I don't want to read what you didn't write

isn't it more like: i have 1000 bits, i transfer 1000 bits but depending on the person, it might be lossy, so they only understand about 700. they then come up with the 300+- on their own, potentially putting them over 1000 or they come back and ask questions to fill in the blank. the bits don't ever have to be bit identical.

Depends. If the 700 bits were arrived at by the LLM while spending a lot of tokens, and the result is "good", I may want it through you as a middleman because it used up your tokens and won't eat my subscription usage limit to ask the AI to supply those 700. If you spend the tokens and put the result online, plenty of people can spare their tokens because they don't have to ask the AI to derive it. Bonus if that result was run through some kind of testing and verification.

Obviously this doesn't really apply to super simple questions that the LLM can just spit out the answer to right away.

For a long time after the internet arrived on the scene, a lot of online news stories would reference websites, papers, polls, etc. without linking to them. There are still news sources doing this today. Sometimes such articles interpret or place context around their hidden references, but a lot of the time they just summarize.

Giving someone the text output of a LLM is very similar to publishing a summary without links to the referenced material. When you were querying your LLM, you could have asked specific questions or asked for a custom focus or point of view. Your intended audience might have questions or different concerns, but they're unable to interact with your LLM. What you have delivered is static and unresponsive. It has all the disadvantages of being machine output without the advantage of being interactive, the way your LLM was for you.

It may have to wait until compute is cheap enough that tokens are essentially free, but we need a system to pass "hyperlinks" to LLM's primed with context, ready to be interactively queried on a chosen context. It's being overly generous to assume that people are putting even 300 bits into a LLM for every 1000 bits of regurgitated writing they try to pass off as their own. When people post LLM output as if it were their own, I have no choice but to assume they had zero knowledge of the subject, but this query taught them what they wanted to learn, and now they're sharing that. That's fine, but please pass an interactive LLM link rather than static text.

Once we have "hyperlinks" for LLM sessions, perhaps we can share LLM output a little more usefully and honestly.

I love this example, partially because it jives with my conviction that LLMs are the ultimate translation machine. Ever since the embedding model days, it is clear that these models are amazing at representing meaning as math. The fact that LLM's most salient use is for coding somewhat agrees with that. After all, what is a programming language but another language? We instruct people with words and machines with code.

I will give you a use case where this is absolutely not the case.

I have a bunch of CLI utils I run for various clients and their peculiar setups. They now have man pages with descriptions and examples in them because the LLM went and read my code and did the needful.

I no longer have to re read my own code, rather I can just use the manual page.

Format and description came from semantics and context that (barely) existed elsewhere and I was not going to retain or transmit, but I have now.

I was going to write a post disagreeing with this on the basis of the fact that the reader lacks the background information the LLM has. For example, if I were to prompt "explain the proof of quadratic reciprocity using Gauss sums" most readers would need the entire LLM's answer (and much more, probably) and not just the prompt.

But then I realized that the reader can prompt the LLM with the same prompt for the same or equivalent expanded text. Most people don't do this as it's extra effort, but it's interesting to imagine a world where this is the default way of engagement with a text, assumed by both writers and readers alike.

I don’t quite think this tracks. Perhaps you want to communicate 1000 bits that are well known and can be referenced with a 300 bit key. Then the LLM can easily retrieve the remaining information. It’s like sending someone a link to the Wikipedia page instead of explaining something yourself.

No, I don’t want to read LLM writing because it is BAD at it. It doesn’t really understand how humans think (because it thinks differently), and doesn’t seem to understand core principles very well (presumably due to the lack of world model), so it can’t write something humans enjoy yet.

If you have 1000 bits of semantic information that you want to transfer but your default communication combines it with 10,000 bits of noise. Giving it all to an LLM and iterate on reducing that noise while making sure the 1000 bits is still present would enable you to communicate more effectively.

Overall, ideas are ideas. I'm not overly concerned with the fact that it was you who had the idea, as long as the idea is interesting. I don't know most of the people who write the things I read, so it seems to be of no consequence to me at all if they wrote it, as long as it is interesting. LLMs are notorious at creating things that are bland and vacuous, but they by no means have a monopoly on it.

Be the source human, machine, or dolphin, if they write a good article, I'm prepared to read it.

I like this example and it made me think, there is an analogy here to spec-driven development and vibe coding.

Rather than send 300+700 bits, like you said, send 300 (or less!) and let the human intelligence on the other side generate the result. Which supports the even older perspective: “If I had more time, I would have written a shorter letter.”

