So google will eventually be mostly indexing the output of LLMs, and at that point they might as well skip the middleman and generate all search results by themselves, which incidentally, this is how I am using Kagi today - I basically ask questions and get the answers, and I barely click any links anymore.
But this also means that because we've exhausted the human generated content by now as means of training LLMs, new models will start getting trained with mostly the output of other LLMs, again because the web (as well as books and everything else) will be more and more LLM-generated. This will end up with very interesting results --not good, just interesting-- akin to how the message changes when kids the telephone game.
So the snapshot of the web as it was in 2023 will be the last time we had original content, as soon we will have stop producing new content and just recycling existing content.
> Google any recipe, and there are at least 5 paragraphs (usually a lot more) of copy that no one will ever read, and isn't even meant for human consumption. Google "How to learn x", and you'll usually get copy written by people who know nothing about the subject, and maybe browsed Amazon for 30 minutes as research. Real, useful results that used to be the norm for Google are becoming more and more rare as time goes by.
> We're bombarding ourselves with walls of human-unreadable English that we're supposed to ignore. It's like something from a stupid old sci-fi story.
When I read comments today I wonder if there is a human being that wrote them or an LLM.
That, to me, is the biggest difference. Previously I was mostly sure that something I read couldn’t have been generated by a computer. Now I’m fairly certain that I would be fooled quite frequently.
Mm. To me, I think ChatGPT has a certain voice, not sure about the other LLMs.
But perhaps I'm wrong. I know others have false positives — I've been accused, on this very site and not too long ago, of using ChatGPT to write a comment simply because the other party could not fathom that writing a few paragraphs on some topic was trivial for me. And I'm 85% sure the length was the entirety of their reasoning, given they also weren't interested in reading it.
>You’re definitely right about that. CharGPT is almost too accurate/structured.
I think a lot of the material was from standardized testing.
This very structured writing style. Many paragraphs, each discussing one aspect, finished by a conclusion. This is the classic style taught for (American at least) standardized testing, be it SAT, GRE, TOEFL, et al.
Was going to post something similar. There may be a need for a way to confirm ( not detect, which is its own field ) organic content. I hate the thought, because I assume I know where that goes privacy-wise.
Mm. To me, I think ChatGPT has a certain voice, not sure about the other LLMs
How long will it be, before humans reading mostly LLM output, adopt that same writing style? Certainly, for people growing up today, they will be affected.
I remember an HN comment six months or so ago by someone who said they were intentionally modeling their writing on ChatGPT's style. The person said that they were not confident about writing and that they were trying to get better by imitating AI.
One of the many surprising things to me about ChatGPT when it was first released was how well, in its default style, it imitated the bland but well-organized writing style of high school composition textbooks: a clearly stated thesis at the beginning, a topic sentence for each paragraph, a concluding paragraph that often begins "In conclusion."
I mentioned that last point—the concluding "In conclusion"—as an indicator of AI writing to a university class I taught last semester, and a student from Sweden said that he had been taught in school to use that phrase when writing in English.
If I see HN comments that have final paragraphs beginning with "In conclusion" I will still suspect that an LLM has been used. Occasionally I might be wrong, though.
I was taught in high school that using "In conclusion" to open your conclusion was cliche and really almost like an unnecessary slap in the face to the reader. Your composition should end with a conclusion, yes. There was a standard formula for that, yes. But it's not necessary to literally label it as such.
Many of the disliked essay writing cliches are good speech tropes. The difference between reading and listening is that in reading you can skim and skip and rewind, so you don't need structured signposts to guide you through the content. In listening you do. You can't see the last paragraph coming when listening to a speech.
An entertaining informative style of speech can detract from clearly communicating substance. (Of course, the audience rarely wants substance.)
I've intentionally changed some parts of I've comments I've written just because upon reading them back, it felt very close to ChatGPT's style at certain sentences.
I understand. A few months ago, I posted a comment here that attracted several down votes. The content, I thought, was completely innocuous, and I couldn’t figure out at first why some people didn’t like it. Only later did I realize that I might have polished it a little too much and it came out reading like ChatGPT.
How long will it be, before humans reading mostly LLM output, adopt that same writing style?
From what I’ve seen (tutoring high school kids), the picture is much bleaker. They use ChatGPT to write for them but they have no writing style of their own. They can barely put a sentence together just to write the prompt!
As humans, we like to copy things we consider good. I know I tried to use writing styles I thought were funny with varying results. I think I will say never ( or as close to never as possible ).
Given how inhumanly capable it is of seeing both sides of a topic, I guess we'll find out when we see "sublety" and "nuance" go the way way of "literally" and radically change meanings to their near opposite.
> How long will it be, before humans reading mostly LLM output, adopt that same writing style? Certainly, for people growing up today, they will be affected.
Just imagine 180M users of chatGPT having an estimated 1B sessions per month. The model is putting 1-2Trillion tokens into people's brains. People don't assimilate just the writing style and ideas, but also take actions into the real world influenced by the model. Sometimes they create useful discoveries or inventions that end up on the internet and in the next scrape. Full cycle.
That’s the first output from ChatGPT. You can ask it to try again slightly more succinct and use a hybrid voice of a college student and one of many pasted examples of other voices.
When I fed that prompt to my Claude 3.5 based CarlinBot, it refused to write a poem and instead wrote a rant about why it hates tangerines (warning, profanity):
Fuck tangerines, man. Those little orange bastards are a pain in the ass to peel. You spend 10 minutes trying to get that leathery skin off, your fingernails getting all sticky with that goddamn citrus juice. And then when you finally wrestle one of those fuckers open, you got all those little white strings hanging off everywhere. It's like dental floss from hell.
