Blog spam was generally written by humans. While it sucked for other reasons, it seemed fine for measuring basic word frequencies in human-written text. The frequencies are probably biased in some ways, but this is true for most text. A textbook on carburetor maintenance is going to have the word "carburetor" at way above the baseline. As long as you have a healthy mix of varied books, news articles, and blogs, you're fine.

In contrast, LLM content is just a serpent eating its own tail - you're trying to build a statistical model of word distribution off the output of a (more sophisticated) model of word distribution.

SEO text carefully tuned to tf-idf metrics and keyword stuffed to them empirically determined threshold Google just allows should have unnatural word frequencies.

LLM content should just enhance and cement the status quo word frequencies.

Outliers like the word "delve" could just be sentinels, carefully placed like trap streets on a map.

2. Given how LLMs work, a prompt is a bias — they're one-and-the-same. You can't ask an LLM to write you a mystery novel without it somewhat adopting the writing quirks common to the particular mystery novels it has "read." Even the writing style you use in your prompt influences this bias. (It's common advice among "AI character" chatbot authors, to write the "character card" describing a character, in the style that you want the character speaking in, for exactly this reason.) Whatever prompt the developer uses, is going to bias the bot away from the statistical norm, toward the writing-style elements that exist within whatever hypersphere of association-space contains plausible completions of the prompt.

3. Bot authors do SEO too! They take the tf-idf metrics and keyword stuffing, and turn it into training data to fine-tune models, in effect creating "automated SEO experts" that write in the SEO-compatible style by default. (And in so doing, they introduce unintentional further bias, given that the SEO-optimized training dataset likely is not an otherwise-perfect representative sampling of writing style for the target language.)

But, insofar as your use-case is against the TOUs of the big proprietary inference platforms, this second cost quickly swamps the first cost. They keep banning you, and you keep having to buy new dark-web credentials to come back.

Given this, it’s a lot cheaper and more reliable — you might summarize these as “more predictable costs” — to design a system around a substrate whose “immune system” won’t constantly be trying to kill the system. Which means either your own hardware, or a “being your own model” inference platform like RunPod/Vast/etc.

(Now consider that there are a bunch of fly-by-night BYO-model hosted inference platforms, that are charging unsustainable flat-rate subscription prices for use of their hardware. Why do these exist? Should be obvious now, given the facts already laid out: these are people doing TOU-violating things who decided to build their own cluster for doing them… and then realized that they had spare capacity on that cluster that they could sell.)

TFA mentions this hasn't been the case.

It’s the same origin. On Slashdot (the HN of the early 00’s) people would admonish others to RTFA. Then they started using it as a referent: TFA was the thing you were supposed to have read.

I miss the old Atomic MPC forums in the ~00s.

So it's classic positive feedback. LLM uses delve more, delve appears in training data more, LLM uses delve more...

Who knows what other semantic quirks are being amplified like this. It could be something much more subtle, like cadence or sentence structure. I already notice that GPT has a "tone" and Claude has a "tone" and they're all sort of "GPT-like." I've read comments online that stop and make me question whether they're coming from a bot, just because their word choice and structure echoes GPT. It will sink into human writing too, since everyone is learning in high school and college that the way you write is by asking GPT for a first draft and then tweaking it (or not).

Unfortunately, I think human and machine generated text are entirely miscible. There is no "baseline" outside the machines, other than from pre-2022 text. Like pre-atomic steel.

Some day we may view this as the beginnings of machine culture.

Have you ever seen someone use their smartphone? They're not "here," they are "there." Forming themselves in cyberspace -- or being formed, by the machine.

I don't think we'll have true machine culture until we have fully autonomous agents in the wild that are interacting with the world independently on its own terms. Right now the substrate is text which comes from a human mind -- it does not arise naturally from nothing. So the machine is a symbiote for now until we solve some difficult robotics problems.

I don't think you need to be independent or autonomous to develop a culture. And a lot of human culture was passed down over generations without understanding why it worked. We just imitate the behaviour and rituals from our most successful ancestors or role models.