I’m not sure if this lands on anything very profound, but what about a pattern where, instead of codifying agent output at all, the only artifacts we share are the prompts. And the rewards (respect) accrue to those who generate the most generative among people and AI

The value of writing isn’t always to communicate new information. It’s often to align everyone’s assumptions. For instance, when I say casually to a colleague or an agent “this change will require a db migration” they understand it’s to my teams primary application database. If I submit a design doc to a company wide review which database is changing is critical information. If the 300 bits were truly enough your agent or junior engineer would implement the wrong thing correctly as they often do.

What already happens: - People give an LLM a bulleted list of points that they want expanded into a professional sounding document. - The receiver doesn't wanna read all that. They put the full document into an LLM and ask it to summarize it into succinct bullet points.

Saying something for years doesn't make it right. Has nobody ever challenged you on that? Our brains prefer to read enjoyable bits, not raw information. If you only got those 300 bits, you aren't going to bother putting them into your LLM to generate 700 bits to make it a more enjoyable read, you're just going to struggle through the 300. Maybe when browsers come with built-in automatic text puff-uppers, then you'll have a case but almost nobody does that now.

Here's a clearer example - would you rather learn a concept from a research paper or a textbook or blog? You say the research paper but they're dense and hard to wade through where-as blogs and textbooks are more wordy but hold your hand, which is something that helps humans learn.

It isn't the transfer of information at all. What's actually happening is you're prompting experiences in a human instead of an AI using text.

Communication only works if you have multiple levels of representation and abstraction, including but not limited to - letter shapes, grammatical structures, style and register, stylometry, and subtext.

All of that is learned, and writers usually assume they can rely on that learning as the context for the text.

So you don't write to 'transfer information' like a network cable, you write to trigger experiences in the human version of latent space.

Factual information is one kind of experience. But even when that's the goal, there are always layers of implied relationship, social register, role, status, and other implications in everything that's written.

In normal communications the context - business emails, personal messages, mainstream journalism, fiction, and the rest - defines what acceptable language looks like.

The content fits inside that. But it has to fit the context, otherwise it lands in a semantic and psychological uncanny valley - like sending LinkedIn speak to a spouse on a wedding anniversary.

The real problem with LLM writing is that it's good at the technical layer - the grammar and spelling - and has some insights into the rest.

But the default content style is marketing and ad speak. And recently it's developed a weird and unique hybrid style which applies marketing fluff and pretension to technical content like code comments.

So you get one register instead of all of them. It can attempt others, but it's still too limited to generate them fluently. Sometimes the results are outstanding, but often it defaults to mechanical clichés.

So that's why it sucks and sounds so hollow.

Can it be fixed? Yes, but it's very hard work, most people don't have the skills, and it takes time - often too much time to be worth the effort.

It's worse than that. If it takes others longer for others to consume and understand what you're producing than it does for you to produce it, you'll never be able to communicate with someone efficiently. Communication breaks down at a fundamental level if if you can't keep up with the other side if outputting and they won't slow to allow you do do so.

I see this all the time now with LLM generated output. It's easy to have an LLM generate a chunk of content that can be dropped into a chat or comment, and when it took you 20 seconds to have something written up based on the shared understanding you and an LLM have about the context of the situation, but it takes other people 3-5 minutes to read and understand that content, that fundamentally doesn't scale. It's bad enough when one or two people are doing it, but if the whole team is doing it, the only way to keep up with the stream of information is to also consume it through an LLM. At that point you're likely to be missing much of the nuance, and the amount of errors will explode.

This can be alleviated by people reviewing the output of an LLM and making sure it both includes fundamental information that might be assumed by context and reducing it to the parts that are essential for the new context it's in. This takes time, but is extremely important.

Having an LLM write gobs of text to send to other people instead of doing it yourself is the equivalent of a low yield cognitive zip-bomb. Don't do it.

If only it were 300:700... it is more like 300:100000 and up. But then you use AI to summarize to hopefully retrieve the 300 bits. We should try doing this in a loop with a simple enough sentence to start with and see what pops out after every iteration. Fully automatic Chinese whispers.

I’m not disagreeing with your preference for human authored writing, but I don’t know that I agree with the content of your argument.

If you consider that what humans are doing during conversation is a form of compressed encoding / decoding from some latent representation through a quantized signal then if you interpret it that as a compressed sensing problem you absolutely can infer to a very close approximation the original latent representation using far fewer than those 1k bits.

