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.
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.