What if AI ends the utility of the Internet?

[ChpBstrd]

https://finance.yahoo.com/news/ais-mad-cow-disease-problem-tramples-into-earnings-season-100005953.html

TL;DR: As more and more internet content is produced by AIs, the datasets AIs use to train themselves become more and more AI-generated. So you end up with a situation where errors (or virus-like replicators) could be perpetuated and would become hard to eradicate.

One metaphor is how feeding cows the remains of other cows helped mad cow disease spread. Another metaphor is how incest can lead to the multiplication of genetic errors.

In practice, this scenario might take years to unfold. The steps would be:

• AI tools capable of mass producing viral social media content become widely available and usable. The AI content is hard to distinguish from real content.
• Would-be influencers utilize the tech to mass-produce content / clickbait. Using AI to create internet content becomes analogous to bitcoin mining, except you're changing the world's understanding of aspects of reality. AI swarms for hire start creating massive amounts of content about how sponsored brands are the best or targeting elections.
• Social media companies and startups begin mass-producing content, using viral content produced by earlier AIs as training material
• Efforts to identify and exclude AI content generally fail, as AI is uniquely adaptive to defeat such efforts. Meanwhile, AI content quality improves to the point video and audio content become indistinguishable from human-produced content. Neither experts nor other AIs can tell the difference.
• Yottabytes of spammy, clickbaity content are created, overwhelming the volume of all human-produced content and straining the power grid. Search tools cannot identify AI content, so human-produced content becomes less and less of what we consume.
• Eventually, humans stop contributing content to the internet because it is relatively inefficient and such content becomes a needle in an exponentially growing haystack. Earning money from human created content becomes impossible - like hand-making things which can be produced by factories.
• As thousands of successive generations of AI content use previous AI content as training data, the internet becomes detached from reality and even more untrustworthy than it currently is as an information source. Sites like Wikipedia are utterly vandalized by swarms of AIs and become useless. Whatever becomes viral (most common) will become truth to the AIs. This may not look like the internet saying humans have six fingers per hand, but it will look something like this in tens of thousands of domains where viral misinformation gets magnified many-fold.
• A generation of people will be gullible and believe any AI content they consume, just as the first generation to experience any technology is always gullible. It is unclear whether democracy will survive said generation, or if the tech oligarchs will use their power for self-promotion the way Vlad Putin used a takeover of all media to establish his dictatorship.
• The backlash starts either when (1) online content becomes laughably false, or (2) people start paying to subscribe to AI "friends" who have good empathy, but of course eventually steer them to products.

This would be a process occurring over several years, maybe a decade, but the endpoint would look like bad data everywhere (Wikipedia, maps, news, reviews, fake photos, fake videos, fake music, etc) to the point these services would be unusable for anything other than entertainment. At this point the internet would be bifurcated into an entertainment function and a business function. The internet's usefulness as a source for usable information would be... questionable.

So we'd have no choice in this scenario but to find other ways to obtain useful information. E.g. if you look on YouTube for a video of how to fix your car, all the results would be AI-generated with the AI's idea of what your car looks like under the hood. In fact, you'd find a completely different reality when you started working. And your first 10,000 results would be like this. For everything.

To be clear, TikTok-like entertainment would still be around, and businesses would still sell stuff and work online, and there would still be efforts made to identify and stamp out fraud. It's just that the internet would become useless for learning or looking up information because there would be so much self-referential bad info. Stated another way, as human-generated content that is tied at least slightly to reality becomes relatively more rare on the internet, machine-generated content tied to nothing more than itself and its own hallucinations becomes more common.

Contrary arguments:
1) Just like computer viruses and spam, this challenge will be dealt with via thousands of little tweaks to our methods and software. If the internet didn't collapse for these reasons, it won't collapse due to AI.
2) People will accept a certain level of incorrectness in their information if it's cheap and easy to obtain the information. We already see this phenomenon today.
3) The expansion of AI generated content will mostly affect large social media platforms and search engines. It'll have less an effect on blogs, company websites, and anything else where a human holds the keys and can cultivate a place dedicated to information quality. Such brands will become more important than the volume of data. Search engines may go back to the old Yahoo model of manually curated links.

Your thoughts? Will the internet become a (worse) trash pile of bad information?

[reeshau]

The only way AI content will grow to a point that it could affect its environment is if it is more useful than current solutions.  If that usefulness is subjectively iffy, both current and AI-driven methods could continue.  Perhaps they will specialize: use AI when you actually want to create something, like writing a memo for you, or making a movie scene.  But fact checking relies on some certainty of your sources, so that the traditional web crawler.

Or, perhaps there will come along a boring/ bean counter / Poindexter AI that is not creative, but just highly connective, better at synthesizing known facts than making shit up.

The other AI's might pick on him, though.

[Morning Glory]

Have you considered writing fiction? You could easily get a novella out of this premise if you have a character deal with all those issues and maybe learn something. I would read it.

[Luke Warm]

Have you considered writing fiction? You could easily get a novella out of this premise if you have a character deal with all those issues and maybe learn something. I would read it.

+1

[FINate]

The only way AI content will grow to a point that it could affect its environment is if it is more useful than current solutions.  If that usefulness is subjectively iffy, both current and AI-driven methods could continue.  Perhaps they will specialize: use AI when you actually want to create something, like writing a memo for you, or making a movie scene.  But fact checking relies on some certainty of your sources, so that the traditional web crawler.

