It's not just predicting content but also the form/shape of the response, highly shaped by RLHF preferences, which seems to encourage longer form responses which need padding with filler to generate the "preferred" amount of verbiage.
It seems most of these models have been trained to interpret questions as wanting a long-form authoritative/comprehensive-sounding response (even when midwit filler is all it has to offer), rather than initiating a short-turn conversation which would better allow you to steer the conversation. Apparently Muse is much better in this regard, but so far I've resisted signing up for any Meta/FaceBook service, and not inclined to break this winning streak!
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.
Not to pick on GenHTTP Lambda, especially because I think it's someone's fun side project, but I was just trying to read though their docs and I was finding it super difficult. I kept getting distracted thinking "why the hell did it write it _that_ way?"
Like, the first sentence of "The control center" section [1] is the exact sort of weird negation that you're describing:
> The editor link opens a control center rather than a text box: most of the code here is written by agents, so the first thing on the screen is how your lambda is doing.
I can't help but think it generated the first round of the page with a text box to paste the code into, got told "no, I want a control center," and then this ended up in the docs.
I also think the phrasing of that sentence (and many others in those docs) is just super weird and unnatural - e.g. "so the first thing on the screen is how your lambda is doing." It also feels like it mirrors a prompt - like, "no, it should be a control center, not a text box, and the first thing the user should see is how their lambda is doing." But in docs I'd never expect it to be written "the status of your lambda" or "the health of your lambda." And what does the code being written by agents have to do with it?
1: PDF warning https://www.cell.com/current-biology/pdfExtended/S0960-9822(...
In that case, model will in fact strive to keep their writing as close to interesting writing as possible.
A short response that's wrong is going to be punished, whereas a long answer that contains the correct information, and then a whole bunch of other stuff surrounding that info, is more likely to be rewarded. Even if you discount the reward by answer length, when there's ambiguity involved it's still going to be a better strategy on average to hedge by saying too much vs risking saying too little.
I also try to keep the llm-enhanced (or even fully llm-generated) content separate from my human writing - because it is true that the failure modes are different. But think LLM writing might be useful even if it is not consumed directly - it can enhance LLM answering your questions.
(this isn't a new observation / thing, see https://randsinrepose.com/archives/nadd/ originally from somewhere in the 2000s)
But now with AI it's easy to generate a lot of prose. I'm thinking for example of the merge request descriptions my colleague submits. And here's the thing, it's not even bad or incorrect - I read it thinking, if I were to really sit down for it, get into The Zone etc, I could crank this out myself.
But I don't because honestly the MR is relatively trivial and can be described in just a sentence or two. Style and accuracy aside, my current take is that LLMs write things that humans wouldn't bother with, because it's too much effort and there's not enough value.
I mean it's an MR for an enterprise library project, the chances that anyone will ever read that MR description again once it's merged is close to zero.
That said: LLMs themselves don't get that, they will read everything you tell them to with the same amount of attention. Granted, their context windows are limited (and performance declines the more you put in it), so they too will use some tools to skim and pick out the important parts, but I think that's a temporary thing. The huge amount of LLM prose will end up being consumed by other LLMs. I don't know enough about the subject whether that's a good or a bad thing though, right now I think it just adds to the noise until it becomes noise in, noise out.
> how awful everything is to read lately > flooded with AI generated content, all of it terrible > everything sounding identical and everything being a rambling unconcise mess.
I can see the benefits of AI and enjoy using the technology - but, like many have said, it's wreaking havoc on the "content" that's out there nowadays. 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 lost me there. Children are pretty intelligent and have no concept of time. I don't see the relation you find obvious and the rest of your points similarly don't follow.
> Therefore, inventing cars was an easily predictable event.
This feels like a benefit of hindsight unless you can call out other obvious and imminent (but not yet existing) inventions.
I don't blame you for immediately rejecting the rest of it. But do hear me out: the point about the inevitability of invention does have real support. Historians call it multiple discovery - tons of things have been invented independently - calculus, the telephone, the jet engine, to name a few.
Neural nets, GPUs, and the scale of Internet text made something like LLMs close to inevitable. Of course, we could debate forever whether any of these on its own was ever inevitable. The point is - society tolerates any cost of a useful invention for the benefit of saving time. So it's reasonable to assume this will be the case with AI as well, which is really happening as we speak.