While I generally agree, the difference is I can also hand an LLM my pile of code and documents which is... a lot more information dense than basically anything I could write. Sure, a one-off prompt in a chat window isn't very helpful.

That would be true if there is only one receiver and if i could exactly predict how the receiver would interpret these 300 bits. But i could use LLM to expand these 300 bits to 1000 bits of more explicit information that gives enough redundancy and context that minimizes risk of misunderstanding and increase easiness of understanding, then validate such expanded message (and possibly do iterative corrections if LLM did not understand those 300 bits correctly), and finally release expanded message to (potentially large and heterogenous) group of receivers.

As with everything, it depends how you use it.

I’ll often put a long stream of consciousness on the page, or jot down rough meeting minutes, then ask ChatGPT to “summarise this for an email”. The result is shorter, clearer and easier to read.

AI amplifies the habits of the person using it. If they’re lazy or dim, then it's like giving a monkey a gun.

So... We should only communicate in compressed formats?

I agree. But there's still a valid use here for LLMs: If I have 1 Million bits of information and I want to communicate a synthesis of 500 bits to you, LLMs can be a viable helper. It is rarely done that way - I agree, but the way I use LLMs: Drop in all the relevant context information (PDFs, HTML, Markdown, Pictures etc. - up to usually 200-300k tokens), then compile a synthesis/summary prompt (usually 1-2 A4 pages of handwritten text. Then copy the output (1-2 A4 Pages), manually edit and send off (often, including an archive of the full original conversation, so the human on the other side can consult the unfiltered original prompt + inferencing, if wanted).

In other words: If I didn't reduce the 500 pages of text for you, you wouldn't know what I mean or what is relevant, or how to filter it yourself.

"Writing is fundamentally the transfer of information from your brain to my brain. If you have 1000 bits of semantic information you want to transfer, you can't give 300 bits of semantic information to an LLM and have it fill in the remaining 700,"

You absolutely can if that information is in the code, which it often is.

There should not be that much in the code that needs further elucidation.

Some stuff definitely - but not much.

Usually you need the code and architectural summary + that stuff.

The AI is not very good at it but it will get better.

I think the debate here is about a few different things.

Yep, I wrote about it here: https://blog.gpkb.org/posts/just-send-me-the-prompt/

I notice this in the attitudes of students towards reading. They will take a 10-page journal article and ask the AI for a summary and then read a summary that's the equivalent of maybe half a page. But if the article could have been half a page, why is it actually 10 pages? It's true that there is some boilerplate, but it's strange to me that people could think that 90% of what they're reading is (to use your phrase) "not true semantic information". It's like if you went to a restaurant and ordered a 10-oz steak and they brought you a little teeny bite of steak and said "Oh, other places will give you a bigger one, but most of that is just filler, we just took out all the superfluous parts." It's a worrying sign for our future if things like this are not tripping people's skeptic sensors and making them wonder if they might possibly be missing something.

you think like such a techie. consciousness is not analogous to IO. if writing were only the transference of something from point A to B then what is the technical function of poetry, a question (the open ended kind), a pondering, a wondering, and so on? Furthermore, language is lossy and introduces a large degree of subjectivity, mystery, and uncertainty nomatter which words you choose. Words are by nature lower res/on a lower ontological domain than thought. So your premise is preposterous on both a practical and theoretical level.

Furthermore not all writing is for another to consume; nor even for the author themselves to consume. That is to say it has meaning ipso facto, not dependent on transference, as ritual.

Communication between people is much more complicated than that.

I get vague statements thrown at me with people expecting me to understand it.

Same with writing, setting up whole context to properly transfer 300 bits is always orders of magnitude bigger then just additional 700 bits.

If you read anything you're reading superfluous bits.

Saying the hard part out loud: we give more value to the people with 1000 bit brains than we do to the 300 bit people. There’s a strong incentive to write like you’re the former and not the latter. Freely available LLMs give people the means to match the incentive.

That's assumes a couple of things that are trivially not true:

- The assumption that both parties know about the same as an LLM does. An LLM know orders of magnitude more.

- The assumption that the output of the LLM is not refined over a few cycles.

The point is that you might give 300 bits of semantic information to an LLM, it fills it to a 1000 with perhaps 400 wrong bits. You correct it half a dozen times. It's now 950. You do the final touch ups. It's now at 1000. And it still took you 20% of the time to do it.

> You might as well transfer those bits to me directly, rather than having the LLM add on an extra superfluous 700 bits that I then have to filter out.