Or, perhaps there will come along a boring/ bean counter / Poindexter AI that is not creative, but just highly connective, better at synthesizing known facts than making shit up.

The other AI's might pick on him, though.

But how do you define 'useful'? It's not about being useful to the end user, that's not how the internet works. Usefulness is really a measure of monetization and cost to produce. If AI can generate clickable content essentially for free, then it is more profitable then human generated content. Which makes it more useful even if the the quality is far worse.

I hope this happens ASAP. It will eventually kill the "free" ad based internet as we know it. Hopefully replaced by fee based content that is AI assisted, but edited and cross checked by actual humans.

[reeshau]

The only way AI content will grow to a point that it could affect its environment is if it is more useful than current solutions.  If that usefulness is subjectively iffy, both current and AI-driven methods could continue.  Perhaps they will specialize: use AI when you actually want to create something, like writing a memo for you, or making a movie scene.  But fact checking relies on some certainty of your sources, so that the traditional web crawler.

Or, perhaps there will come along a boring/ bean counter / Poindexter AI that is not creative, but just highly connective, better at synthesizing known facts than making shit up.

The other AI's might pick on him, though.

But how do you define 'useful'? It's not about being useful to the end user, that's not how the internet works. Usefulness is really a measure of monetization and cost to produce. If AI can generate clickable content essentially for free, then it is more profitable then human generated content. Which makes it more useful even if the the quality is far worse.

I hope this happens ASAP. It will eventually kill the "free" ad based internet as we know it. Hopefully replaced by fee based content that is AI assisted, but edited and cross checked by actual humans.

I think you can throw social media into that cesspool, yes.  Less facetiously, AI could come to dominate entertainment destinations.  But that's not how Amazon works.  Or even Google.  We will have AI hallucination Maps?  Even basic search needs to inform.  The proposition that AI has that down pat seems shaky to me, until they get some way to guarantee the results as genuine or true.  (Or at least, a truth meter-style rating)  That doesn't appear to be a priority, although the lawyers may have something to say about that.

Maybe it will be somewhat like gaming.  There are any number of "free to play" MMO or app-based games with a predictable pattern of hooking you with easy early leveling, followed by increasing spending on random loot boxes or cosmetics.  They "dominate," if only because there are so many, and they advertise heavily.  But old-style gaming is still very much alive and well, and thriving in its own right.  In fact, the tools that were developed for the MMO crowd have made making games easier for everyone, so there are any number of interesting indie games now, in addition to AAA titles.

Lost in the noise to the uninformed is not the same as going away.

[MustacheAndaHalf]

Years ago book publishers sued Google for showing their copyrighted material.  Much like AI has done, AI crawled the web and surfaced information - and the owners of that information sued.  Same thing has happened with AI, which could limit the content AI crawls.

I would view this more as how lawsuits will shape AI.  For example, if you generate a realistic image of someone else without their permission, will you get away with it while movies can't?

("No identification with actual persons (living or deceased), places, buildings, and products is intended or should be inferred.")
https://en.wikipedia.org/wiki/All_persons_fictitious_disclaimer

[FINate]

The only way AI content will grow to a point that it could affect its environment is if it is more useful than current solutions.  If that usefulness is subjectively iffy, both current and AI-driven methods could continue.  Perhaps they will specialize: use AI when you actually want to create something, like writing a memo for you, or making a movie scene.  But fact checking relies on some certainty of your sources, so that the traditional web crawler.

Or, perhaps there will come along a boring/ bean counter / Poindexter AI that is not creative, but just highly connective, better at synthesizing known facts than making shit up.

The other AI's might pick on him, though.

But how do you define 'useful'? It's not about being useful to the end user, that's not how the internet works. Usefulness is really a measure of monetization and cost to produce. If AI can generate clickable content essentially for free, then it is more profitable then human generated content. Which makes it more useful even if the the quality is far worse.

I hope this happens ASAP. It will eventually kill the "free" ad based internet as we know it. Hopefully replaced by fee based content that is AI assisted, but edited and cross checked by actual humans.

I think you can throw social media into that cesspool, yes.  Less facetiously, AI could come to dominate entertainment destinations.  But that's not how Amazon works.  Or even Google.  We will have AI hallucination Maps?  Even basic search needs to inform.  The proposition that AI has that down pat seems shaky to me, until they get some way to guarantee the results as genuine or true.  (Or at least, a truth meter-style rating)  That doesn't appear to be a priority, although the lawyers may have something to say about that.

Maybe it will be somewhat like gaming.  There are any number of "free to play" MMO or app-based games with a predictable pattern of hooking you with easy early leveling, followed by increasing spending on random loot boxes or cosmetics.  They "dominate," if only because there are so many, and they advertise heavily.  But old-style gaming is still very much alive and well, and thriving in its own right.  In fact, the tools that were developed for the MMO crowd have made making games easier for everyone, so there are any number of interesting indie games now, in addition to AAA titles.

Lost in the noise to the uninformed is not the same as going away.