These are ultimately all software systems where updates can be pushed out with the press of a button. We just found an amazing new tool in the LLM and people are trying out all kinds of ways of using it, many of them dumb or stupid, this is discovery that has to be done, we'll keep doing it and pushing updates.
It doesn't take rose colored glasses really to infer that over time we'll cull the dumbest uses and extrapolate on the best ones. There's a huge difference between using a generic chatbot, even one driven by a SOTA model, vs targeting a specific business task with good guardrails, skills, and tool interfaces that call out to real business logic when the model isn't the right thing. The latter will emerge more and more over time.
But, that comes with with the same kind of cultural acknowledgement, as seen with the characters in Star Trek, that such algorithmically generated material is an experimental diversion, and inherently not worth the same respect or focus as experiencing something authored with intentionality by a person.
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.
The AI situation is different. The technology is obviously extraordinarily powerful but the benefits to humanity are still nebulous and there are many very obvious harms like the one we're discussing. So to describe yourself as "aggressively pro AI" you must believe that future, as-yet unrealized benefits will outweigh the harms. I don't think that's a crazy thing to believe but it deserves scrutiny and demands elaboration.
These challenges never go anywhere because it’s essentially a religious conviction. We’re likely in sort of uncharted territory on a macro scale and I reckon people a) either acknowledge that or refuse, and b), either trust or distrust the tech world’s stewardship of this technology. There’s a hundred bazillion data points, and I’m pretty pessimistic about the tech scene at large right now, but I think reasonable people can disagree on this without being factually or logically wrong. But true zealots in the midst of such uncertainty are driven by faith, which can certainly be deftly swayed over time, but definitely can’t be conquered in a debate.
I feel like in my life all I ever see is people mortgaging the future for today though. Hey how about that record storm in California?
You often have to make compromises somewhere in the hierarchy, especially in late-stage capitalism where even access to basic needs comes at a moral cost, and as such those compromises sometimes come with a moral duty. I am writing this via a computer built from limited natural resources, some of which have been obtained through what I see as slave labor. I don't eat meat and still have to acknowledge the harm that clearing and maintaining farmland does to wildlife, even if the result could theoretically be a net gain.
When you start viewing situations through the lens of game theory, you realise some compromises are inevitable for the rational individual. But you must acknowledge the individual is not the only concern. As such, I use AI as part of my work, but I opposed the local buildout of a large data center in my community and tried in vain to make the person responsible for the deal understand the issues it could introduce. This isn't NIMBYism, we need a halt on all buildouts until fundamental environmental and accountability concerns are addressed.
And many other aspects of my existence and quality of life, poor and underprivileged as I was growing up, have depended upon US imperialism. I have also produced a ton of waste just existing and owning pets. It's abhorrent.
I continue to ask myself how I can fulfill the moral duty which I have accrued and reduce my dependence on exploitative economic systems, but I also recognize that AI is the new social leverage and we only have so much time to use it to our advantage before it becomes institutionalized and wielded against the people. I also might only find future income stability by evolving with my industry.
Life is complicated.
AI has made it even worse and lots of content on the web, for both these reasons, is not worth you time (now wonder users don't read most of the text on our apps!)
Soulless Enshitificatoon Oblivion now exercises a huge weight on societies to such an extent, it's ridiculous.
If you built a model only based on long form journalism and The New Yorker voice, including both the template structure of the article and the grammar and syntax, the model would replicate the patterns it is built upon. You could do the same for high vocabulary rap.
When you stuff all of English or all of human thought into it with somewhat equal weight, including very technical writing the sum is a lack of style and elegance and personality.
The second problem is that writing often doesn’t happen linearly. You write bookends and fill in, edit and revise. You sort of know where you’re going and can then make decisions how to get there. Language models may benefit from being redesigned to perform passes. Select a template of some sort, fill in the mad lib, then run a detail pass, then run a persuasion / emotion pass, then run a style pass. Current model writing very much sounds like I would expect autocomplete to.
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.
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.
"You can be pro AI productivity but anti-AI creativity. Automating creativity is the death of art."
