LLM Voice (martinfowler.com)

How is someone "aggressively pro AI"? This is the first time I read such concept.

Gentle reminder that most AI token spend is used for TPS report type writing not Cutting edge fiction writing. Creative writing, or general human-to-human communication was never really the target. It's not surprising the first majority use of this was code.

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.

I suspect it’s partially an average / regression to mean.

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.

A model trained to minimize statistical prediction errors is going to sound ... predictable (aka boring). A (good) human author tries to keep their writing interesting, while an LLM strives for predictable instead.

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!

Don't these sort of conflict with being "aggressively pro AI"?

> 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.

Do you have some personal financial or other incentive to be “aggressively pro AI”?

The part I find grating is the attempts to punch everything up. ("It's not X. It's Y"). Like, its a react badge component, it doesn't need punching up. You didn't find the smoking gun in a murder mystery, just a misplaced CSS rule.

> I've also noticed how awful everything is to read lately.

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 got the same bright red line for 2023. E.g. I treat a handful of youtube channels with long electronic jams almost like oldies at this point - keep listening to the same stuff from at least two years ago because it’s the authentically good stuff.

LLMs have made the internet unreadable. It used to be just Google search results were SEO-gamed empty prose, but now it's every blog post, news article, product announcement ... everything. I've stopped reading basically everything online except user comments on a few websites. It's the last place remaining to actually connect with other human minds, but even that seems like it won't last forever.

to AI verbosity and weird use of diction is natural, to humans it makes it hard to parse. The number of time i start reading long winded AI written design docs, and i just can't get into the flow after reading a few sentences and have to keep re reading certain sentences.

Agree but also want to point out that this started with the SEO boom of the last decade where sites started rambling so they would keep you on the site longer and get more SEO "points", they were articles meant for a robot to catch humans.

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!)

> sounding identical

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.

This applies to code btw, it's so much nicer to read a human made pull requests. It's also much faster as people tend to be lazy and produce denser code.

Personally, for blog posts, documentation (unless I flat out vibe code the entire thing) etc are all written by me by hand. If the model can save me hours of "oops that was really dumb of me" why can't I take time to read and document things by hand? I think everyone's embracing AI in the wrong ways. Maintain your knowledge of things, and keep the documents nice. There's still room for letting your agent manage agent specific docs, but for the humans, just write it yourself, ask the model for feedback and tweak it by hand. Let the machine talk to the machines.

> it's a mixture of everything sounding identical and everything being a rambling unconcise mess.

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

So, you're still aggressively pro AI even though you recognize how quickly it is destroying the quality of online content?

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.

100 percent agree. It's like the consensus for tone and vocabulary has been "midlevel digital marketer from 2015" and now everything is starting to sound like a lazy sales pitch. I'm with you on AI but man, this is such an annoying side effect.

And yet you remain aggressively pro AI? Why?

Honestly I think it's a good thing it's so hard to get AI to talk normal and I hope they keep making the voice more and more distinct. Agents should run as background tasks and automations. They should be understandable so they can communicate with you about what they did, and for observability/ai safety. Im perfectly fine reading llm speak when it's an ai telling me what work it did. When in reading a blog post with a human listed as the author though, I do not want to read an AIs thoughts that the human is laundering. Honestly I'd probably tolerate it if it was framed as "I asked my Claude to do a cool thing here is its report on what it did." But in general it's really nice to have an instant filter of - this post is a waste of time - ah this post has an actual human's thoughts

This is true across every topic. Anything at all that someone can put into an LLM and get out an emoji-laiden wikipedia article sized post to slap on a heavily SEO-ed website is absolutely flooded with it.

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.

Sorry, meant to reply to your comment, but my post above...

"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.

Maybe stop being pro ai then.

My guess is that tokenization is typically not (strictly) syllabic.

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.

My model is about gaps - all ideas start with gaps. LLMs can fill in these gaps quickly - but some ideas are just bad - so the filling in does not improve them. Maybe it would be useful to treat that filling in as a kind of hallucination - the LLM trying to meed impossible requirements. But on other cases the ideas are good and the LLMs fill in the gaps in the right way. It is a kind of skill to see when LLMs are hallucinating in the gaps. I use LLMs very extensively and I am adding checks to that should catch when the filling in is weak - but it is still not fully autonomous.

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.

It's (at least) two things I think; one is that thanks to the internet, the ability to focus on things for long is shot - I struggle to read long form content on the internet, generally speaking. I also did my whole career with Google and skim reading documentation looking for solutions for an immediate problem more than getting a wholesale end-to-end understanding.

(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.

I can imagine that one of the aspects of being social creatures is that we are drawn to the entropy/surprise factor of natural language produced by a brain that isn't our own. Deriving a shared understanding without being able to precisely model the language output of someone else is like, the fundamental substance of conversation.

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.