I don't think the internet itself goes away, but the utility will be greatly diminished. Yes, Amazon will remain as a niche, but this isn't an ad based ecosystem and it's mostly dealing with physical goods. Similarly, subscription based services with curated content will do fine -- they may even do better as the quality of "free" ad based content plunges.

Mapping is an interesting question. It's already the case that Google Maps Directions in my area has gotten progressively worse over the past 1-2 years. Not sure what they're doing, but it seems like they're somewhat randomizing routes perhaps to train AI. DW and I often get in the car and plug in the same destination at the same time and get very different routes. Many of these are just awful, taking us through known problem areas, dangerous left turns onto busy roads, etc. In theory this could improve directions as they gather data, but for now it's much much worse. My overall concern with mapping is that it's expensive to normalize and import public data to keep geo up to date. So it's not hard to imagine Google deciding to use location data from phones fed to an AI algorithm to update this data w/o expensive human intervention... I'm just not confident this will produce reliable information as drivers do all sorts of questionable things like illegal turns and so on. In other words, AI could in fact make the real world experience of maps much worse.

[partgypsy]

You have articulated it much better than I have, but this is something I've been thinking about. Already, the noise to signal ratio is getting that social media and general Internet browsing is less rewarding and useful. Short term, it has not been good because there is now no shared "truth" or agreement. And disinfection from things that to me, actually matter. I don't know what the long term effects will be.

[reeshau]

It somewhat follows the path of cable TV.  When you first went from maybe 3 or 4 channels to 25, it was amazing.  When they started adding channel 599, nobody noticed, but the bill went up.  How many ways do people need to share pictures, or video?  After some saturation point, it's just dividing the pie into more, smaller slices.  And as the slices get smaller, they get worth disproportionately less.

Maybe pay services will rise in the middle of the chaos, like Netflix did for video content.  Both are playing out now, so who knows what the next thing will be?  If AI is simply working on providing the same solutions we have today, just with incremental utility, then it's impact will also just be incremental.

Whatever breakthrough may come, it's unrecognizable to us today.

[Michael in ABQ]

One of my favorite books by Neal Stephenson, Anathem, dealt with this issue to some extent.

It's a science fiction book set in an alternative universe similar to Earth but the dominant religion/philosophy is based around science and so you have monasteries filled with monks studying science but forbidden to use advanced technology (i.e. computers). There are computers and the internet in the outside/secular world, but one character describes it as basically 99.99999% noise due to much bad information.

The way to access good information is restricted to those who can prove they are actually humans and maybe have contributed as well. So, for example, not just scanning your driver's license to verify your identity but people checking security cameras to see if you really are in the city you claim to be, or you have someone personally vouch for you. Or as in this forum, spend literally years contributing to the point that you establish you are a real person with insight into a particular subject.


I could see a future where the valuable information is all gated and restricted where AI can't access it or contaminate it. This already occurs to some extent. Within the ecommerce industry for instance there is a very well-regarded private forum that is restricted to owners of businesses with at least $1M in revenue. There are only about 1,200 members but within that private forum is a bunch of extremely relevant and useful information for people in that industry.

[GilesMM]

Spotting and correcting AI mistakes would be a great use for.... AI!

[RetiredAt63]

It is starting to happen.

Over on Ravelry AI has been collecting information - and people are seeing e-books on Amazon that are obviously AI generated based on that.  When someone who knows the fibre arts sees 12 ebooks on Amazon with almost identical covers and horribly wrong cover images, it is obvious.  The issue is that those who know the skills know those books are fake, but a beginner is not going to know that.

[Log]

The internet still has utility?

[NorCal]

My probably unrealistic hope is that the free content on the internet becomes so bad that we start increasing consumption of higher quality (possibly paid) content. There is value in having trusted sources, and a lot of that has been lost with the expectation of free.

I also hope we quickly establish legal precedent for liability with harmful AI content. Section 230 has protected Google, Facebook, etc from legal liability, as they are not “publishers”. Generative AI probably counts as a publisher, and can face liability for what it says. But this needs to be established in a court.  Having LLM’s be responsible for the consequences of their words would go a LONG ways towards limiting the potential harm AI can cause.

[ChpBstrd]

My probably unrealistic hope is that the free content on the internet becomes so bad that we start increasing consumption of higher quality (possibly paid) content. There is value in having trusted sources, and a lot of that has been lost with the expectation of free.

I also hope we quickly establish legal precedent for liability with harmful AI content. Section 230 has protected Google, Facebook, etc from legal liability, as they are not “publishers”. Generative AI probably counts as a publisher, and can face liability for what it says. But this needs to be established in a court.  Having LLM’s be responsible for the consequences of their words would go a LONG ways towards limiting the potential harm AI can cause.

Excellent point. 230 needs to go before it becomes a cover for AI-for-hire business models. But even then I suspect it will eventually become possible to install havoc-wrecking AI's on web servers anonymously.

[SmashYourSmartPhone]

It accelerates the current trends of people using the internet less and less, trusting it less, etc.

My understanding is that Gen Z's use of the internet has dropped significantly (at least this is what I hear from people who do business social media marketing for a living), and the generation below (Gen A?) is basically not connected at all, in the ways we consider normal.

