AI paid for by Ads – the GPT-4o mini inflection point

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.

So long, web, we hardly knew ya!

That's an inflection point, all right. OpenAI's customers can now at least break even.

Of course, it means a flood of crap content.

> 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.

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.

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"?

I don't know who these people are who can't even do basic arithmetic.

> an estimated annual revenue of $1,550 for 50,000 monthly page views.

> This is approximately ~$0.00022 earned per page view.

No, this is $0.002583 earned per page view, a ~12x difference. Looks like the author divided by 12 twice.

> 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.

This analysis implicitly holds supply constant. But supply isn't constant, it will balloon. So the price per impression will tank.

So, on the margin, this will drive human created content out since it is now less profitable to do it by hand than it was before.

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.

Lot cost models just lowered the bar of entry.

This has been possible for well over a year, just not with OpenAI’s API specifically.

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?

I read that title _very_ wrong as injecting ads directly into ChatGPT responses. How hilariously dystopian would that be?

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.

Do we really want ads that suggest the wrong product or make up imaginary products and prices ?

I 'd rather be exploited by google

Ad blockers, 30-40% of internet users. getting traffic... if it was that easy everyone would do it, and diminishing returns.

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

Won't people who hate ads just choose to cut out the middleman and use 4o mini on their own?

The rate limits though 15m tokens per month in the top tier isn’t really scale

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

Don’t think you’re getting 50k views pm in finance space with some “you’re an expert blog writer” AI spiel

Stuck culture? Meet REAL stuck culture.

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.

websim looks cool but requires a google login to even try it, i hate the internet in 2024

We're doomed.

Definitely the future of Twitter

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.

[0] https://youtu.be/hwn6-8XpIuE

[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.