And don't even get me started on how those tangerine slices always shoot juice in your eye when you try to eat them. It's like getting maced by a tiny orange asshole. You ever get tangerine juice in your eye? Shit burns like the fires of hell itself. Makes you want to claw your own goddamn eyeballs out.
Nah, fuck tangerines and their whole stupid family tree. Oranges, clementines, satsumas - they can all go straight to fruit hell as far as I'm concerned. Give me a nice safe banana or an apple any day over those sadistic little citrus grenades. Tangerines are the work of the devil, plain and simple. Fuck writing poems about them little torture devices.
On what does a tangerine wait
Each morning below the winters sun as it awakes?
Do twisted dreams linger, of what it might mean to be a taste on the memory of a forgotten alien tongue?
Is its sacred role seen -- illuminated amongst the greens and unique chaotic chrominance bouncing ancient wisdom between the neighboring leaves?
The tangerine -- victim, pawn, and, ultimately, master ; its search for self in an infinitely growing pile of mixed up words truly complete. There is much to learn.
I was listening to a podcast/article being read in the authors' voice and it took me an embarrassingly long time to realize it was being read by an AI. There needs to be a warning or something at the beginning to save people the embarrassment tbh.
I think it will eventually be good public policy to make it illegal to post massive amounts of texts produced by AI without disclosing it.
As with all illegal things on the internet, it's difficult to enforce, but at least it will make it more difficult/less likely
How about articles written by human charlatans? Claiming they are 'doctors' or 'scientists'. Or posters claiming something that didn't happen? Like a... pro bullshtter claiming he was denied apartment renting because of his skin color. He could make a lot of money if that was true. But poster is still taking ads place, payed by poor 'suffering' minority. Another example 'influencers' who pretending, or really being, experts advise you on forums about products. The tell mostly the truth, but avoid some negative details and competing products and solutions. Without disclosing their connections to businesses.
Shorter version: intentional bullshtting never ends, it's in human, and AI, nature. Like it or not. Having several sources used to help, but now with flood of generated content it may be not the case anymore. If used right this has real affect on business. That's how small sellers live and die on Amazon.
Sure, but for me there isn't anything fundamentally different between a LLM reply and a spammers reply / SEO-vomit. Both are low quality useless junk that gives the masquerade of resembling something worth engaging with.
In fact the really bad spammers were already re-using prompts/templates, think of how many of those recipe novellas shared the same beats. "It was my favorite childhood comfort food", "Cooked with my grandma", blah blah blah
Really? People want to have discussions with other people. I don’t want the output of aggregate data that some tech company worth billions (or the wannabes) might offer. It is truly weird that this needs to be said.
I don’t want this to come across as too negative of a sentiment, but (…) a lot of online discussions are just people repeating opinions they heard elsewhere they agree with. AI is, in this regard, not that different. And marketing is a big part of it, so there are already companies with lots of weight behind making sure that people talk about only certain topics with certain viewpoints (i.e. the Overton window).
Actually original commentary in a discussion is bloody hard to come by.
Sure but the output of an LLM is _never_ original.
Human output signal might be wildly different from person to person if judged on originality. But LLM output is then pure noise. The internet wad already a noisy place but humans are “rate limited” to a degree an LLM is not.
OP is pretty on point. While internet is full of SEO junk, it was far more prevalent back in 2010-2014-5, where the main SEO strategy was to dump 500 words articles in web directories.
The difference is that back then there was an effort from companies like Google to fight the spam and low quality content. Everyone was waiting Matt Cutts( back then head of web spam and search quality at Google) to drop a new update so they can figure out how to step up their game. So at one point you could't afford to just spam your domain with low quality content because you would be penalised, and dropped from the search engines.
There is nothing like that today everybody is on the bandwagon of AI, somehow chatting with pdf documents is now considered by the tech bro hype circle as a sign of enlightenment a beginning of a spark of intelligence...
this is mainly to prolong time on site / impressions that can be served. of course 98% of the banners on those pages are served by doubleclick (google) and thus google makes more money, the crappier the page.
> A recipe is a statement of the ingredients and procedure required for making a dish of food. A mere
listing of ingredients or contents, or a simple set of directions, is uncopyrightable. As a result, the Office cannot register recipes consisting of a set of ingredients and a process for preparing a dish. In contrast, a recipe that creatively explains or depicts how or why to perform a particular activity may be copyrightable. A registration for a recipe may cover the written description or explanation of a process that appears in the work, as well as any photographs or illustrations that are owned by the applicant. However, the registration will not cover the list of ingredients that appear in each recipe, the underlying process for making the dish, or the resulting dish itself. The registration will also not cover the activities described in the work that are procedures, processes, or methods of operation, which are not subject to copyright protection.
Recipes were an easy way to avoid some copyright claims. Copy the list of ingredients, and write a paragraph about how your grandmother made it from a secret recipe that turned out to be on the back of the box.
----
I can still think of content farms and the 2010s and the sheer bulk of junk they produced.
> The former “content creator” — that’s what Demand CEO Richard Rosenblatt calls his freelance contributors — asked to be identified only as a working journalist for fear of “embarrassing” her current employer with her content farm-hand past. She began working for Demand in 2008, a year after graduating with honors from a prestigious journalism program. It was simply a way for her to make some easy money. In addition to working as a barista and freelance journalist, she wrote two or three posts a week for Demand on “anything that I could remotely punch out quickly.”
> The articles she wrote — all of which were selected from an algorithmically generated list — included How to Wear a Sweater Vest” and How to Massage a Dog That Is Emotionally Stressed,” even though she would never willingly don a sweater vest and has never owned a dog.
