a complete example from ChatGPT when the question calls for it, it's just usually an incorrect complete example.

[samstave]

>the well of information for AI to train on starts to dry up

and WRT to the eddy-like model-self-incestuation - I am sure that the scope of that well just becomes wider - now its slurping any and all video and learning human micro emotions and micro-aggressions - and mastering human interpersonal skills.

My prediction is that AI will be a top-down reflection of societies' leadership. So as long as we have these questionable leaders throughout the world governments and global corps - the Alignment of AI will be bias to their narratives.

[bobthepanda]

It didn't take very long for the first lawyers to get sanctioned for using ChatGPT-made-up cases in legal briefs. https://www.reuters.com/legal/new-york-lawyers-sanctioned-us...

It would be hilarious if the end result of all this would be to go back to a 1990s-2000s Yahoo style of web portal where all the links are curated by hand by reputable organizations.

[tudorw]

Re-visiting this might be a good idea, it's a different set of tools we have available, perhaps there is something out there that can distribute this task and manage reputation.

[nbates80]

I mean, this was already heading to be the case pre llm.

The internet was already becoming ad farms. This is the final blow and now the internet as we knew it will die.

I’m not that pessimistic about llm generated content. I’m starting to use it to rewrite my online and slack comments for grammar, I’m also using it for brainstorming, enhancing things I create, code (not as in “ok ai write me an app” but as in “change this code to do this, ok this is not considering x and y edge cases, ok use this other method, ok refactor that” it is saving me a lot of typing and silly mistakes while I focus on the meat of the problem.

[taberiand]

They find a way to validate the utility of the information instead of the source.

It doesn't matter if the training data is AI generated or not, if it is useful.

[rurp]

The big problem is that it's orders of magnitude easier to produce plausible looking junk than to solidly verify information. There is a real threat that AI garbage will scale to the point that it completely overwhelms any filtering and essentially ruins many of the best areas of the internet. But hey, at least it will juice the stock price of a few tech companies.

[rs999gti]

> You no longer have to hire a copywriter

Has anyone tested a marketing campaign using copy from a human copywriter versus an AI one?

I would like to see which one converts better.