I work to become more like that, and take long breaks from the internet.  There's just less and less of value on it, and even deeply technical queries for particular error messages and codes often lead to generic "Windows 10 Code 0x1234567A [SOLVED]" style pages that regurgitate pointless crap like "reboot" and "install updated drivers if any exist" and such.  Or the car versions of it.

There is no guarantee that the internet has to remain in the current forms indefinitely, and it almost certainly won't.  People are rejecting that which it comes with - see the rise in vinyl, as people realize that physical media they own is better to yet another streaming service informing them that their "purchased content" has been removed, again.  Who would have believed, growing up in the 90s, that Walmart and Best Buy would be selling new LPs in the mid-2020s?

But that's not how Amazon works.  Or even Google.  We will have AI hallucination Maps?  Even basic search needs to inform.

Does it?  Google has demonstrated they believe they can return utterly nonsensical results for historical queries.  They got laughed at extensively and claim they're modifying Gemini to be less overtly absurd, but "search results returning nonsense" doesn't seem to be anything they're concerned about, as long as they feel they're driving people in the right direction with it.  The important takeaway from the debacle is that they believed it was both perfectly fine to do what they did, and that nobody would care.  I guarantee it is not the only topic they mess with results in.

Amazon simply sells fraudulent and scam products from vendor of the week.  They're Walmart, without the upstanding business ethics and morals.  I despise both.  Yet between the two, I trust Walmart over Amazon, because if I buy a toaster from Walmart, there's a good chance it's actually passed UL testing.  From Amazon, a UL listing guarantees nothing but "some Chinese company told them to put this mark on it, because it needs this mark to sell."  See teardowns of "UL listed" (fake) chargers for how hazardous this sort of device can be.

And we already have AI Hallucination Maps.

Mapping is an interesting question. It's already the case that Google Maps Directions in my area has gotten progressively worse over the past 1-2 years.

In my area, they've started renaming formerly correct road names to be simply wrong.  The road is no longer called what their maps and instructions say it is.  It's annoying, and I've taken to making diagrams before I head out, so I can navigate the final steps even if the road names Google claimed were there are something different.  It's very bad at following roads around corners where the road name changes, and asserting that even around the corner, the road is still the old name.

[ChpBstrd]

Solid points @SmashYourSmartPhone.

It's hard to imagine the internet being a different thing than it was when we grew up, but enshittification already seems to be throwing the business textbooks for a loop. Tech stocks are going UP as enshittification worsens, because they're laying off expensive workers and creating a pretext to persuade people to pay for the quality information services they used to get for free. Maybe an accurate Google Maps is worth $5 per month?

AI might follow a similar trajectory, becoming the Google circa 2010 of usefulness and then a few years later having quality walled off behind a paywall - as happened in journalism. Imagine people in the late 2030s waiting several minutes for an answer to a question, or having to listen to an ad to unlock the answer, which might not be correct. It could happen!

Another mindfuck is the idea that the internet may be so pervasive here in 2024 that it is no longer the growth industry it once was. Client-side hardware, website services, SAAS, and social media are now growing earnings by raising prices and enshittifying rather than by attracting hoards of new users or creating new products. That will eventually deplete demand, as you noted.

In The Jetsons, the (intact nuclear) family would pile around the TV (to watch the same thing) and eat dinner around the dinner table because that's the only way people could envision the future. Such visions look quaint nowadays, but we should pick up on the lesson about things not always being the way they've always been. Things don't always get better. The fast-changing internet would be my last candidate for something that will be the same in 10-20 years.

And yes, the charm has worn off of Amazon just as it did with WalMart.

[RetiredAt63]

AI has a long way to come.  Someone on another forum was watching a video about Queen Elizabeth the first, and realized the image was a woman in Tudor dress with Queen Elizabeth II's face.  ;-)

[SmashYourSmartPhone]

It's hard to imagine the internet being a different thing than it was when we grew up...

No, it's not.  It's very easy, because it's radically different.  The smartphone changed the internet, and not for the better.

https://web.archive.org/web/20240416235606/https://www.bloomberg.com/features/2024-sextortion-teen-suicides/ is just one such read (worth reading if you have teenagers, and certainly worth talking about with them) on what sort of "innovation" has been enabled by internet everywhere, "global communication/global town hall" mediated by large tech companies focused on Engagement Uber Alles.  What's a bunch of teen suicides and wrecked mental health, in the face of these wonderful profits?

Tech stocks are going UP as enshittification worsens, because they're laying off expensive workers and creating a pretext to persuade people to pay for the quality information services they used to get for free. Maybe an accurate Google Maps is worth $5 per month?

Tech stocks are going up.  Gold is going up.  Silver is going up.  Bitcoin is going up.  Groceries are going up.  Energy prices are going up.  Ooooorrrrr...., maybe something else is going down.

The problem is that the tech companies are incredibly adverse to providing services like that for pay - because they have two options.  Either they make far less money per user, or they make it clear to users just how much their data is worth in the vibrant adtech and influence economy that they'd rather people not know about.

In the US, Facebook is making nearly $70/user/year.  And that's an average, so consider how many people don't use Facebook much, or run ad blockers, or such.  In order to allow a heavy user to pay for an ad-free experience, Facebook would probably be charging them north of $200/year - at which point said users may very reasonably ask how Facebook makes so much on them.  Which is a question the tech companies do not want people asking.