> “Never trust anything you read on eHow.com,” she said, referring to one of Demand Media’s high-traffic websites, on which most of her clips appeared.
Be VERY careful using Kagi this way -- I ended up turning off Kagi's AI features after it gave me some comically false information based on it misunderstanding the search results it based its answer on. It was almost funny -- I looked at its citations, and the citations said the opposite of what Kagi said, when the citations were even at all relevant.
May I ask how you know those 999 answers were correct, and how would you have been sure to catch a mistake, misinterpretation or hallucination in any of those?
It's not only Kagi AI but Kagi Search itself has been failing me a lot lately. I don't know what they are trying to do but the amount of queries that find zero results is impressive. I've submitted many search improvement reports in their feedback website.
Usually doing `g $query` right after gives me at least some useful results (even when using double quotes, which aren't guaranteed to work always).
Happens about 200 times a day (0.04% of queries), very painful for the user we know, still trying to find root cause (we have limited debugging capabilities as not storing much information). it is on top of our minds.
Yeah, that's totally fair. I just think about all the people to whom I've had to explain LLM hallucinations, and the surprise in their faces, and this feature gives me some heebie-jeebies
Eventually the only purpose of AI as is the only purpose of computers is to enhance human creativity and productivity.
Isn't an LLM just a form of compressing and retrieving vast amounts of information? Is there anything more to it than that?
Don't think LLM itself will ever be able to out compete competent human + LLM. What you will see is that most humans are bad at writing books so they will use LLM and you will get mediocre books. Then there will expert humans that use LLM and are experts to create really good books. Pretty much what we see now. Difference is future you will a lot more mediocre everything. Even worse than it is now. I.e, if you look at Netflix there movies all mediocre. Good movies are the 1% that get released. With AI we'll just have 10 Netflix.
> Don't think LLM itself will ever be able to out compete competent human + LLM
Perhaps, perhaps not. The best performing chess AI, are not improved by having a human team up with them. The best performing Go AI, not yet.
LLMs are the new hotness in a fast-moving field, and LLMs may well get replaced next year by something that can't reasonably be described with those initials. But if they don't, then how far can the current Transformer style stuff go? They're already on-par with university students in many subjects just by themselves, which is something I have to keep repeating because I've still not properly internalised it. I don't know their upper limits, and I don't think anyone really does.
Oh man. Want to know an LLM's limits? Try discussing a new language feature you want to build for an established language. Even more fun is trying to discuss a language feature that doesn't exist yet, even after you provide relevant documentation and examples. It cannot do it. It gets stuck in a rut because the "right" answer is no longer statistically significant. It will get stuck in a local min/max that it cannot easily escape from.
This is a limit of an LLM's architecture. It is based on statistics and can only answer statistical questions. If you want it to provide non-probable answers, an LLM won't work.
Your brain is also based on statistics. We also get stuck in a rut because the "right" answer is no longer statistically significant.
And yet this is not what limits our cognition.
Current LLMs are slow to update with new info, which is why they have cut-off dates so far in the past. Can that be improved to learn as fast (from as little data) as we do? Where's the optimal point on inferring from decreasing data before they show the same cognitive biases we do?
(Should they be improved, or would doing that simply bring in the same race dynamics as SEO?)
Even humans are not good at this. The US military has a test (DLAB) to figure out how good you are at taking in new information in regards to language -- to determine if it is worth teaching you new languages. Some humans are pretty good at this type of thing, but not all. Some humans can't even wrap their heads around algebra but will sell you a vacuum cleaner before you even realize you bought it.
The problem with LLMs is that there is one and it is always the same. Sure, you can get different ones and train your own, to a degree.
> They're already on-par with university students in many subjects just by themselves, which is something I have to keep repeating because I've still not properly internalised it.
That’s because it’s not really true. There are glimpses of this but it trips up too often.
This is a weird take. The paren comment said that, the Internet will not be the same with LLM generated slop.
You're differentiating between LLM generated content and LLM + human combination.
Both will happen, with dire effects to the internet as a whole.
Yeah, but the layout of singular value decomposition and similar algorithms and how pages rank among it is changing all the time. So, par for course. If aspect become less useful people move on. Things evolve, this is a good thing
My experience is that AI tends to surface original content on the web that, in search engines, remains hidden and inaccessible behind a wall of SEOd, monetized, low-value middlemen. The AI I've been using (Perplexity) thumbnails the content and provides a link if I want the source.
The web will be different, and I don't count SEO out yet, but... maybe we'll like AI as a middleman better than what's on the web now.
> So the snapshot of the web as it was in 2023 will be the last time we had original content, as soon we will have stop producing new content and just recycling existing content.
I’ve seen this take before and I genuinely don’t understand it. Plenty of people create content online for the simple reason they enjoy doing it.
They don’t do it for the traffic. They don’t do it for the money. Why should they stop now? Is not like AI is taking away anything from them.
You ask an LLM to do it. Not sarcasm, they’re quite good at ranking the quality of content already and you could certainly fine tune one to be very good at it. You also don’t need to filter out all of the machine written content, only the low quality and redundant samples. You have to do this anyways with human generated writing.
I just tried asking ChatGPT to rate various BBC and NYT articles out of 10, and it consistently gave all of them a 7 or 8. Then I tried today's featured Wikipedia article, which got a 7, which it revised to an 8 after regenerating the respose. Then I tried the same but with BuzzFeeds hilariously shallow AI-generated travel articles[1] and it also gave those 7 or 8 every time. Then I asked ChatGPT to write a review of the iPhone 20, fed it back, and it gave itself a 7.5 out of 10.
