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.

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.

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.

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.

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.

> 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 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.

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

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

If you read anything you're reading superfluous bits.

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.

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

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.

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!)

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

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.

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.

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.

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.

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.

"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.

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.

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

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’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.

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.

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

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

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

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

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

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.

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

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.

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.

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?

Whichever it is, I stopped reading regardless

>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.

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.

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.

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.

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.

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

There's nothing AI about that sentence.

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.

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

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.

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 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?

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.

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.

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

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.

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

Same here. I also used to love writing with em dashes, but don't anymore out of fear people will call my writing AI and stop reading. It'd not hard to do in most word processors, typing two hyphens "--" does it in Word and Pages, so I don't get why people think it's such an obvious AI tell. It's not that exotic.

Maybe we will get to a place where we normalize AI authorship or co-authorship in our writing?

I appreciate what the Palm Springs Post does with its news articles. It has an "AI Assist" author and gives it credit in bylines as appropriate:

https://thepalmspringspost.com/author/ai-assist/

That's nice. So when I'm reading I know whether it's a human author or not. And if it's a dry municipal meeting, seems like a perfect place to send in the bot!

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.

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