I know what you mean. The latest models are _better_, but there's still something off about the prose they generate. Obvious tells are mostly gone now, but some internal spider-sense still makes me narrow my eyes.

One of my favourite rules: perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away. This is invaluable to writing. So much text from an LLM feels superfluous, like I could've said the same wall of text with just a few sentences. Writing is tightly coupled with thinking: refining thoughts, arguments, ideas, discarding them, mulling over them. LLM text always raises this nagging question in my mind: am I reading what you think, or what your agent parroted back and you accepted?

>its information density is often really low

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).

"I just can't seem to stay interested and read LLM text."

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.

Sorry to single you out, but it’s surprising that this is such a consensus opinion around here. If an LLM isn’t giving me exactly what I need in the format I want, I tweak skills or system prompts until it does.

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 also started to just abandon writeups obviously written by LLMs. They feel (and usually are) just cheap marketing/selling posts. Not to mention why would I read stuff nobody wrote and likely never read themselves? There's little value in such content. I don't even point LLMs to such websites to get a TLDR. Part of why reading is interesting is the author, their history, beliefs and style. In the world of LLMs content, the author is some undefined diffuse alien with a made up personality.

> All human communication is based on internally simulating the presumed mind that was/will-be at the other end

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.

---

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.

I see many Youtubers who are certainly still writing their own scripts, but every once in a while you can tell they absolutely used a Claude correction in one place.

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.

Potentially dumb question, but could you expand on how you use LLMs as writing critics? Do you just ask them to flag bad writing? I'm asking because I don't think i'm skilled in writing.

Tangential rant: I find myself getting viciously angry when these support bots lead with, "I'm very sorry to hear that. I understand how [rehashed summary of complaint] would make you feel frustrated." Mother ****er, you don't understand what it means to feel anything.

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.

It could be argued that automating hand-crafted works — aka pretty much anything that is automated — 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).

[delayed]

You’re not talking to someone from Japan, are you? If so perhaps they’re using an LLM to translate?

It's no mystery: solving LLM voice requires making the model worse at other things they value above all other things

I highly disagree, though I admit I am one of those towards the top of literacy and cognitive skills. I absolutely enjoy when someone uses LLMs to learn. However, I don't think that's what we'll be seeing.

Look at programming. What's happening there is we are having agents produce 10-100x the code to where people give up on reading it and just hit OK. Look at math. Plenty of people there want to learn from the LLMs and all these crazy new math proofs. But those are the people who will probably be out of a job soon because the money in math is in producing proofs, and that's now an LLM-exclusive job.

I think LLMs are weird in that the first order effects are those of an incredible and amazing tool that anyone can learn things from, but the second order effects are going to be the exact opposite - people giving up on bothering with ideas. Even at my job, which is one of the few that has not gone crazy with AI, anytime I ask anyone a question, all I hear is "have you asked Claude?". There's no communication anymore, no pushback, no struggle. I worry that one day we'll look around and realize we've built an entirely separate second civilization of machines and robots that acts more like a neighboring country than a tool we can use to build humanity up.

-Just my rambling thoughts, not refined by an AI.

It is not any different from knowing the difference between good writing and bad. I can't help it that I have taste and preferences. I can avoid telling you yours are wrong. But I don't think this is a minority position and nor is it different from people who reflexively hate AI "art" or music

(I have to admit, I truly do like some of the music, so feel free to scorn me. If you want to be infested with a vicious ear-worm check this out: https://www.youtube.com/watch?v=zONGVaONKe8&list=RDzONGVaONK...).

Add it to your system or developer prompt, and it still can be.

Given the amount of angst being expressed in this thread about particular LLM-isms, I wonder why people don't simply prompt against the traits they hate. Yes, it will reduce output quality very slightly in some circumstances, but if it really gets your goat like this, surely it's worth it.

I read a lot of old material from 1-1900AD where I can pick upon the author's 'voice'. There are subtle clues, like if they had a regional accent, where they use a specific word or type of cadence.

AI does the same with responses where the cadence is some type of formula that becomes identifiable.

Each of the big models has their own but an example of the cadence is

A) Agreement or Acknowledgement with the User

B) Specific detail gets a paragraph of detail related to the previous response

C) Overview of the conversation's current state

D) Prompting or Direction for this conversation to go

E) Landing on a beat. Wrap up the vibe. Recap with a zinger.

I cannot unsee it and a few of the models are annoying. I have attempted to build out the Agents and context to avoid the 'last beat one liner' without success because after 2-3 more turns, I get "That challenge was accepted and you stepped up. You're the winner this round".

There is this article on this topic, which I found quite interesting: https://arxiv.org/pdf/2610.29162

Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation. [...] We propose and test a hypothesis as to why: while human language is optimized for successful communication, AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for information density rather than successful communication.

I'm hanging on to the typewriter em-dash. They'll never take it away from me!