If new LLMs can access the past generation's knowledge of how to please human evaluators, they will use it. It's not a deliberate decision by an "agent", it's just the best text source to copy from. This is a new feedback loop between generations of assistants, and it bypasses whatever the human designer had in mind. Phrases like "it is always best to ask an expert" will pop up just because you tuned the LLM to sound like a helpful assistant, and that's what helpful assistants sound like in the training data. You'd have to actively steer the new generation away from using their ancestral knowledge.

I guess it comes down to what your definition of "culture" is. There is no targeted teaching of the next generation, for example - but is this a requirement? I agree that talking about "machine culture" right now sounds like a stretch, but now I wonder what pieces are actually missing.

Far from saying the pool of language is now polluted, I think we now have a great data set to begin to discern authentic from inauthentic human language. Although sure, people on the fringes could get caught in a false positive for being bots, like you or I.

The biggest LLM of them all is the daily driver of all new linguistic innovation: Human society, in all its daily interactions. The quintillions of daily phrases exchanged and forever mutating around the globe vastly eclipses anything any hardware could ever emulate given the current energy constraints. Software LLMs are just a state machine stuck in a moment in time. At best they will always lag, the way Stalinist language lagged years behind rhe patois of average Russians.

If there is one distinctly human trait, it's sniffing out anyone who sounds suspiciously inauthentic. (Sadly, it's also the trait that leads to every kind of conspiracy theorizing imaginable; but this too probably confers in some cases an evolutionary advantage).

I'm a little more cautious though. I think GPT will be way more integrated, simply because it's useful. Stalinist language was artificial, in the sense that it was basically imposed on you from outside for no good reason. When you wanted to get real stuff done (either talking to close friends, being productive with colleagues, etc) you wouldn't use socialist newspeak because it got in the way. GPT will be imposed by the outside world, but it's actually a useful thing to be able to converse with a language model; you'll do it every day at work, when buying things, when using your phone/PC.

And also, unlike in USSR times, so much of our communication is online and visible. It would not surprise me if we develop a model that can train continuously on the firehose. Text is small. Data rate of every person on earth speaking simultaneously:

- 150 words per minute spoken

- 150 words × (5 characters/word + 1 space) = 150 × 6 = 900 characters per minute

- 1 byte per char = 900 bytes/min = 15 bytes/sec

- 15 bytes / sec * 8,000,000,000 people speaking continuously = 120 gigabytes/second

That's a lot but it's not even the bandwidth of a single consumer GPU.

  Too deep we delved, and awoke the ancient delves.

GOGI.

At least in Googles case, they're having so much difficulty keeping AI slop out of their search results that I don't have much faith in their ability to give it an appropriately low training weight. They're not even filtering the comically low-hanging fruit like those YouTube channels which post a new "product review" every 10 minutes, with an AI generated thumbnail and AI voice reading an AI script that was never graced by human eyes before being shat out onto the internet, and is of course always a glowing recommendation since the point is to get the viewer to click an affiliate link.

Google has been playing the SEO cat and mouse game forever, so can startups with a fraction of the experience be expected to do any better at filtering the noise out of fresh web scrapes?

Google has been _monetizing_ the SEO game forever. They chose not to act against many notorious actors because the metric they optimize for is ad revenue and and those sites were loaded with ads. As long as advertisers didn’t stop buying, they didn’t feel much pressure to make big changes.

A smaller company without that inherent conflict of interest in its business model can do better because they work on a fundamentally different problem.

A useful Search would ideally send a user to the site with the most signal and the fewest noise. Meanwhile, ads are inherently noise; they're extra pieces of information inserted into a webpage that at best tangentially correlate to the subject of a page.