I personally give this experiment a 7, maybe 8 out of 10.
ChatGPT has a giant system prompt that you have no control over. Try using Llama and create a system prompt with clear instructions and examples. If you were going to use a model in a production system you would also want to either fine tune it or train a BERT-like model as a classifier that just outputs a score. Maybe even more than one for ranking along different dimensions.
Except AI in search is taking away significant traffic from everywhere, and it hits small blogs as well as nonprofits like encyclopaedias the hardest, while misrepresenting and “remixing” the actual content.
I’ve given up on the internet as a place to share my passions and hobbies for the most part, and while LLM’s weren’t the only reason, this current trend is a significant factor. I focus most of my attention on talking directly with people. And yes that does mean the information I share is guaranteed to be lost to time, but I’d rather it be shared in a meaningful manner in the moment than live on in an interpreted zombie form in perpetuity.
I have a blog. Been writing on that for 7 years. Should I care if AI in search is taking away traffic? If yes, why? I’m not writing for traffic. I write because I enjoy doing it. People find my way mostly thanks to other people linking to my site. And a solid % of traffic comes from RSS anyway.
I think giving up on the web because of AI is the wrong move. You should still create and focus more on connecting with others directly, when online. Get in touch, write emails, sign guestbooks.
I’m personally having great exchanges daily with people from all over via email and that won’t stop because of stupid ChatGPT or whatever.
And don’t get me wrong, it’s awesome to spend more time offline so if you want to do down that path it’s great.
The only reason to put things you write online is to make it available to others. If writing simply for my own enjoyment or reference I write in my notebooks, as I do all the time. I never stopped doing that.
A lot of people who create content don't want their content to feed AI. They love what they do and they don't want their work to support a system whose purpose is to debase and commoditize that work. The only way to avoid that is to never publish to the web, everything published to the web feeds AI. That is the web's purpose now.
Also there are plenty of people who create content because they love it, and also need to be able to make a living at it, because doing so at the level of quality they want is time consuming and expensive.
But mostly because even people who produce content because they love it want to share that content with the world and that will be nigh impossible when the only content anyone sees, and that any platform or algorithm surfaces, is AI generated. Why put in the effort and heart and work to create something only for an AI to immediately clone it for ad revenue? Why even bother?
> The only way to avoid that is to never publish to the web, everything published to the web feeds AI. That is the web's purpose now.
And in doing that you also prevent real humans from accessing that same content. Look, I have no simpathy for AI companies. I wrote about it before on my site, will probably write again. The current situation sucks. But giving up is not the right answer imo.
> Also there are plenty of people who create content because they love it, and also need to be able to make a living at it, because doing so at the level of quality they want is time consuming and expensive.
Fair but those are the minority. I'd argue the vast majority of people create content because they enjoy the process and earn a living in other ways. I run a newsletter where I interview people with blogs and so far, after a year running it, not a single person has told me they blog for a living. Every single one is doing it for passion. And I suspect that's true for the vast majority of people out there. The bulk of internet content (when it comes to creative content that is) is created by people who do it as a hobby.
> But mostly because even people who produce content because they love it want to share that content with the world and that will be nigh impossible when the only content anyone sees, and that any platform or algorithm surfaces, is AI generated. Why put in the effort and heart and work to create something only for an AI to immediately clone it for ad revenue? Why even bother?
Why even bother? Because there are people out there who care. And the assumption that "the only content anyone sees, and that any platform or algorithm surfaces, is AI generated" is a wrong one imo. I can assure you that there are PLENTY of people out there who still value original content, still value connecting with real human beings doing things because they love the craft. Assuming everything is doomed is not helpful.
Is it going to be harder? Yes. Are there solution? Yes.
In an infinitely large world with an infinitely large number of monkeys typing an infinite number of words on an infinite number of keyboards, "just index everything and threat it as fact" isn't a viable strategy any more.
We are now much closer to that world than we ever were before.
> new models will start getting trained with mostly the output of other LLMs
That is a naive, flawed way to do it. You need to filter and verify synthetic examples. How? First you empower the LLM, then you judge it. Human in the loop (LLM chat rooms), more tokens (CoT), tool usage (code, search, RAG), other models acting as judges and filters.
This problem is similar to scientific publication. Many papers get published, but they need to pass peer review, and lots of them get rejected. Just because someone wrote it into a paper doesn't automatically make it right. Sometimes we have to wait a year to see if adoption supports the initial claims. For medical applications testing is even harder. For startups it's a blood bath in the first few years.
There are many ways to select the good from the bad. In the case of AI text, validation can be done against the real world, but it's a slow process. It's so much easier to scrape decades worth of already written content than to iterate slowly to validate everything. AlphaZero played millions of self games to find a strategy better than human.
In the end the whole ideation-validation process is a search for trustworthy ideas. In search you interact with the search space and make your way towards the goal. Search validates ideas eventually. AI can search too, as evidenced by many Alpha model (AlphaTensor, AlphaFold, AlphaGeometry...). There was a recent paper about prover-verifier systems trained adversarially like GANs, that might be one possible approach. https://arxiv.org/abs/2407.13692v1
Communal spaces are fine, communal spaces will continue to be fine. Forums are fine. IRC is fine. The only thing that's dying is Google. Google is not the Internet.
> this also means that because we've exhausted the human generated content by now as means of training LLMs, new models will start getting trained with mostly the output of other LLMs
There is also a rapidly growing industry of people whose job it is to write content to train LMs against. I totally expect this to be a growing source of training data at the frontier instead of more generic crap from the internet.