Exactly this. Just send me the prompt! ;)

I’m not sure the 300-bit → 1,000-bit framing applies in all instances. The 300 bits may be a compressed cue to a much fuller idea. The AI can combine that cue with its prior knowledge to help reconstruct what the prompter was trying to express, with the prompter then verifying whether it’s right. Without the relevant prior knowledge for reconstruction, or the prompter for verification, it becomes much harder to know whether you’ve reconstructed the intended idea.

This is a great analogy, thanks.

That means folks using LLMs starting off with 300 bits KNOW that they lack the full payload of information to transfer to you. IOW, they know they need to transfer much more than 300, so they use LLM to fill those gaps. That's the crux of the slop universe out there. Folks are using LLMs for the 700 bits on top of their 300 bits and passing off the full 1000 bits as their own.

I echo the writer's sentiment. "I don't want to read the clanker's 700 bits. I want only your synthesis." (I can get the clanker to generate those 700 myself. Unless ... unless this whole LLM slop market is all about saving you the time to get an LLM to generate those 700!)

Just use LLM to filter out those extra 700 bits, and you'll happily end receiving 100 bits or original information and 200 bits of slop.

> Writing is fundamentally the transfer of information from your brain to my brain. If you have 1000 bits of semantic information you want to transfer, you can't give 300 bits of semantic information to an LLM and have it fill in the remaining 700, because it doesn't know what those 700 bits are.

Sure you can. LLM doesn't know what those 700 bits are, but you do. You may not realize it, and may not even know it at the time of prompting, but you do by the time you're sending.

Typical case is like this: you have 500 bits of semantic information to transfer. You give 300 of them to LLM, and get back the 500 bits you knew you have, and extra 500 you can quickly confirm are correct and relevant. Some of them are just dereferences of your input - where you recalled a pointer, but not what it pointed to. Some of it is information you never had before, but are able to easily validate.

You send that to me. I likely immediately realize the message was AI-assisted, but I trust you to be a decent human being, and not an asshole that lobs unverified LLM vomit over the fence for others to deal with. End result: you communicate 1000 bits of information to me, instead of planned 500, and you yourself learn extra 500 bits.

This is the optimistic scenario, but it does happen when LLM operator is not an asshole.

(Excuse the strong language, but I spent a lot of effort every day on both dealing with inconsiderate people lobbing LLM output at me, and making sure never to act like one myself, so it's a topic close to my heart.)

I have 300 bits of information in rude direct form with some obscene language as well. I ask LLM to wrap it into a nice polite message. Resulted 1000 bits are sent in the email.

I bet no one would like to get direct rude "source" instead.

>If it's able to guess those 700 bits correctly, then they aren't true semantic information, and you really only have 300 bits you want to transfer.

A math teacher only has "this class is about math" to transfer, the rest is known. ;-)

Jokes aside, I think we're in a weird transition now where AI is used to generate text that looks good but is bad. In a few years people will know that and be more critical.

I think we went through a similar phase when DTP had it's breakthrough. Suddenly school papers were laser printed 300ppi times new roman and got more attention than better papers written by hand. But eventually that became the baseline.

I think Ai will make it so that well-written texts with clarity, good layout, correct illustrations, callouts etc become the norm, and will no longer impress anyone unless the information itself is actually good.

And if the information is good, it won't matter if it's AI generated or not.

> you can't give 300 bits of semantic information to an LLM and have it fill in the remaining 700, because it doesn't know what those 700 bits are. If it's able to guess those 700 bits correctly, then they aren't true semantic information, and you really only have 300 bits you want to transfer.

It doesn't actually follow, because maybe the LLM is smarter than the original writer (at least in the domain the writing is about) and hence really is able to complete the ideas in a way the writer can't. As an existing example, consider formulating a conjecture and having an LLM prove it. But I agree; if I wanted to read an LLM's output I'd simply ask it myself rather than read someone's supposedly-human writing.

I think there are a lot of situations unaccounted for here

- people writing in a non native language

- people insecure in their writing

- people not used to writing in industry terms

- people with the curse of knowledge that are aware that they can’t write for a general audience well

Surely others too. None of those mean you have to read it, but I have gotten immense value from reading some things people have had ai write (and I’ve seen a ton of junk as well)

I don't think that necessarily holds, but it needs nuance. I tried to convey this within my company by giving a "guide" as to how to use LLM's for writing, across three modes:

1. Transliteration - roughly keeping the number of characters or bits, but translating to a different lingo, language or mental model (e.g. metaphors). Roughly the safest mode, but can still yield catastrophic results - it's safest if the author still provides taste and editing.