Meaning you reject the creative value of AI but values it's ability to drive a car safely, or write code for your job, or turn off the lights when you leave the room. objective goal oriented task completion great.
But AI draws your a picture? What's the point. AI writes you a song? No feeling. AI creativity isn't interesting.
There isn’t any doubt that AI can be a useful tool. So what you’re essentially saying is that we should avoid using a useful tool because many people, without any relevant training, use it badly.
By that standard, you should not be pro-car, pro-airplane, pro-boat, etc. etc.
In AI-generated entertainment, the names "Marcus", "Chloe", "Sterling", "Meridian", and "Apex" keep appearing. Those aren't common names, but LLMs seem to prefer them.
OP’s linked article mentions the importance for them of reading their writing aloud. Humans construct sentences with a (mostly unconscious) sonic influence. How things sound really matters.
My theory is that LLM prose sounds slop-like because the base unit for an LLM is not a single sound.
Thus, perhaps, you get the uncanny, meandering, hollow-feeling outputs, which feel vapid to human readers, without the reader really knowing quite why.
I think this is also part of the reason why LLMs are rewarded for and thus more strongly exhibit their faux-punchy constructions, such as the infamous “it’s not [X], it’s [Y].” The transformer model can, with current tokenization, articulate some features of “good writing”, but not quite all.
I think there are a number of fairly simple SOLNs to this (or ways to test this hypothesis) without making tokenizers purely syllabic across the board, which would probably be a mistake.
Whether prose output quality even matters for models which are not explicitly trained for creative writing is highly debatable, though.
For me it's mostly the latter, which is a natural consequence of LLM writing. When the (time) cost-per-word goes down, there is less incentive to keep signal-to-noise ratio high, so it ends up verbose and bloated. I also believe that this is (at least partially) upstream from everything sounding identical: take two articles on two different subjects - when SNR is high, there isn't really much room for them to become similar (The ingredient list for a risotto will never read the same as the ingredient list for pizza), but when 80% of them is fluff, they can blend together more easily (To take the recipe comparison, this would be the background/story part above the recipe), especially when written by the same LLM
It makes no sense on being "aggressively pro AI" without any payment. It is like being an unpaid shill.
My parents live on a rural property, and I'll help them do research on things like livestock practices or occasionally "prepper" related things to give them a sanity check before they spend money. I have to weed through a dozen or more obvious ChatGPT-created pages that are the first page or two of results on every search engine before I get to anything a human has touched. The vast majority of the LLM-created pages are either massively verbose and vague, or often wrong. But they're getting clicks. Very often, when asking an LLM the same question and to give citations, I'll get a combination of useless, irrelevant sites with no connection to the material, maybe one of these LLM-generated sites, and maybe one online forum.
I'm not aggressively pro-AI, I'm in the middle, but stuff like this has me worried that we're looking at the equivalent of the Dark Ages all over again. But instead of simply no one bothering to write anything down as was the case in the early Middle Ages, it's going to be the task of searching for real human work and thought in a sea of aggressively spammed LLM-generated slop pages.
A doomer isn't skeptical; they're nearly as credulous as boosters, or maybe more so in some ways.
(But I agree with you fwiw)
Endless headlines like, "aArON SORkIn SAYs hE WiLL Never UsE ai to WRite" doesn't change that base fact. No one asked him to, other than clickbait journalists.
If you find this hard to believe, just think about relative market sizes.
Yeah, thanks dude. We appreciate your efforts to enshittify everything.
Edit: Sure, downvote, but the parent comment is really not very different from "arsonist puzzled and upset that so many buildings have burned down".
I've noticed this a lot in conservative spaces especially, when they say "I love Trump, but..." -- why the need to say how much you love AI before criticizing it? Who are you trying to convince?
I use LLMs constantly for work and they have genuinely improved my life. But I absolutely despise generated "content". Text, audio, video, images -- for the reasons you describe.
It's ok to have a nuanced and multi-faceted opinion and approach to this and not have to call yourself a raving fanboy.
I find the same thing with written results out of AI "research", it is a slog to read through it and most of it can be skipped even if it can sometimes be useful its an unenjoyable read.
