Human writers also use specific constructions for definite (though not always consciously known) reasons, and mapping those correlations onto writing models is only getting done on accident. This is a big part of why LLMs doing pointless negations (it's not X, it's Y) is so jarring: people use that construction in order to preempt an expected misconception, whereas LLMs try to use linear algebra over the learning corpus to generate it in appropriate quantities and places according to RLHF/fine-tuning/etc.

In other words, it isn't just verbiage that has a "preferred" quantity, but also higher-level language features, and the appropriate contexts for those features feel like the sort of thing that will be extraordinarily difficult to replicate without a theory of mind.

edit after coming back and re-reading: the "in other words," and rephrase is because I'm trying to explain something complex enough that I'm not even sure I fully understand it, so I try typing out multiple explanations and leave the ones that make the most sense to me. This might be idiosyncratic to my writing and not a significant feature of training corpuses, but even if it wasn't it's a second example of the sort of thing I'd suspect LLMs to be bad at using appropriately.

I vaguely remember a study that I was able to find when fact-checking this comment before publishing it [1], which found that music needs periods of high predictability to nurture listener expectations for the sake of later subverting those expectations, and the most enjoyable parts of music are those moments where expectations are subverted. This sounds like a similar effect.

1: PDF warning https://www.cell.com/current-biology/pdfExtended/S0960-9822(...

Gemini models are so, so much better in this regard. Their outputs are actually readable.

Interesting, I was thinking the walls of text are better for the watermarks.

> Gotta take some bad with the good, I suppose.

Well, it's quite simple if you think about it. Intelligence is all about awareness of the imminence of death and the finite amount of time given to any of us. So, we humans intrinsically evolved to make the most of it. The majority of the tools we create have the ultimate purpose - to save time. Even money is nothing but a way to cut some time. It's not really about time but experiences - we want to have as many and as rich as we can get, and we innately try to cut time somewhere else, so we can use what we saved for better experiences.

Therefore, inventing cars was an easily predictable event. There's not a single hypothetical timeline or dimension where humans don't end up with cars - to transport things and to travel faster. Similarly, it was always simply inevitable - we would have invented something like LLMs, sooner or later.

Cars required significant sacrifices - they are absolutely deadly machines, they kill thousands of people every single day, kill wildlife, endager our own existence, and we're kind of okay with that. I can absolutely assure you, if someone tomorrow invents a gravitational jump travel that allows moving things between continents in minutes, even with a success rate as low as 10%, we would still be using it.

So, yes. We can't un-invent cars and AI, the only way to move forward is to adjust to this new reality and extract the best possible experiences, preferably by not robbing other people of their - i.e., enjoy life while it lasts, it still remains finite.

You can be pro-"thinking AI is going to get really good and solve many problems" (i.e. be the sort of person who would "go long" on AI), while also being anti-"applying AI right now to domains outside of the few it's been specifically optimized for [e.g. coding]" (because in these other domains, AI is yet to approach, let alone surpass, "authentic quality" equivalent to that of a skilled junior professional.)

I can like the concept of the engine, and the leverage it provides, but I hate traffic and get anxious about having an accident on the freeway. And the people behind the industrial revolution have destroyed the climate.

I love computers and the internet, and the internet has ruined a lot of people and things around me.

It's possible to be interested in a new technology and enjoy its leverage while still being extremely apprehensive about how it intersects with game theory and politics. The race to the bottom of the barrel keeps coming for the jugular of every cool new thing.

In a prior evolution of this, if you look into the Get Rich Quick marketing copy of the early 2000s (before SEO had taken over content formatting), you can see a parallel of sorts in those marketers doing human sales optimization Impact with bold and underline. It derives from research into the human response to ads from decades of print ads and centuries of printed matter prior, and I think there is a direct lineage from ‘how can we bias humans through restructuring text’ to ‘how can we bias machines through restructured text’. This isn’t to disagree, but to call out how rapidly humanity (merely twenty years!) has switched from marketing to people, to marketing to algorithms, in search of a favorable bias.

Ah yes I remember meeting this end of spontaneous expression to bend toward whatever might please some algorithm.

Soulless Enshitificatoon Oblivion now exercises a huge weight on societies to such an extent, it's ridiculous.

It’s heavily RL’d and it’s gotten worse because they’re using LLM judges, that’s why it’s gotten worse. The last good writing model was GPT 4.5.

I suspect they’re leaning heavily on some kind of auto verifier, and that’s why the writing sucks, because good writing is so subjective.

And they’re probably nerfing what it can get from training as they’ve lost lawsuits for scraping copyrighted work, so they’ll purposefully avoid anything that references a specific writers style, etc.

Isn't this easy for frontier 'labs' to fix? If the user requests lots of writing and prose, then use an auxiliary model which has been trained on 'award-winning' writing instead of all the crap on the internet?

The "regression to the mean" or "it's just the training data" hypothesis doesn't work. Pretrained models do not have these weird writing styles and sound much more human. It's only after instruction tuning -- SFT, DPO, etc. -- that LLMs develop weird writing styles.

That is, the original training on massive text corpora seems to produce an average of human writing, but that average looks pretty average. It doesn't feel like LLM slop. But the instruction tuning on specific tasks (answer this prompt, summarize this text, do this math problem), which turns text predictors into chatbots, seems to create strange style features as a side effect. And because that tuning step depends on company-specific training data (rather than just a crawl of the entire Internet), each company's LLM has a different style.

It's not that difficult to measure this. Linguists have had measures of style variation for a long time, though NLP and AI researchers seem to have forgotten. We've generated parallel human and AI corpora to measure it, and can show specifically which writing features change: https://arxiv.org/abs/2410.16107

Currently we're working on tweaking the instruction tuning process to see if we can pin down which steps and what types of tasks tend to produce the weird style features.

Do you have more info on this?

I've noticed this as well, I've been comparing it to 4Chan Greentext memes. Everything is a short punchy sentence, and every sentence has the same "weight" to it. There's no flow or rhythm, it's like reading off bullet points.

They will never admit it, but they give clues.

It makes no sense on being "aggressively pro AI" without any payment. It is like being an unpaid shill.

Is typically called ai booster. The opposite is ai doomer.

Leans into AI adoption.