AI might follow a similar trajectory, becoming the Google circa 2010 of usefulness and then a few years later having quality walled off behind a paywall - as happened in journalism.

Journalism has suffered mightily at the hand of large tech companies and their lies (see Facebook and "pivot to video" for the worst example), though lately, they simply have done themselves in, because it turns out that when "reporting the facts" turns to "telling people what they ought to think," then a lot of people tune out.

Another mindfuck is the idea that the internet may be so pervasive here in 2024 that it is no longer the growth industry it once was.

This has been a problem for a while.  It's why so many tech companies have "How we're innovating to bring internet to the unserved 3 Billion!" projects, or at least had them.  Because that's a customer base they couldn't otherwise abuse.

Throw in people rejecting it largely, and I think we're going to see a decline in the number of people connected, and the ever-increasing squeezes to get more profit out of those who remain - which will just drive more people off.  And it can't happen too fast.

I very much look forward to the wave of digital consumer tech heading back out to sea.  It's just done a lot of damage in the process we have to rebuild from.

[blue_green_sparks]

Would not surprise me if all this fakery results in sophisticated analysis tools to determine what is real vs what is AI-generated. AI will leave certain detectable markers for now at least. May even raise the market demand for genuine media.

[dcheesi]

I wonder if it could lead to the end of online anonymity?

Between AI fakery and various gov'ts pushing for age verification etc., we could end up in a situation where the only safe spaces for human interaction are in walled gardens w/ mandatory ID verification. Initially just on an ad-hoc voluntary basis, to prove you're human; if the open 'net got bad enough, many people would willingly submit just to gain access to un-junkified content.

But once it caught on, those gov'ts would likely piggyback on the trend to make it official, and make age-verification mandates enforceable. And if it also just so happened to be a boon to law enforcement going after individuals for what they say online, I'm sure that'd be just a happy accident...

[ChpBstrd]

I wonder if it could lead to the end of online anonymity?

Between AI fakery and various gov'ts pushing for age verification etc., we could end up in a situation where the only safe spaces for human interaction are in walled gardens w/ mandatory ID verification. Initially just on an ad-hoc voluntary basis, to prove you're human; if the open 'net got bad enough, many people would willingly submit just to gain access to un-junkified content.

But once it caught on, those gov'ts would likely piggyback on the trend to make it official, and make age-verification mandates enforceable. And if it also just so happened to be a boon to law enforcement going after individuals for what they say online, I'm sure that'd be just a happy accident...

China is already there with their Great Firewall, mass surveillance, and social credit systems.

Americans are arguably begging for such a "product". Experience proves we will expose ourselves to mental health risks, fraud risks, and propaganda from foreign bots in exchange for free "content". TikTok and X are just some of the earliest forms of surveillance capitalism/feudalism.

[Just Joe]

Would not surprise me if all this fakery results in sophisticated analysis tools to determine what is real vs what is AI-generated. AI will leave certain detectable markers for now at least. May even raise the market demand for genuine media.

https://www.popsci.com/technology/ai-running-for-office/

[LennStar]

AI can't end the utility of the internet, because the utility of the internet is not endless text generation for selling products.

At least it used to be.

I am pretty sure this forum will not be filled by useless AI halluzinations. Spam generated by AI maybe, but that has nothing to do with AI, since you can do that with AI as easily.

[innkeeper77]

This is already starting- model collapse is a real issue, Amazon is starting to have some categories filled with AI junk, and many research papers are exhibiting the signs of being GPT generated. To my perspective, this is a continuation of the decline of the internet. The "post forum" era..

Generative AI can never become a general intelligence, it just isn't how it works. It can also never get rid of the problem of "hallucinations"- it is a technical limitation of operating purely on mathematical probabilities rather than real datasets.

I would suggest you look into Ed Zitron's podcast, better offline. It is a podcast that dives into the general rot in the tech industry overall, with a lot on AI because AI is the current bubble we are experiencing. https://linktr.ee/betteroffline

[reeshau]

Generative AI can never become a general intelligence, it just isn't how it works. It can also never get rid of the problem of "hallucinations"- it is a technical limitation of operating purely on mathematical probabilities rather than real datasets.

I agree with this point. AI based on large language models is an attempt to solve human interaction in the same way supercomputers attempted to beat chess grandmasters; by computing all possible combinations.  But John von Neuman said about chess:

"Chess is not a game. Chess is a well-defined form of computation. You may not be able to work out the answers, but in theory there must be a solution, a right procedure in any position. Now real games are not like that at all. Real life is not like that. Real life consists of bluffing, of little tactics of deception, of asking yourself what is the other man going to think I mean to do."

Real life isn't strictly a game, but it does include elements of intention and limits of information.  So, emulating intelligence through total knowledge is not strictly possible, because a good deal is unknown and unknowable.  Playing at it can work for well-covered sections with lots of examples.  But when you use The Onion as a data source, without somehow communicating to your model that it is entirely false or ironic, just polluted your model, and degrades your capability.

[Ron Scott]

The internet has already morphed into a big marketing environment and group-think echo chambers that give you more of the same and where like-minded souls gather to tell each other how cool and good looking they are. In this way both the monetization and the politicization of information has already been accomplished. I suspect AI tools will be used for ulterior motives to perfect the approach. We are all entitled, it seems, to our own opinions AND our own facts. Pity the stupid.