Smaller models will probably stay trained on bigger models, however.
> growing industry of people whose job it is to write content to train LMs against
Do you have an example of this?
How do they differentiate content written by a person v/s written by LLM, I'd expect there is going to be people trying to "cheat" by using LLMs to generate content.
> How do they differentiate content written by a person v/s written by LLM
Honestly, not sure how to test it, but this is B2B contracts, so hopefully there's some quality control. It's part of the broad "training data labeling" business, so presumably the industry has some terms in contracts.
ScaleAI, Appen are big providers that have worked with OpenAI, Google, etc.
They just give you such an insight into another human being in this raw fashion you don’t get through a persona built website.
My own blog is very similar. Haphazard and unprofessional and perhaps one day slurped into an LLM or successor (I have no problem with this).
Perhaps one day some other guy will read my blog like I read Makoto Matsumoto’s. If they feel that connection across time then that will suffice! And if they don’t, then the pleasure of writing will do.
And if that works for me, it’ll work for other people too. Previously finding them was hard because there was no one on the Internet. Now it’s hard because everyone’s on it. But it’s still a search problem.
Print-on-demand means that paper books will be just as flooded with LLM sludge as eBook stores. I think we are at risk of regressing back to huge publishers being de-facto gatekeepers, because every easily accessible avenue to getting published is going to get crushed under this race to the bottom.
Likewise with record labels if platforms like Spotify which allow self-publishing get overwhelmed with Suno slop, which is already on the rise (there's some conspiracy theories that Spotify themselves are making it, but there's more than enough opportunistic grifters in the world who could be trying to get rich quick by spamming it).
> The Fifty Shades trilogy was developed from a Twilight fan fiction series originally titled Master of the Universe and published by James episodically on fan fiction websites under the pen name "Snowqueen Icedragon". Source : https://en.wikipedia.org/wiki/Fifty_Shades_of_Grey
The AI is already tainted with human output.... If you think its spitting out garbage it's because that's what we fed it.
There is the old Carlin bit about "for there to be an average intelligence, half of the people need to be below it".
Maybe we should not call it AI rather AM, Artificial Mediocrity, it would be reflection of its source material.
The issue is that the AI shit is flooding out anything good. Nearly any metric you can think of to measure "good" by is being gamed ATM which makes it really hard to actually find something good. Impossible to discover new/smaller authors.
Scale matters. The ability to churn out bad writing is increasing by orders of magnitude and could drown out the already small amount of high quality works.
While it's true that the volume of bad writing is increasing, our ability to analyze and refine this sludge is also improving. Just as spell check and grammar check give instant feedback why not AI instant feedback about writing quality / originality / suitability / correctness / … ? If instant feedback can improve spelling and grammar why not these other things?
> But this also means that because we've exhausted the human generated content
Putting aside the question of whether dragnet web scraping for human generated content is necessary to train next gen models, OpenAI has a massive source of human writing through their ChatGPT apps.
I use LLM output from kagi too. But given the rate of straight-up factually incorrect stuff that comes out of it, I need it to come with a credible source that I can verify. If not, I'm not taking any of it seriously.
AlphaGo learned to play Go by playing with itself. Why couldn't LLM do the same? They got plenty of information to be used as a starting point, so surely they can figure out some novel information eventually.
LLMs aren't logically reasoning through an axiomatic system. Any patterns of logic they demonatrate are just recreated from patterns in input data. Effectively, they can't think new thoughts.
> Effectively, they (LLMs) can't think new thoughts.
This is true only if you assume that combining existing thought patterns is not new thinking. If they can't learn a certain pattern from training data, indeed they would be stuck. However, their training data keeps growing and updating, allowing each updated version to learn more patterns.
Google really missed the opportunity of becoming ChatGPT. LLMs are the best interface for search but not yet the best interface for ads so it makes sense for them to not make the jump. ChatGPT and Claude are today what Google was in 2000 and should have evolved to.
Mind you they will be trained on what humans have filtered as being acceptable content. Most of the trash produced by ML that hits the web is quickly buried and never referenced.
The incentives will be largely gone when SEO-savvy AI bots will produce 10K articles in the time it takes you to write one, so your article will be mostly unfindable in search engines.
Human generated content will be outpaced by AI generated content by a large margin, so even though there'll still be human content, it'll be meaningless on aggregate.
We can adapt. There's already invite-only and semi-closed online communities. If the "mainstream" web becomes AI-flooded, where you'd you like to hang out / get information: the mainstream AI sludge, or the curated human communities?
The LLM is trained by measuring its error compared to the training data. It is literally optimizing to not be recognizable. Any improvement you can make to detect LLM output can immediately be used to train them better.
GANs do that, I don't think LLMs do. I think LLMs are mostly trained on "how do I recon a human would rate this answer?", or at least the default ChatGPT models are and that's the topic at the root of this thread. That's allowed to be a different distribution to the source material.
Observable: ChatGPT quite often used to just outright says "As a large language model trained by OpenAI…", which is a dead giveaway.
This is the result of RLHF (which is fine-tuning to make the output more palatable), but this is not what training is about.
The actual training process makes the model output be the likeliest output, and the introduction phrase you quoted would not come out of this process if there was no RLHF. See GPT3 (text-davinci-003 via API) which didn't have RLHF and would not say this, vs. ChatGPT which is fine-tuned for human preferences and thus will output such giveaways.
This seems like it would only work if you deliberately rank AI-generated text above human generations.
If the AI generations are correct, is it really that bad? If they're bad, I feel like they're destined to fall to the bottom like the accidental Facebook uploads and misinformed "experts" of yesteryear.