And the discipline gets attacked even more when all you read is LLM text. For example in code reviews there is no noticeable value in answering LLM generated review comments. I remember taking great care answering human-written reviews comments with much thought to either further build on their suggestions or to forward knowledge about the system when I had to defend a particular change.
Now, some bot will read my carefully written response and forgets about it. There is no value anymore.
No wonder that our pull request comments feel like the dead internet theory right now.
I am absolutely pro AI assisted coding. But communication between humans HAS to stay human. I am currently experiencing the opposite and it’s bad.
While I think your argument has merit, I would be careful to discount any human skill based on people being too lazy to do it.
But I don’t know where to draw the line yet. It’s difficult to decide what parts you are offloading and what parts you keep for your own brain. Because theoretically you can offload nearly everything when you get a well-defined ticket and just give it to an agent.
Maybe it’s also just me who is having a hard time staring at LLM generated texts all day and missing the past brainstorming, pair programming or “just talking about the project” sessions with colleagues and managers.
I fear that we are moving in the direction of silos again with lots of gate keeping as everyone has their own little rubber duck to talk to now.
Systems from today (if not worked on with discipline) will be no more than an agglomeration of features glued together. But as you said, guess our fight to fight will be to be that disciplined person.
s/time/effort
Maybe that's the key ingredient that's missing from LLM-generated text. Even when there is actual content there, there's this lack of surprise / uncertainty in its delivery that we require to stay engaged. We may not know what will be communicated, but knowing exactly how it will be communicated just kills our innate interest.
Communication expresses an understanding. LLMs arent communicating.
LLMs generate code well enough. Code has no natural language semantics. Programming's condition of satisfaction is that the machine executes correctly. The condition of satisfaction of communication is that the communication expresses what there is to be communicated.
In the limit where an LLM is trained to repoduce an exmplar of communication, it seems as if it does so. In that limit, LLMs can reporduce anything.
If frontier labs wanted to improve "LLMs for communication" they would pay for reasoning traces to be produced which reproduced the thought process of writers/speakers thinking about their audiences. Rarely do such traces exist in prior datasets .
While I don't like LLM voice either, I'd strongly disagree with this. LLM writing can be very tight these days -- sentences do not "go on" and are not "joined irrelevantly". I'm not sure when you last used an LLM, but it must have been a very long time ago; this hasn't been the case since late 2024 for frontier models, in my experience.
Obviously there's no underlying consciousness behind it all, but the verisimilitude is good enough for stimulating conversations on the recent research literature in my field, which I appreciate. For me, at least, the communication level is absolutely fine.
(YMMV, of course, and perhaps my bar for this is lower than yours.)
They seem to be getting better at the O(paragraph) level reporting on technical updates.
They're not yet at O(pages), and certainly no where near O(book).
For foreign text, sure. But I never was never able to read New York Times articles about some invention that started with Lucy leaving her home at seven o'clock.
If you have a specific question, reading LLM text can be more interesting than anything in the world. Even Wikipedia.
I've noticed this as well. When I get a bot to generate something I am happy to read it (although I often skim read it). But god forbid I have to read someone's LLM generated report! Come to think of it, if someone marks the text as LLM generated, especially after they have plonked a hand typed summary or sentence at the top, I am quite happy to read it. Maybe my beef is being expected to read LLM text the same way I would read high effort human prose? I certainly read LLM text differently to human text - you know the first sentence is going to glaze you and the second paragraph will just repeat the first paragraph so you can skim those annoyances.
I came to think of it, LLM output is not worth to be read like human written text. It's best to just skim it and move on. The problem is when this LLM output finds its ways everywhere on the internet. Being able to skim text requires knowing the context beforehand and knowing that you're dealing with LLM generated output. If LLM output that is published was labeled as such it'd probably make us less frustrated.
These aren’t deep AI hustle bro hacks, it’s literally just editing some text. If a big part of your day is reading AI-generated words, slap something together to make it do words gooder.
I find that it often isn't just too low, but also too high at the same time. It has a lot of information that is irrelevant, burying far too little relevant information for the work needed to extract it. I wonder if this is a side effect of attention mechanism, where the AI's ability to track information in a document is far less hampered by other information, leading to it organizing information to be both dense per word but any specific concept to be sparsely spread throughout the writing. I also wonder if this is why the order feels off, because the AI doesn't need for the information to be well ordered to the same extent that a human does, so the order of storing information is often overlooked when storing it into a document (and even information that needs to be ordered might have the ordering information separate from how it is ordered in the document).