That aside, I am already getting tired of strangely guard-railed chat-buddy tools like Gemini, ChatGPT, Claude, and Perplexity and am hoping the evolution of AI moves more quickly into personal agents, robotics/world-models, and domain-specific expertise in areas like health, law, etc. where accuracy is outcome oriented and less a matter of opinion or something subversive. I also have interest in tools that allow me to express creativity in domains I’ve struggled to master due to basic technical/technique incompetence, like art and music. I actually think Apple’s play is like this and is one of the more interesting approaches out there today. And this is where I imagine the business cloud providers are headed. Give us something useful, that makes us more productive.

AI is bigger than the internet.

[LennStar]

I would certainly like it if I could use AI to augment my "paintings" e.g. of characters of the stories I write. Currently it looks like the scratches of a 6 year old, with the difference that I put double the time in it.

Or maps. Imagine I would just put in a certain color and tell the AI "make that mountains with 2 "here be dragons" paintings."

[Ron Scott]

Arts is certainly an area where AI could help us. I would think of it more as a grand tool than a creator and publisher but I suppose it could go both ways. But the concept of getting what’s in my head into the physical world—alleviating failures in developing physical technique—is exciting to me.I think the effort in learning techniques (which is the least interesting part of the process) hamstrings many would-be artists from expressing themselves.

That and the utility of AI to accelerate and expand science and engineering are where the value is IMO.

If AI eats the internet so be it. Probably will. But that’s besides the point for AI.

Oh, and when quantum computing comes online and marries AI? Sky’s the limit.

[reeshau]

This thought came across my LinkedIn feed recently.  I may have already posted it elsewhere.

[Ron Scott]

I suppose we’d have to see her art, writing, laundry, and dishes to opine here.

[Ron Scott]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

[2sk22]

One of my favorite books by Neal Stephenson, Anathem, dealt with this issue to some extent.

It's a science fiction book set in an alternative universe similar to Earth but the dominant religion/philosophy is based around science and so you have monasteries filled with monks studying science but forbidden to use advanced technology (i.e. computers). There are computers and the internet in the outside/secular world, but one character describes it as basically 99.99999% noise due to much bad information.

The way to access good information is restricted to those who can prove they are actually humans and maybe have contributed as well. So, for example, not just scanning your driver's license to verify your identity but people checking security cameras to see if you really are in the city you claim to be, or you have someone personally vouch for you. Or as in this forum, spend literally years contributing to the point that you establish you are a real person with insight into a particular subject.


I could see a future where the valuable information is all gated and restricted where AI can't access it or contaminate it. This already occurs to some extent. Within the ecommerce industry for instance there is a very well-regarded private forum that is restricted to owners of businesses with at least $1M in revenue. There are only about 1,200 members but within that private forum is a bunch of extremely relevant and useful information for people in that industry.

Delighted to find another Neal Stephenson fan here on this forum - I have read Anathem at least a dozen times. What's amazing about this book is there are many tangential passages that deserve to be expanded into a separate book on its own! Also note that this book was written 2006-2008, long before the current batch of LLMs became important. Back then, I was still using classical statistical linguistic methods to built text classifiers :-)

In any case, here is the bit from the book about crap in the Reticulum (which is the name for the Internet in that world) that you were referring to. I think Stephenson was absolutely prescient on this.

[reeshau]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

This is actually very similar to training autonomous vehicles.  An autonomous car generates 5 TB of data per day; even 5G wireless could not handle a small fleet transmitting everything.  A large part of the on-board compute is to determine what is "interesting" (novel) and only send that back to the mother ship.  Long sections of straight road with no other vehicles are essentially garbage, in terms of learning.

[NorCal]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

This is actually very similar to training autonomous vehicles.  An autonomous car generates 5 TB of data per day; even 5G wireless could not handle a small fleet transmitting everything.  A large part of the on-board compute is to determine what is "interesting" (novel) and only send that back to the mother ship.  Long sections of straight road with no other vehicles are essentially garbage, in terms of learning.

I would counter that what we call “high quality” inputs like research journals and scientific papers probably count as garbage data for those seeking to monetize the outputs of LLM’s.

[reeshau]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

This is actually very similar to training autonomous vehicles.  An autonomous car generates 5 TB of data per day; even 5G wireless could not handle a small fleet transmitting everything.  A large part of the on-board compute is to determine what is "interesting" (novel) and only send that back to the mother ship.  Long sections of straight road with no other vehicles are essentially garbage, in terms of learning.

I would counter that what we call “high quality” inputs like research journals and scientific papers probably count as garbage data for those seeking to monetize the outputs of LLM’s.

Depends if you are trying to build a chat bot, or discover the next cure for cancer.

[NorCal]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

This is actually very similar to training autonomous vehicles.  An autonomous car generates 5 TB of data per day; even 5G wireless could not handle a small fleet transmitting everything.  A large part of the on-board compute is to determine what is "interesting" (novel) and only send that back to the mother ship.  Long sections of straight road with no other vehicles are essentially garbage, in terms of learning.

I would counter that what we call “high quality” inputs like research journals and scientific papers probably count as garbage data for those seeking to monetize the outputs of LLM’s.