Where would the AI get the data necessary to generate correct answers for novel problems or current events? It's largely predictive based on what's in the training set.
> Where would the AI get the data necessary to generate correct answers for novel problems or current events?
In a certain sense, it doesn't really need it. I like to think of the Library of Babel as a grounding thought experiment; technically, every truth and lie could have already been written. Auguring the truth from randomness is possible, even if only briefly and randomly. The existence of LLMs and tokenized text do a really good job of turning statistics-soup into readable text.
That's not to say AI will always be correct, or even that it's capable of consistent performance. But if an AI-generated explanation of a particular topic is exemplary beyond all human attempts, I don't think it's fair to down-rank as long as the text is correct.
Are you suggesting that llms can predict the future in order to address the lack of current event data in their training set? Or is it just implicit in your answer that only the past matters?
We would lose the long tail, but if I were a search engine, I would have a mode that only returned results on a whitelist of domains that I would have a human eyeball every few months.
If somebody had a site that we were not indexing and wanted to be, they could pay a human to review it every few months.
You can make as many sites as you like, but I would still ask a human to review them and make a judgment call on whether other humans might be interested in the content before indexing them.
You can record as many albums as you like as well, but the DJ needs to like your music before they play it on the radio.
I guess what I’m saying is I don’t want the Internet to become a Top 25 radio station cranking out scam entertainment for the masses. I want “small pirate and indie radio” to be the norm. If you want top 25’s, go back to centralized, curated media.
The thing with the AI content boom is that if there’s 1000x more of it than there is genuine indie stations, it gets harder to find the real content. Piping things through a top25 filter doesn’t fix that, or actively makes it worse due to the incentives to monopolize / plan the system.
For example, i've been appending "site:reddit.com" to some of my Google queries for a while now —especially when searching for things like reviews— because, otherwise, Google search results are unusable: ads disguised as fake "reviews" rank higher than actual reviews made by people, which is what i'm interested in.
I wouldn't be surprised if we evolve some similar adaptations to deal the flood of AI-generated shit. Like favoring closer-knit communities of people we trust, and penalizing AI sludge when it sips in.
It's still sad though. In the meantime, we might lose a lot of minds to this. Entire generations perhaps. Watching older people fall for AI-generated trash on Facebook is painful. I hope we acted sooner.
Ok, i'll concede that you have a very good point there. Trust can be (and is being) exploited.
I guess for me, so far at least, some sites feel much more legit and human that the obviously bot-ridden mess that are the likes of Twitter/Instagram/FB. Like for example here or on Lobsters (and more on the latter) i have the feeling that it's mostly people talking with people. On the couple of relatively-small subreddits i visit, i feel the same too.
But i could be wrong of course. Maybe the tone of an HN poster is super easy for an LLM to copy; there's a reason why "shit HN says" exists after all. The only reason i have to believe otherwise is that, in comparison, Instagram or Twitter bots are so obvious and bland, and those companies have way more resources to throw at AI than HN or reddit :P
It writes content that's worth reading, but it's extremely expensive to run. It requires chain of thought, a RAG pipeline, self-revision and more.
I spent most of yesterday testing it and pushed it to beta, but the writing feels stilted and clearly LLM generated. The inflection point will come for content people actually want to read, but it's not going to be GPT-4o mini.
The point isn't to generate good content "that's worth reading". The point is to generate an endless stream of slop which looks plausible enough to get you ad impressions.
That's picking up pennies in front of a steam roller: Google is incentivized to punish you when the content is garbage, and people are disincentivized to share what you generate.
It's an entirely different game once you can generate useful content worth reading with AI. People will even pay you good money for it.
I don't know what makes you think that Google is incentivised to punish garbage. They certainly don't seem to mind seeing up an endless stream of slop for certain kinds of queries. I don't understand why they'd be more incentivised to serve SEO'd human generated slop than SEO'd machine generated slop.
If you actually look into the "SEO slop", they're constantly fighting a battle with Google to keep their place.
It's all garbage so no one notices when some of it suddenly disappears off the face of the earth and gets replaced with other garbage: but for the ones making it, their revenue essentially goes to 0 overnight.
I had a lot of fun with NovelAI. I believe at the time it was using GPT2, and I loaded in fine tuned models for the canon of choice I wanted to experience (trained on fanfic, and things of that sort).
Spellbound is an instruct model to NovelAI's completion model: you enter commands which in turn dictate what happens to your character, then the AI models how others would react to you
Given the younger generations increasing ambivalence to the non-stop fire hose of bullshit that the vast majority of the platform internet already is, and given that we're now forging the tools to make said fire hose larger by numerous factors, I don't think this is going to be the boon long-term that a lot of people seem to think it is.
It is extremely ironic that computers which operate by the logic of boolean arithmetic and algebra are now used to generate bullshit instead of adding rigor and checking existing written content for basic falsehoods and logical fallacies.
Itch.io has almost no crap filters so all you find is crap. Steam lets anyone publish but you rarely come across any crap. Many PC game devs know that the income overwhelmingly comes from Steam vs every other site put together.
Unfortunately, this just gives more power to the walled gardens.
so, status quo? this sort of content only has value because google links to it when people search, and because google runs an ad network that allows monetizing it. google is also working furiously to provide these same AI-generated answers in their SERP, so they can eliminate this and monetize the answers directly instead of paying out to random third parties.
i'm pretty skeptical that this ai-generated content will ever be monetizable in the way the article suggests, simply because google is better at it. if you're a human making your living by writing articles that are indistinguishable from ai-generated content, then you might be harmed by this but for most people this inflection point is not going to be a noticeable change.