Up until ~5 years ago, Google was able to strike a balance on keeping these two stable; you'd get results with some Ads but the signal generally outweighed the noise. Unfortunately from what I can tell from anecdotes and courtroom documents, the Ad team at Google has essentially hijacked every other aspect of the company by threatening that yearly bonuses won't be given out if they don't kowtow to the Ad teams wishes to optimize ad revenue somewhere in 2018-2019 and has no sign of stopping since there's no effective competition to Google. (There's like, Bing and Kagi? Nobody uses Bing though and Kagi is only used by tech enthusiasts. The problem with Google is that to copy it, you need a ton of computing resources upfront and are going up against a company with infinitely more money and ability to ensure users don't leave their ecosystem; go ahead and abandon Search, but good luck convincing others to give up say, their Gmail account, which keeps them locked to Google and Search will be there, enticing the average user.)

Google has absolutely zero incentive to filter out generative AI junk from their search results outside the amount of it that's damaging their PR since most of the SEO spam is also running Google Ads (since unless you're hosting adult content, Google's ad network is practically the only option). Their solution therefore isn't to remove the AI junk, but to instead reduce it enough to the degree where a user will not get the same type of AI junk twice.

A search engine isn't a two-sided market in itself but the ad network that supports it is. A better search engine is a technological problem, but a decently paying ad network is a technological problem and a hard marketing problem.

I've noticed that lately. It used to be the top google result was almost always what you needed. Now at the top is an AI summary that is pretty consistently wrong, often in ways that aren't immediately obvious if you aren't familiar with the topic.

The problem is that, of the signals you mention,

• the highly-informative ones (posting a new review every 10 minutes, having affiliate links in the description) are contextual — i.e. they're heuristics that only work on a site-specific basis. If the point is to create a training pipeline that consumes "every video on the Internet" while automatically rejecting the videos that are botspam, then contextual heuristics of this sort won't scale. (And Google "doesn't do things that don't scale.")

• and, conversely, the context-free signals you mention (thumbnail looks AI-generated, voice is synthesized) aren't actually highly correlated with the script being LLM-barf rather than something a human wrote.

Why? One of the primary causes is TikTok (because TikTok content gets cross-posted to YouTube a lot.) TikTok has a built-in voiceover tool; and many people don't like their voice, or don't have a good microphone, or can't speak fluent/unaccented English, or whatever else — so they choose to sit there typing out a script on their phone, and then have the AI read the script, rather than reading the script themselves.

And then, when these videos get cross-posted, usually they're being cross-posted in some kind of compilation, through some tool that picks an AI-generated thumbnail for the compilation.

Yet, all the content in these is real stuff that humans wrote, and so not something Google would want to throw away! (And in fact, such content is frequently a uniquely-good example of the "gen-alpha vernacular writing style", which otherwise doesn't often appear in the corpus due to people of that age not doing much writing in public-web-scrapeable places. So Google really wants to sample it.)

I’m guessing that the association between “pagers” and “Hezbollah” ended up creating the latter two tabs, but who knows. Maybe some AI video out there did a product review of Hezbollah.

Is it of your own coinage? When the AI sifts through the digital wreckage of the brief human empire, they may give you the credit.

https://en.wiktionary.org/wiki/nontent

Increasingly I find that for in-depth explanations or tutorials Youtube is the only place to go, but even there the search results can lead to loads of videos which just seem… off. But at least those are still made by humans.

It seems to me that the fact it's so cheap and relatively easy for people with dreams of becoming wealthy influencers to put content out there has more to do with the flood of often mediocre content than AI does.

Of course the vast majority don't have much real success and get on with life and the crank turns and a new generation perpetuates the cycle.

LLMs etc. may make things marginally easier but there's no shortage of twenty somethings with lots of time imagining riches while making pennies.

It sucks, because sharing recipes seemed like one of those things the internet could be really good at.

The only remaining reliable source - now that many newspapers are axing the remaining staff in favour of LLMs - is pre-2020 print cookbooks. Anything online or printed later must be assumed to be tainted, full of untested sewage and potentially dangerous suggestions.

I joke with folks that my assumption with "one clove of garlic" is that they really mean "one head of garlic" if you want any flavour. (And if the recipe title has "garlic" in it and you are using one clove, you’re lying.)

I read Spanish and Italian fluently and stumble my way through Japanese (with translation). It's easier to find a good recipe in these languages, provided you can find the ingredients or substitutes.