Depends if you are trying to build a chat bot, or discover the next cure for cancer.

I expect the monetizatation of both those paths will be dystopian in their own unique way.

I’ve become a little cynical about Silicon Valley if you can’t tell.

[LennStar]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

The current stance is that the internet as a whole does not have enough data. I can't see how 0,01% of it could be enough.

[reeshau]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

The current stance is that the internet as a whole does not have enough data. I can't see how 0,01% of it could be enough.

I think the story of Google Gemini recommending non-toxic glue to keep cheese on a pizza shows the flaw.  Gemini naively consumed content from The Onion as sincere information, rather than satire.  The models need some way to understand intent, to be able to prioritize and filter what they include in their model.  Otherwise, this content was pollution, damaging its capability.

[LennStar]

Leopold Achenbrenner takes a different view of the internet and its breadth in his paper Situation Awareness.

A quick excerp, where he is commenting on the need for AI developers to stop breezing through the internet and deep-dive on the important stuff:

“Current frontier models like Llama 3 are trained on the internet-and the internet is mostly crap, like e-commerce or SEO or whatever. Many LLMs spend the vast majority of their training compute on this crap, rather than on really high-quality data (e.g. reasoning chains of people working through difficult science problems). Imagine if you could spend GPT-4-level compute on entirely extremely high-quality data—it could be a much, much more capable model.”

The current stance is that the internet as a whole does not have enough data. I can't see how 0,01% of it could be enough.

I think the story of Google Gemini recommending non-toxic glue to keep cheese on a pizza shows the flaw.  Gemini naively consumed content from The Onion as sincere information, rather than satire.  The models need some way to understand intent, to be able to prioritize and filter what they include in their model.  Otherwise, this content was pollution, damaging its capability.

Looking at how many humans seem to be unable to differentiate between sincere inofrmation, satire, propaganda and hilariously easy to see lies makes me not optimistic that we will be able to make a AI do this, not to mention a LLM.
Though if it happens I might be convinved that it is indeed an artificial intelligence. 

[ChpBstrd]

One of my favorite books by Neal Stephenson, Anathem, dealt with this issue to some extent.

It's a science fiction book set in an alternative universe similar to Earth but the dominant religion/philosophy is based around science and so you have monasteries filled with monks studying science but forbidden to use advanced technology (i.e. computers). There are computers and the internet in the outside/secular world, but one character describes it as basically 99.99999% noise due to much bad information.

The way to access good information is restricted to those who can prove they are actually humans and maybe have contributed as well. So, for example, not just scanning your driver's license to verify your identity but people checking security cameras to see if you really are in the city you claim to be, or you have someone personally vouch for you. Or as in this forum, spend literally years contributing to the point that you establish you are a real person with insight into a particular subject.


I could see a future where the valuable information is all gated and restricted where AI can't access it or contaminate it. This already occurs to some extent. Within the ecommerce industry for instance there is a very well-regarded private forum that is restricted to owners of businesses with at least $1M in revenue. There are only about 1,200 members but within that private forum is a bunch of extremely relevant and useful information for people in that industry.

Delighted to find another Neal Stephenson fan here on this forum - I have read Anathem at least a dozen times. What's amazing about this book is there are many tangential passages that deserve to be expanded into a separate book on its own! Also note that this book was written 2006-2008, long before the current batch of LLMs became important. Back then, I was still using classical statistical linguistic methods to built text classifiers :-)

In any case, here is the bit from the book about crap in the Reticulum (which is the name for the Internet in that world) that you were referring to. I think Stephenson was absolutely prescient on this.

Thank you for the reference! It's an interesting thought that the over-production of garbage information has something in common with enshittification. In both cases, monopolists make their products or the world worse so that you'll pay them to clean up the mess they created. It's like a polluter selling gas masks.

There are less-conspiratorial ways to explain the phenomenon, such as the psychological biases of users and the tendency for the internet's heaviest users to also be the least rational or epistemically disciplined, so I'll keep this one in the back pocket. In the meantime, I burned an Audible credit on Anathem.

[2sk22]

Thank you for the reference! It's an interesting thought that the over-production of garbage information has something in common with enshittification. In both cases, monopolists make their products or the world worse so that you'll pay them to clean up the mess they created. It's like a polluter selling gas masks.

There are less-conspiratorial ways to explain the phenomenon, such as the psychological biases of users and the tendency for the internet's heaviest users to also be the least rational or epistemically disciplined, so I'll keep this one in the back pocket. In the meantime, I burned an Audible credit on Anathem.

I am really worried about the pollution of knowledge. Consider even arXiv for example - I heard about someone using arXiv to submit bogus papers in support of some fake technology as a part of a pump-and-dump scheme. There is another passage in anathema where Stephenson mentions an interesting idea for a way to consider the reputation of the source of information - I don't want to give away too much of the plot :-)

Also, while Anathem will be great as an audio book, you really need to read it in paper form (not even as an ebook) to get the full experience in my opinion.

[NorCal]

One of my favorite books by Neal Stephenson, Anathem, dealt with this issue to some extent.

It's a science fiction book set in an alternative universe similar to Earth but the dominant religion/philosophy is based around science and so you have monasteries filled with monks studying science but forbidden to use advanced technology (i.e. computers). There are computers and the internet in the outside/secular world, but one character describes it as basically 99.99999% noise due to much bad information.