> For example, putting in 50k page views a month, with a Finance category, gives a potential yearly earnings of $2,000.
> I'm going to take the median across all categories, which is an estimated annual revenue of $1,550 for 50,000 monthly page views.
> This is approximately ~$0.00022 earned per page view.
The problem is... this doesn't take into account a million AI generated sites suddenly all competing for the same amount of eyes as before, driving revenue to zero very quickly. It'll be worth something for a bit and then everyone will catch up.
Many people in the history of the internet have made a lot of money by doing something that was "worth something for a bit and then everybody caught up"
The same number of eyes will still be driven to a subset of content by algorithmic influence. Whether search engines, algorithmically-generated "viral" popularity, or whatever. Most people are consuming whatever is placed in front of their faces. That content will still have value, the trick will be getting your content into that subset.
This article is weird clickbait which, even weirder, worked.
It seems to assume a world where SEO entrepreneurs where ready to churn out million-page sites, but the cost per query were blocking them. There is no marginal cost, no SEO cost to adding another page, as long as a couple people visit it and "pay it off".
In the real world, it doesn't work like that. Whatever monstrosity was created like this would not do well in the search engines. So no meaningful threshold has been passed, in terms of the cost for AI generation.
People are creating lots of AI content, but not like this - not bottom tier generic SEO pages which will barely rank and aren't that compelling in an already saturated Internet.
Incidentally the real money seems to be in generating AI images and, eventually, video: much better return for your money.
This assumes a future where users are still depending on search engines or some comparative tool. Profiting off the current status quo. I would also be curious how user behavior will evolve to identify, evade, and ignore AI generated content. Some quasi arms race we'll be in for a long time.
True, but ChatGPT has been interviewed by a national television broadcaster in the UK at least, so I think it broke out of our bubble no later than December 2022: https://youtu.be/GYeJC31JcM0?si=gdmlxbtQnxAvBc1i
This has already been happening for quite some time with users ignoring Google search and searching Reddit directly. The irony is that, I assume, most of Reddit's income right now is coming from content licensing deals with AI companies.
So it will now be cost-effective to connect the exhaust of ChatGPT to its inlet and watch as the quality of output deteriorates over time while making money off ads. Whatever rocks your boat, I guess. How long before the answer to every prompt is "baaa baaa baaa"?
You’re sadly misinformed if you think training an LLM consists of dumping the unfiltered sewage straight from the web into a training run. Sure, it’s been done in early experiments but after you see the results you learn the value of data curation.
That article itself might be part of the degradation. It mentions at least four times that the contract was canceled as if it's something new. I wonder if someone just dumped a bunch of facts and ran it through a spin cycle a few times with AI to get a long form article they didn't expect anyone to read.
It's clearly working because the models are only getting better, believing that the performance of these models would fall at some point in the future is just very delusional.
Weren't they just getting better mostly because they were being scaled up? There's no way to do that once you've exhausted all of the data. Besides progress has slowed down at this point anyway.
Not only. Look at the subject of this thread, GPT-4o mini.
I'm optimistic about synthetic data giving us another big unlock, anyway. The text on the internet is not that reasoning dense. And they have a snapshot of pre-2023 that is fixed and guaranteed not to decay. I don't think one extra year of good quality internet is what will make or break AGI efforts.
The harder bottleneck will be energy. It's relatively doable to go from 1GW to 10GW but the next jump to 100GW becomes insanely difficult.
GPT-3 was 173B parameters and it's very bad compare to much smaller models we have nowadays, the data and the compute play a giant role, also I doubt you would need to train a model further after you have trained it on absolute everything (but we are very far from that).
> Will the future of the internet be entirely dynamically generated AI blogs in response to user queries?
I still enjoy commenting on HN and writing some thoughts on my blog. I'm pretty sure that there are many other people too.
At some point everything that is not cryptographically singed by someone I know and trust needs to be considered AI generated.
Maybe AI-generated content might have better quality than generated by humans. But then it's likely that I'm under the influence of some bigger corporation that just needs some eyeballs.
Generating content on the fly is already happening, has been for a while. Word spinners used with a script that grabs the content of the first 5 Google results, Wikipedia, etc, has been around a long time and Google indexed the incomprehensible garbage it created.
From what I can tell, all scalable automated work falls in value towards zero over time.
For example, a person could write a shareware game over a few weeks or months, sell it for $10, buy advertising at a $0.25 customer acquisition cost (CAC) and scale to make a healthy income in 1994. A person could drop ship commodities like music CDs and scale through advertising with a CAC of perhaps $2.50 and still make enough to survive in 2004. A person could sell airtime and make speaking appearances as an influencer with a CAC of $25 and have a good chance of affording an apartment in 2014. A person can network and be part of inside deals and make a million dollars yearly by being already wealthy in a major metropolitan city with a CAC of $250 in 2024.
The trend is that work gets harder and harder for the same pay, while scalable returns go mainly to people who already have money. AI will just hasten the endgame of late stage capitalism.
Note that not all economic systems work this way. Isn't it odd how tech that should be simplifying our lives and decreasing the cost of living is just devaluing our labor to make things like rent more expensive?
It's only odd if you model economics as a cooperative venture between a society trying to build better collective outcomes, and not as a competitive system. Additional capability and information can never hurt a single actor taken in isolation. But added capability and information given to multiple actors in a competitive game can make them all worse off.
As a simple example, imagine a Prisoner's Dilemma, except neither side knows defecting is an option (so in effect both players are playing a single-move game where "cooperate" is the only option). Landing on cooperate-cooperate in this case is easy (indeed, it's the only possible outcome). But as soon as you reveal the ability to defect to both players, the defect-defect equilibrium becomes available.