2. Compression - taking out redundancy to make the text more dense and more salient. The LLM chooses what to take out - and might take out the wrong things. More dangerous - but if you're happy with the salience and you believe the reader won't have time to read the uncompressed - it's probably safer than having the reader LLM compress without the benefit of your editing process.

3. Decompression - using the salience of your idea to add detail to the reader who wants to understand it fully, by utilising knowledge that is common to you and not common to the reader. This can be very powerful when there's no time to fully write the thing by a human - but it's the easiest to get wrong and to create slop. As an example - you could try explaining concept X + illustrate it through 3 examples. You know the examples are in public memory and easily retrievable - so you write your explanation of concept X, list the examples you want - and the LLM can take all of them, synthesise and bring the full package from your 300 bits to 1000 bits.

You are right that those are not the exact 1000 bits from the original brain, but they could contain 900 of the 1000 - which is still better communication efficiency than transferring 300.

I am however, more and more in the camp of fleshy brains writing everything, as my slop allergy rises.

How do you explain professional speech writing?

> If you have 1000 bits of semantic information you want to transfer, you can't give 300 bits of semantic information to an LLM and have it fill in the remaining 700, because it doesn't know what those 700 bits are.

LLMs do inference or computation among other things, so the remaining 700 bits can be something like that. The hidden implication in your claim is that computation adds no information content, which leads to an interesting philosophical discussion.

So for example, if I ask an LLM to give a proof or derive a new theorem from a set of axioms, according to your assumption, if it answers correctly, then I haven't learned anything new.

I am not really sure how to resolve this paradox in information theory.

I like your premise but I will point out that communication isn’t just about conveying information.

It is often about persuasion and that sometimes benefits from framing effectively, which I think an LLM can help with given the key points you’ve got.

So I spent a full workday investigating an issue with an agent and then I spent another hour prompting the agent to summarize the findings so that I can send them to a colleague. How does what you're saying come anywhere close to describing my work?

Correct.

The real benefit is that the use of an LLM allows me to convey to it 1747 bits of scrambled information in an order that fits what's actually inside my head, and then have it unscramble that and convert it down to the 1000 you need. It can do that better and faster than I can.

It massively reduces the time it takes to make a short letter.

No, thanks. Depending on the type of information being transferred and just how uncertain the recipient's knowledge is, I may or may not need to add some amount of information to whatever I have to share. That information would be extra context you may need to have, on top of what new thing I am trying to explain. If I know I'm talking with someone who's all caught up, great, no need for context setting. Though even in those cases, there might be some terminology I want to define, or to provide a list of acronyms or whatever. That is boilerplate work, but is at least useful, if not needed. I know people tend to view this sort of comms as "technical" - they shouldn't be. So much misunderstanding happens because people are just not aligned on their shared context or medium. Of course, AI is not fully reliable to generate that context, you as a writer need to put in the work to review and align the result with your own understanding, or drive it thoroughly before.

All that being said, I acknowledge that people (me included) love to be sloppy in their comms (with or without AI) and then blame others for misunderstanding. It's also unlikely we'll change soon. What can you do.

What if the bits I'm transferring are "I don't remember the precise syntax, but here's a link to the relevant documentation."

I think you need to take information compression into account. Compression naturally created by shared understanding of concepts and acronyms.

Can you explain to a child what it means when "GitHub is down" with less characters?

While I acknowledge that this is meant to be framed in the context where the message is fully compressed, I think particularly when the medium is language (usually pretty compressable), then actually anti-aliasing the transfer of information is still 100% relevant! After all, AA is really about making better use of the information that you actually transmit. Compressed or otherwise, 1000 bits of aliasy garbage is not in any way the same as 1000 bits of perfectly antialiased signal.

I like to think of this idea of transfer of information (via words, let's say) from speaker/writer to listener/reader, as actually a fairly normal signal processing situation. One where the constructor of the message during it's 'rasterisation' from continuous thoughts to discrete words, can do the job well (ie: band-limit and 'antialias' the message, fully considering the target sampling domain), or badly (ignore the target domain and just speak/write from the source context, leading to 'aliasing' in the listener's received message).

I know this isn't how many people think about words, but I think the idea of antialiasing is relevant and something that people should think more about (including with respect to AI-generated text).

> Writing is fundamentally the transfer of information from your brain to my brain.

Writing is fundamentally an artifact of intellectual work that has been done.