That is, when baking, you can usually (again, exceptions for creaming the sugar in butter, etc.) take all of your dry ingredients and mix/sift them together, and then you pour your wet ingredients in a well you’ve made in the dry ingredients (these can also usually be mixed together).

But for shortbread or fork biscuits those three could indeed all go in the bowl in one go (but that one admittedly doesn't really need a bracket because the recipe is "put in bowl, mix with hands, bake").

When I saw them, they blew my mind. Short to store and easy to understand.

The top of search results is consistently crowded by pages that obviously game ranking metrics instead of offering any value to humans.

For a while I thought email as a medium was doomed, but spammers mostly lost that arms race. One interesting difference is that with spam, the large tech companies were basically all fighting against it. But here, many of the large tech companies are either providing tools to spammers (LLMs) or actively encouraging spammy behaviors (by integrating LLMs in ways that encourage people to send out text that they didn't write).

There's probably an analogy to be made about the open decentralised internet in the age of AI here, if it gets to the point that search engines have to assume all sites are spam by default until proven otherwise, much like how an email server is assumed guilty until proven innocent.

However, the frontline of the email war has shifted lately. Now the most important part of the war is being fought over emails that look just like ham, but aren't. Business frauds where someone convinces you that they are the CEO or CFO or some VP and they need you to urgently buy this or that for them right now no time to talk is big business right now, and before you get too high-and-mighty about how immune you are to that, they are now extremely good at looking official. This war has not been won yet, and to a large degree, isn't something you necessarily win by AI either.

I think there's an analogy here to the war on content slop. Since what the content slop wants is just for you to see it so they can serve you ads, it doesn't need anything else that our algorithms could trip on, like links to malware or calls to action to be defrauded, or anything else. It looks just like the real stuff, and telling that it isn't could require a human rather vast amounts of input just to be mostly sure. Except we don't have the ability to authenticate where it came from. (There is no content authentication solution that will work at scale. No matter how you try to get humans to "sign their work" people will always work out how to automate it and then it's done.) So the one good and solid signal that helps in email is gone for general web content.

I don't judge this as a winning scenario for the defenders here. It's not a total victory for the attackers either, but I'd hesitate to even call an advantage for one side or the other. Fighting AI slop is not going to be easy.

Advertising in your mails isn't Google's.

I'm not saying this is impossible but that's going to be an uphill sell for me as a concept. According to some quick stats I checked I'm getting roughly 600 emails per day, about 550 of which go directly to spam filtering, and of the remaining 50, I'd say about 6 are actually emails I want to be receiving. That's an impressive amount overall for whoever built this particular filter, but it's also still a ton of chaff to sort wheat from and as a result I don't use email much for anything apart from when I have to.

Like, I guess that's technically usable, I'm much happier filtering 44 emails than 594 emails? But that's like saying I solved the problem of a flat tire by installing a wooden cart wheel.

It's also worth noting there that if I do have an email thats flagged as spam that shouldn't be, I then have to wade through a much deeper pond of shit to go find it as well. So again, better, but IMO not even remotely solved.

Looking at the last days’ spam¹ I have three 419-style scams (widows wanting to give away their dead husbands’ grand piano or multi-million euro estate) and three phishing attempts. There are duplicate messages in each category.

About fifteen years ago, I did a purge of mailing list subscriptions and there’s very little that comes in that I don’t want, most notably a writer who’s a nice guy, but who interpreted my question about a comment he made on a podcast as an invitation to be added to his manually managed email list and given that it’s only four or five messages a year, I guess I can live with that.

⸻

1. I cleaned out spam yesterday while checking for a confirmation message from a purchase.

I already find myself mentally filtering out audible releases after a certain date unless they're from an author I recognize.

How do we know what content LLMs were fed? Isn't that a highly guarded secret?

Won't the quality of the content be paramount to the quality of the generated output or does it not work that way?

I don't see how Google's SEO rules being written or unwritten has any bearing. Spammers will always find a way.

Based on the process above, naturally, the third iteration then is LLMs writing for corporate bots, neither for humans nor for other LLMs.