This will matter more when code quality starts to matter. Lately it hasn't. I hope it does at some point and LLMs could help there too if we only cared about this metric. And lastly, people need to remain the loop.

One of the main differences is that LLM-generated code can be tested. You can prove that it does what it's supposed to do — and if it fails the tests, you can tell the model to go back and fix it. So there's an objective criterion that can say "this is acceptable" vs "this is unacceptable". Now, it does rely on your tests being good, and not everyone has written those good tests. But it exists.

There are no objective criteria that can say "this text is good" or "this text is bad". Sure, spelling and grammar errors... but we're talking about writing style, "voice", and whether the text is easy for humans to read.

Now, there are plenty of problems with LLM-generated code that can't be solved by having good tests. Is it maintainable? Is it written in good coding style? Those are harder to quantify. But since LLMs are such good mimics, if you tell an LLM to add code to your existing project, then it's usually going to do a decent job of matching your coding style. The question of maintainability still rears its head; my personal experience is that the smaller the chunk of code the LLM generated for me, the more likely it is I can understand it fully and therefore maintain it in the future. An entire vibe-coded project? Good luck maintaining that. A 100-line class with five 20-line methods with names that make sense? Yeah, I can take that and maintain it in the future, even if the LLM tool goes away.

But mostly, the difference between the acceptance of LLM text or of LLM code, IMHO, boils down to "are there objective measures that I can use to accept this"? Code has to work, which is a much higher bar than text (which just has to have correct spelling and grammar), and therefore it's easier to get the LLM to spit out decent-ish code than decent-ish text.

Even more importantly: it's not an either/or situation.

You can love to cook and enjoy every second of it, but sometimes you just need fuel and that $3 microwave burrito is sufficient. Just the same, 99% of your means could be frozen TV dinners, but that doesn't mean cooking a romantic meal for Valentine's Day is off the table.

There's a weird level of fear mongering occurring around AI that it will make you incapable of independent thought if you use it. People will throw around very light research that kind of points in that direction, but they'll never highlight any of the nuances in such an argument. You can use AI and still "think" if you choose to, just like how you can avoid AI and still opt out of "thinking" if you choose to. It has less to do with AI and more to do with the context of the individual using AI.

But isn't one problem here that it's not just personal consumption? People bemoan the proliferation of food-service (Sysco et al) "slop" at restaurants. It seems like LLM writing for others is serving that $3 burrito for $15 without even trying to doctor it up. If you're not willing to put in the effort to think about and produce your writing it's probably best to keep it for personal consumption.

Probably he should have actually said "what's the point of writing then?", because there isn't one. Just like food is fuel, writing is thinking.

Analogy: it’s fine to write stories for yourself mad libs style with AI

It’s not ok to write stories with AI and post them for others to read claiming “I did that”

Get the difference?

> to some people, food is art, creativity, life to others, food is fuel

right. he is talking about the first group. where did you get the 'shame' part?

I like your analogy. Microwaved empty calories perfectly describes llm output. You are going to pay the price sooner or later.

This analogy looks quite good. Heavily processed food is not heathy for your body and heavy use of AI and other dopamine shortcuts like social media are not healthy for your brain. You have to approach both with causion.

I think the idea is if you are going to share LLM writing then why not just share the prompt which is shorter and has the message you want to convey anyway.

That is not to say that I don't think there is a place for LLM assisted writing. Writing is an exercise for the author and translates to oral argumentation so I don't think that fades away. But LLMs could do some "audience normalization". What if I could make a blogpost that is dynamic based on experience of the reader? There are lots of opportunities, but versus other fields where optimizing for laziness can be beneficial I don't think lazy writing currently has much value add.

Largely speaking eating is an activity that only involves yourself. Whereas, writing and thinking are largely activities to communicate ideas between yourself and others. Basically it matters more how this is done.

Sure, let's do food analogy..

It's your turn to host supper club this week and you serve frozen diners on your nice tableware.

I’m fine with my neighbour eating TV dinners, but if they invite me for a meal I expect to not be served a TV dinner, or to at a minimum be informed beforehand that that’s what going to be served.

Same thing with LLM-assisted writing. It’s one thing to do it for yourself, quite another to share it with other folks without warning.

I agree. I honestly think convenient food is perfectly fine for people most of the time. Recently I have been playing with "eternal stews" which I guess are logically the opposite of fast foods, but at the same time are focused on availability and convenience over intentionality and mastery (at least the way I make them).

Most food is literally just fuel. I think that is fine. And at some level, eating 3$ burritos for a quick lunch on the daily really makes you appreciate the depths of a home made risotto even more. :]

To stick with that analogy, you also wouldn't serve a tv dinner to a guest unless you were somehow incapacitated. You certainly wouldn't do it if you were running a restaurant.

Add to that the fact that AI is significantly worse for you than TV dinners, and has an outsized harm to humanity, and the conclusion is plain:

AI should not be imposed on others.