Statistically even or repetitive over a chapter, chapter after chapter. So make an exaggerated point prove it come up with a conclusion and then have an exception.
Repeater 4 to 6 times a chapter in a book of 45 chapters.
So I wrote something to remove it and yes it is better but still not very good.
But what happens when everything is persistently, subtly wrong? When we cannot smoothly map the text to a simulation? An inner-sim keeps throwing errors and mis-predicting what comes next is unnerving and draining. Like talking to a human who is not currently sane.
Sometimes I imagine LLM text like a plastic animal model put out to confuse local wildlife, where it hits just enough instinctive triggers that they try to mate with it before leaving in frustrated bafflement.
______
Half-Life 2, Dr. Kleiner: "Never fear, Gordon, she's de-beaked and completely harmless. The worst she might do is attempt to couple with your head. Fruitlessly!"
Which requires a bunch of things an LLM lacks, such as theory of mind, knowing who it is talking to, and willingness/ability to engage in short turn conversation to assess and steer the conversation according to the feedback it is getting from the person it is talking to.
IMO, it's still better to not let an AI critique your writing style at all, only the facts and claims you're making against its context / MCPs. Your own writing style is a reflection of your mind's thought process and, especially on forums like this one, is the only way to imbue your ideas with your personality.
Now there's an idea. LLMs may have taken the em-dash from us, but the em-en dash is unknown to Strunk & White and might be human-exclusive for some time yet.
Us oldies learnt to type these with multiple dashes automatically, as unicode hadn't been invented yet.
-- - --
The Darude-Sandstorm dash
I know this because the topic is usually an obscure issue, and I have already gone back-and-forth with multiple LLMs to see where their boundary of knowledge is on it, and it will often include hallucinations along the way (e.g. pointing to old/nonexistent settings pathways)
So when the first reply from the official 'support' channel contains the exact same nonexistent settings pathways, it is easy to conclude they are sending me the LLM output without doing the verification steps that I have already done on my end. When I point out what doesn't exist, the reply is always a similar flavor of the usual LLM "you're absolutely right" response after being corrected.
I'm all for having these bots process inquiries if it means a speedier resolution, but, please, anyone out there using these things, turn off the simulated sympathy. It's infuriating how condescending it comes across.
(yes I understand this is likely not legal or practicable)
A very important point, which was also made more forcefully in "Human Routers of Machine Words" by Fernando Borretti, shared on this site previously. It's good that Fowler brings up a very legitimate use case in the next paragraph, though.
Automating creativity is the death of art.
The typewriter was the beginning of the end of the art of handwriting.
Factories was the beginning of the end of hand-created goods.
AI has killed the art of handwritten code (which itself automated away other forms of art).
We should actually restructure our organizations/society in such a way that this stops being necessary. If no one reads a report, it was almost certainly a waste of effort. The only possible benefit was for the report-writer (which of course is also flushed if the "writer" uses an LLM). These things were already a plague before LLMs.
I always had a high faculty for language, and when I was young, even very young, I was able to speak in full sentences much like I'm doing now. What I didn't have was context. I didn't know how to relate to people socially, and then later on, I wasn't as experienced an engineer as my peers, but I was trying to act adult and mature in both ways.
So I tried to write in ways that were catchy and unnecessarily complex, because I thought that conveyed sophistication and intelligence, and that I knew what I was on about more than I did. It's the kind of writing that gets you through a 30-page final paper for a class you don't really like, or a research project you didn't finish. The term I often use when complaining about my colleagues' LLM output is "flashy but meaningless."
As my confidence grew, I started writing in simpler (though still not the most concise, that's just not who I am!) ways, because it became easier to convey ideas simply in a way that made it obvious I understood them.
When you think about what LLMs are - word machines without (or with very limited) context - you can see why they write the way they do. They are always acting; trying to replicate a human contextual experience with only weights and tokens to carry them through it. That fact is part of why I am skeptical of their ability to suprass humans in reasoning based on their own output.