If real work has been done, there should be little work left for me to do.

If no real work has been done, and all the work of validation is now thrown to me, then I am being asked to carry someone else's burden of work.

Thanks but no, I have work to do.

I'm naturally verbose due to my ADHD. I'm trying to learn to be more succinct.

In the meantime, for business communication, I use AI to shorten my text, to make it more concise.

I changed the readme in my repo to say you have to send me a personal email if you want me to merge.

https://github.com/josephmisiti/awesome-machine-learning

It's helped a lot. Agents haven't figured out how to do that yet, or sendgrid, sns, etc are doing the hard work for me.

Maximizing the amount of documentation was never the goal, and people who assume that have entirely misunderstood what the point of documentation is.

I’ve started just asking it to touch up text to make it sound more human and follow technical writing guidelines and it converges on convincing pretty quickly. It also makes it easier for me to verify that it’s correct because all of the stupid LLM noise and formatting goes away and it gets distilled down to the important bits.

You might be surprised how many organisations/ teams now don’t even review pull requests. Claude does coding, Codex do code review, and if developer did 5 round of this before pull request, we could very well just let CI merge.

Summarizing large PR documentation sounds like a job for AI...

The one thing that gets with those PRs is that they want you to read stuff they haven’t even read themselves! Yeah, no.

I just have my LLM read the PR, then give me a summary of what the PR is about, how important it is, and how good the PR itself actually is. Then I have a conversation with the LLM about specific points, especially things where I get that feeling that I don't have a 100% understanding.

Until I fully understand what's going on, the PR doesn't move and my interrogation of the LLM doesn't end. My interaction is littered with "Explain X" and "How does this square with Y?" and "What if Z happens?"

The interrogation is the point, without me having to wade through hundreds of lines of irrelevant code to get at the meat of the matter.

Lately on my hobby projects I've been lazily committing work with the prompt 'add the relevant files to git and commit with a message that explains why' - only because I remember 15 years ago reading an HN comment from someone complaining that too many commit messages answer 'what' but not 'why' haha

So far I haven't had a reason to go back through commits to isolate any issues but if I do hoping the 'why' messages may come in handy for my LLM lol

Same problem with postmortems. You get a clean AI writeup with root cause, timeline, and action items all in the right sections, but none of it tells you what was actually checked or ruled out along the way. Reads like it explains the incident, but you can't tell if the fix was ever tested or just sounds plausible.

To be quite generous, that could be tongue in cheek.

That's a classic calling card of LLM prose but that doesn't mean they actually used an AI. Though sure, it would be nice if we gave this style a name and then abolished it.

FWIW Pangram [0] "We believe that this entire text is human-written."

I've only just discovered pangram, but I've seen it referred to a few times in HN recently wit nothing obviously pejorative about it. Take it with all required grains of salt, though

0: https://www.pangram.com/dashboard

I believe that to be a very obvious joke. The kind that the british are fond of.

The not this but that is bad when it is out of nowhere combination. In here, it has actual meaning.

I was about to comment on the exact same thing. Stopped reading after this. AI writing patterns has started to show itself in my AI allergy. Don’t get me wrong I like AI but it’s a pain to read the same style over and over again.

I'll be even more contrarian: I don't care if someone didn't write it as long as it's good (that's the major problem). I don't mind people using quotes when writing either. It's just that typically people (used to) curate their words before presenting them as writing. Although I'd prefer it if they also credited/explained what they used to write it in the case of heavy AI generation.

> One of the instructions I've hammered into Claude is "Write like a human. …" It works wonders.

If this actually works, it's absurd. The implication is that Anthropic could trivially make Claude sound less obnoxious, but chooses not to. I don't see a way this could be justified as a safety feature or anything, so.

Better metric: your LLM enhanced version cannot be larger than the prompt. Otherwise, you are just reducing your signal to noise ratio.

> So yeah, if you proofread and iterate on your AI's output until you feel you'd be proud if you had written it yourself, I'm happy to read it, too.

I agree, using LLMs is disproportionately frowned upon, but what matters is if you invested your own attention in the process. I use LLMs a lot for sparing, usually ask it to assume some opposing persona or use web search, not relying on its defaults.

How is ai giving them voice? All that generated slop does that you stop to want to read anything. And it is not their voice anyway. It is not their words, their ideas are mixed with ai generated points.

As for second paragraph, the writing "had no overall meaning". There is nothing to respond to if there is no overall meaning.