If you read the Black Swan by Taleb, it stops being weird. He points this out and dubs it Extremistan, where small advantages accrue to oversized returns.
We will need humane solutions to this, because the non humane ones are starting to become visible (armed drone swarms driven by AI).
Well, computer hardware was stagnating without a forcing function. Running LLM’s locally is a strong incentive to get hardware more powerful and run your own local models without any ADs.
Why would ad companies not generate themselves the content ? It make no sense what he is saying. They pay for ads because today they can’t write said content. If now they can why pay other people ?
I wouldn’t be surprised to see ads injected in LLM answers, that’s the logical way to go. Free LLM with ads
There may well be tiers like on NFLX, where some are ad-sponsored and some are not. But seeing as how rapidly the free/open models catch up to prior-generation proprietary models, I doubt there will be much margin or room for ads on anything but the latest/greatest.
I strongly assume there are higher rate limits, more than once I've seen the Right Kind of Startup, (buzzworthy, ex-FAANG with $X00m in investment in a market that's always been free, think Arc browser), make a plea on twitter because they launched a feature for free, were getting rate limited, and wanted a contact at OpenAI to raise their limit.
Arc is an excellent example because AFAIK it's still free, and I haven't heard a single complaint about throttling, availability, etc., and they've since gone on to treat it as a marketing tentpole instead of experiment.
I tried it on technical queries and it hallucinated like crazy. Probably ok for narrow tasks, but I wouldn’t expose it through the main UI like they did - people expect some degree of intelligence there.
I am enjoying this Moment in time where I can ask chatgpt product related questions and not get ad biased suggestions.
I think there is ~ half a year left
The enshitification of search will drive queries directly to AI, either local or centralised. This will provide a before unknown nexus of opinion/ perception / idea control as the primary research tool will no longer return a spectrum of differing ideas and references, but rather a consolidated opinion formed by the AIs operators.
This has really dystopian vibes, since it centralizes opinion and “factuality” in an authoritative but potentially extremely biased or even manipulatively deceptive manner.
OTOH it will provide opportunities for competitive solutions to query answering.
So this has been an inflection point that has concerned me, specifically in regards to a few types of sites: news and instruction sites.
News sites are already often shit and parasitic. I mean parasitic because if you go to a free news site (say Yahoo news, etc) you often see rewritten articles that originated from paid sites (e.g. NYT). The pure ad-supported sites are typical enshitification that degrades journalism and increases sensationalism because they don't need to write unique articles, but you should sensationalize them to drive up views. You also don't have to hire journalists to get story details. So news most people read degrades and you get very limited views.
The problem here is that this paradigm barely works because you have to pay real people to write those rephrased articles. So while it costs more to run the NYT where you need to hire investigative journalists and send people to physical places, there is a bound on that difference. But if you paste in a NYT article into GPT4 and ask it to summarize it, you'll get very similar quality to yahoo news (or even CNN, MSNBC, or Fox. Which all also do this leeching, but less of an issue). I'm sure people realize how easy it is to scrape NYT and then post the GPT output. This is in spirit no different than if you just used archie.is, but large scale.
The same is true for many tutorial sites or cooking sites, etc. I'm sure many of you also get annoyed at the google search results that are just stackover flow posts embedded on a different site or the Medium articles (especially paid ones) that are also just SO posts and can show up higher in the listing.
The issue becomes: how do we generate and disseminate new information in this paradigm? Okay, free blog posts aren't "hurt" because they have no income, but people build reputation through them and it gets many people jobs. But what about others that do make a living through this? Is this not similar Jack Conte's (Patreon co-founder/CEO and 1/2 of the band Pomplamoose) argument about creating content "for the algorithm" vs for "yourself/your fans/fun/etc". That it is taking some of the human elements out of the art/entertainment/content. (Can totally disagree with his argument btw). Personally I'm on the side of Jack. Our goal shouldn't (now) be to just serve people search results or just generate content for content's sake, but to now focus on serving people high quality content and high quality results. Google indexed the entire internet. People gamed the system (SEO) and now google results are shit, youtube results are shit, and everything is shit. We don't need more content (who uses page 2 on Google?), but we need to have better content. [1]
I think we need to ask: is this what we want? If not, then what are we going to do about it?
If we are okay, then I think someone should create a super-website where you just have information about just about everything. There definitely is utility in it. But the question is at what cost.
[1] I think most people want this. But the problem is you're not going to find market forces showing this because there is no product doing this. Or if there are, they aren't well known and could be confusing to use and/or a wide variety of problems (UI/UX do matter). But it requires reading between the lines and market research a la talking to people and finding out what they want, not a la data. You need both.
When views are low the math doesn't make sense but it is very possible to get a lot of views through AI generated + human reviewed content.
We're trying to do that with PulsePost (https://pulsepost.io) and the biggest challenge is unique content. Given a keyword or a niche topic, AI models tend to generate similar content within similar subjects. Changing the temperature helps to a degree but the biggest difference comes from adding internet access. Even with same prompt, if the model can access the internet, it can find unique ideas within the same topic and with human review it becomes a high value article.
But this also means that because we've exhausted the human generated content by now as means of training LLMs, new models will start getting trained with mostly the output of other LLMs, again because the web (as well as books and everything else) will be more and more LLM-generated. This will end up with very interesting results --not good, just interesting-- akin to how the message changes when kids the telephone game.
So the snapshot of the web as it was in 2023 will be the last time we had original content, as soon we will have stop producing new content and just recycling existing content.
So long, web, we hardly knew ya!