The way to access good information is restricted to those who can prove they are actually humans and maybe have contributed as well. So, for example, not just scanning your driver's license to verify your identity but people checking security cameras to see if you really are in the city you claim to be, or you have someone personally vouch for you. Or as in this forum, spend literally years contributing to the point that you establish you are a real person with insight into a particular subject.


I could see a future where the valuable information is all gated and restricted where AI can't access it or contaminate it. This already occurs to some extent. Within the ecommerce industry for instance there is a very well-regarded private forum that is restricted to owners of businesses with at least $1M in revenue. There are only about 1,200 members but within that private forum is a bunch of extremely relevant and useful information for people in that industry.

Delighted to find another Neal Stephenson fan here on this forum - I have read Anathem at least a dozen times. What's amazing about this book is there are many tangential passages that deserve to be expanded into a separate book on its own! Also note that this book was written 2006-2008, long before the current batch of LLMs became important. Back then, I was still using classical statistical linguistic methods to built text classifiers :-)

In any case, here is the bit from the book about crap in the Reticulum (which is the name for the Internet in that world) that you were referring to. I think Stephenson was absolutely prescient on this.

Thank you for the reference! It's an interesting thought that the over-production of garbage information has something in common with enshittification. In both cases, monopolists make their products or the world worse so that you'll pay them to clean up the mess they created. It's like a polluter selling gas masks.

There are less-conspiratorial ways to explain the phenomenon, such as the psychological biases of users and the tendency for the internet's heaviest users to also be the least rational or epistemically disciplined, so I'll keep this one in the back pocket. In the meantime, I burned an Audible credit on Anathem.

I've been around corporate america long enough to know that this type of thing usually isn't malicious intentions.  Just a malicious indifference to externalities. 

I did add Anathem to my reading list.  It sounds like the type of book I'd enjoy.

[Ron Scott]

I would find it depressing if all we got out of AI was an internet-level experience.

[MrGreen]

It actually makes sense that this would happen when you consider that we have enshittified (?) the utility of major pieces of the internet. AI would have the ability to enshittify the internet as a whole, writing articles, making fake videos, etc. If past pattern is repeated, AI will start out as a useful tool but will eventually destroy the utility of the thing it was supposed to improve in the name of the almighty dollar.

[ChpBstrd]

Perhaps printing presses will one day be as sought-after as vinyl record presses are today.

[Ron Scott]

I think the concern about AI ending the utility of the internet will go the way of the concerns of those in the early 20th Century that the automobile would end the utility of horses.

Those who think progress in AI will ebb because “they’re running out of internet data” need to consider what Aschenbrenner and most other AI lab leaders are saying about adding efficient compute, dramatically improving algorithms, and deep diving on important documents vs. speed-reading internet data dumps. All that is happening at the same time as Nvidia, Deep Mind et al are developing world models in which AIs and robots explore more naturally on their own and learn through exploration of both simulated physics and the real world. Bottom line: AI certainly does not depend on an infinite, high quality.internet.

I mean—just consider Alpha Go. It trained on people playing games and got pretty good. Then it learned by playing a million games against itself and became the champion of the world.  It MOVED BEYOND human supplied training datasets…

AI is still nursing. Stay tuned.

[LennStar]

dramatically improving algorithms,

How can you do that if you don't know what algos are used?
It's one of the major concerns. Nobody knows how/why e.g. ChatGPT came to that halluzination.

[ChpBstrd]

I think the concern about AI ending the utility of the internet will go the way of the concerns of those in the early 20th Century that the automobile would end the utility of horses.

Those who think progress in AI will ebb because “they’re running out of internet data” need to consider what Aschenbrenner and most other AI lab leaders are saying about adding efficient compute, dramatically improving algorithms, and deep diving on important documents vs. speed-reading internet data dumps. All that is happening at the same time as Nvidia, Deep Mind et al are developing world models in which AIs and robots explore more naturally on their own and learn through exploration of both simulated physics and the real world. Bottom line: AI certainly does not depend on an infinite, high quality.internet.

I mean—just consider Alpha Go. It trained on people playing games and got pretty good. Then it learned by playing a million games against itself and became the champion of the world.  It MOVED BEYOND human supplied training datasets…

AI is still nursing. Stay tuned.

The concern is NOT that there won't be enough data or bandwidth. The concern is that there will be too much low-quality or incorrect data that will increase the frequency of AI's output being at least partially false. Adding computational speed or making the dataset larger does not address this problem.

There are millions of ways for AI output to be incorrect, and only one way to be correct. Combine this factor with the observation that AIs will be producing most of the data on the internet that become their own training datasets, and you get a high likelihood of incorrect data multiplying within the dataset, which will perpetuate the production of more incorrect data.

This is how the world works in meatspace, with false rumors, superstitions, or propaganda flooding out the much more scarce high-quality journalism, high-quality science, or accurate political information. For specific reasons, misinformation is more likely to be produced or spread in certain domains. In meatspace, we promote "critical thinking" or a "scientific worldview" as ways to filter out unsupported ideas. I'm most interested if anyone has information about AI's being trained to use these sorts of solutions.