But how is that a positive that people that cannot give a fuck about writing coherent thoughts down to share them can now hog your attention with something that has no clever meaning whatsoever? Is talking just for the sake of talking a good thing in your book?

This is the equivalent of someone sending me an unsigned birthday card with just the Hallmark writing. They thought of me, but it's not the same.

I still want the imperfect human.

This is grotesque.

Everyone is an elitist, when you've enplaned do you assume the pilot is qualified to fly it or do you hope everyone gets a turn? Admittance to the canon is no less a responsibility, would you trust the yoke of human culture to the artless, craftless, and naive?

>not everyone does their thinking in articulate prose

not everyone does their thinking, if you dont I dont want you to have voice

Whichever it is, I stopped reading regardless

I'm curious if there have been any formal studies on this. I suspect that people have become better over time at identifying LLM writing styles, and are also combining that with selection/confirmation bias. Dunno what fraction of the perceived drop in quality comes from that.

Have they got worse or have we simply gone past the novelty state and are more critical of their outputs?

Have you seen the episode "Darmok" of Star Trek The New Generation series? It is about a people with an impenetrable language aimed at packing as much information as possible by using references and metaphors, not descriptions. I guess this is where LLM language output is heading.

> At least that feels human.

I feel the same. I've actually started to appreciate things I used to dislike. Like typos, or grammatical errors. I used to see it as a lack of attention to detail. But more and more it now feels like "hey, something written by a fellow human!".

Same thing for video voice overs. Things like a bad quality microphone, or someone who doesn't pronounce things very clearly. Now I go: for sure a human!

For better or for worse I’ve always preferred this, and IME a large portion of people (especially younger or who don’t write/communicate much for recreation or their work) will not read it or assume it was AI.

IMO it is a costly/goodhart-resistant way to “show your work” and help other people understand or challenge your mental model. (IE a justification for something you believe to be true). Overly polished writing is performative, it’s hard to take seriously once you’ve read The Economist/LW enough to see how poorly “well written” correlates with truth

To a certain extent are all wrong or ignorant about almost everything because our knowledge/time are very limited. But it is really important to understand what other people think in order to coordinate with them/align human goals and understanding.

It’s good that the average person taking the lowest-friction path to using an LLM in bad faith is easy to identify now. The more obvious and disliked it becomes, the more they’ll be hit with the stick to actually know things and not bother people. It’s so much worse to be “bad and stupid and not care” than “possibly cringe or wrong”

Nobody is going to read anything regardless of whether you used a LLM to write it or not (but if you use a LLM they'll use that as their excuse for the virtue signalling aura farm), so honestly why bother writing.

There's nothing AI about that sentence.

I’ve been away. Does anyone have a collection of these posts?

There is one person who is so good at writing with AI assistance.

Everyone else, including me, not so much. I envy his talent and ability to so successfully use the new tools. It's just something to think and talk about regularly. Maybe one day we will all be able to use the tools as well as that guy.

im sensing about a 20 month lag between hn sentiment and normy-coworker sentiment. managers and business-types pointing out the benefits and challenges of working with an llm i told them about early 2025. im just politely nodding along waiting for them to catch up. theres some emperors new clothes happening at my company so maybe im complicit

"Hey, the comments are too wordy. Can you clean it up so its only important information?"

As much as you want to say "actually do your job, jackass." Being goal oriented has had the best results for me.

It's something that my company has been struggling with. Developers are generating mountains of code, documentation, and Jira ticket comments. It's incomprehensible and overwhelming.

For the team I lead, my guideline is AI generated is fine but it needs to be human-edited and/or summarized. You want me to read what you're offering? Put some effort into it and meet me halfway. I don't want AI generated gibberish with made-up terms. You'd better also understand what you are presenting as your work. It's been fairly well-received though we're still working on it.

I have some co-workers on other teams who use AI to generate responses to literally everything. Ask a simple question? Get pages of AI generated nonsense in response. They are proving a tougher nut to crack.

In my current position(DevOps) we are getting this kind of messages or issues a lot recently. While forwarding these issues to my team lead, i don't have to deal with them which feels great but once its assigned to me... Its another long process to understand what issue actually means and what i need to do.

Thats said, i don't know any professional way for this but currenlty trying to communicate with peers in order to work without a problem.

I’ve seen the problem crop up in several places.

At a broad level, the issue isn’t generated content, it is the ratio of verification capacity to generation capacity (V/G). Your pain is because generation capacity has increased significantly, while verification is laborious and capacity has not (and can not) catch up.

Unlike spam which is from external sources and can be ignored, messages from other employees have to be responded to. I guarantee this is creating bottlenecks all across the firm, outside of the individuals who are feeling productive.

For fixes, theres theoretical approaches that might work?

If you need leadership to help you, then this issue has to become something that is on their radar, which means that something needs to go wrong or costs need to be registered.

The shortest conversation for that is to make people aware that generation has improved individual productivity, while moving the costs of that production to the rest of the firm.

If leadership is not at the stage to listen, then you need to move the costs you are incurring to the people who are sending them to you. Maybe set time aside to sit down with whoever sent a PR and then read what they sent together, to understand it.

It also makes a difference if tokens are being subsidized or not. If the firm doesn’t care how many tokens are being used, then you are naturally going to have over production.

Ask questions about how the audience is taking the blog posts. Say that many people seem to really not like LLM writing and ask how are they listening to their readers on this point.

You may find that brings a better and more human focused discussion. Who is the audience and what is best for them? Don't frame it about your preferences but about the real persons that your colleagues seek to talk to.

>the only thing that bothers me about AI-generated writing is if it's too verbose and obscures the important information by leaning ineffectively on jargon and high-minded verbiage

But so many READMEs and design docs are degraded to this level by uncritical/unmonitored use of Claude. You make it sound like it’s an exception, not a rule.

In my experience, that brain damage inducing Claude-style technical writing is everywhere. It sucks not only because of its verbosity, but because it includes completely pointless tidbits extracted from random extended LLM sessions, with no attempt to prune redundant or gratuitous details.

Apparently most readers have read too much AI output that they don’t recognize a bit of irony as a joke.

Yes, you can easily prompt this writing style.

As someone who is usually on the other side of this interaction (software engineer at a company receiving bug reports), it's much easier when people write in their own words (even if they sometimes use the wrong terms or say something which can't be true) than when they use an LLM to make things clear and correct. I want to know what you're trying to do, what you expected to happen, and what actually happened - that doesn't benefit from having an LLM rewrite it.

Sure, I use precise and technical language, but I'm still a human and I can understand what people are getting at when they use imprecise terms. But if the text passes through an LLM I can't tell whether the result actually matches the original problem or whether the LLM has made guesses or filled in blanks.

On Windows, use wincompose. Space followed by a hyphen gives you an emdash with the cute half-width spaces around it. On Android, I can get an emdash by long pressing hyphen. There's also a keyboard shortcut on the mac for it. No idea about Linux/BSD or iOS, unfortunately.

on a Mac, it's Option+Shift+-

— (em dash)

Option+- gives you an "en dash": – which is longer than a normal hyphen.

On Linux, it's whatever you say it is, after reading through `man xcompose`.

It's curious that you're pinning the blame of "excusing apathy" on the reader and not the writer. Why doesn't it go both ways? Given that producing LLM text is so cheap, your proposal amounts to an asymmetric war which is ultimately a DDOS on my finite time and attention. Or in other words: if it wasn't worth your time to write it, it's not worth my time to read it.

A lot of my problem with AI writing is that the facts aren’t checked. You get a big cheesy dramatic arc but the core story isn’t factually correct. There’s only so much time in a lifetime, too little to waste fact checking someone else’s work because they couldn’t be bothered to.

The lazy ones are the people spending 60s prompting then hitting copy/paste. They dump it on me as is and call it a job well done, which leads to me doing their job for them or calling them out to stop the behavior.

I don’t like doing either of these things. I don't want to do your job and I don't want to lecture about how it isn’t “your work” if you don’t touch the content after an LLM spits it out. It is rude and selfish to put me in that position.

The title is the conditional that underpins the claim.

Do you read our AI output end-to-end and decide that it’s the accurate content communicated concisely? If yes, this is not an issue.

If you are sending people generated text you haven’t read, how do you know what you sent doesn’t fall in the category of writing the author describes?

The problem is, some of us work with colleagues, or even bosses, who throw AI replies at us in nearly every conversation — so we don’t have a choice to NOT read it.

You don't have the choice to read or not read except when you read for leisure

It was very clearly on purpose, if you look at the surrounding prose.

It's clearly a joke. The rest is readable and well written, and free of LLMisms

I'd like to know how to type it with voice dictation.

On Macintosh keyboards it's ⌥⇧-. On the iOS soft keyboard, it's a long-press on the hyphen key. People often point out that various software will turn a double hyphen -- into one by default. We're overflowing with different ways to type an em dash, for those with any curiosity to learn.