[cautiouscat]

This quote from the authors friend really hit home with me.

> “If you’re going to use an LLM to write me an email, I’d much rather you just send me the prompt; at least then I’d have an idea of what you actually meant to say.”

I’m not saying there’s no merit in adding a bit of politeness and professionalism to your communication, which I’m sure the prompt itself lacks. However the root of what you’re trying to convey is the prompt, wrap that in a header and a signature. Not only are we talking as humans, we’re also communicating directly.

Also I just find it a little insulting if someone sends me an AI response. I don’t know why, maybe because it feels not genuine.

[roenxi]

Yeah, that's one of those sentiments where people say it but they probably don't mean it literally. Much like if your boss asks for honest feedback giving it to them with both barrels is a career limiting move.

You make subtle mistakes in how you perceive the world, the interlocutor makes similar mistakes and the damage those mistakes do is limited if you follow some structural rules of how to communicate (aka politeness).

AIs only rewrite what is in the prompt with more words so it can be insipid but I'd expect to do better on average with that then sending out raw prompts. I'd suggest the real ask from the friend is "put more time into communicating with me than a short prompt".

[tsunagatta]

I am not the author but I have used that quote before, and no, I mean it literally. I would prefer any words that came from your own brain over the output of a statistical model. I don't mind if it's short, or didn't take much time, or doesn't have much thought put into it, I want it to be yours. The subtle mistakes in how my interlocutor perceives the world and communicates themselves are the entire point.

[kigiri]

I had a case in my job of an HR guy that was so bad a communication everything was super confusing and we always had issues and tension. Once GPT came out he clearly just copy / pasted LLM output and I was very conflicted because it felt insulting but it was the first time I was actually able to understand what he intended to say...

I sadly would prefer LLM output over prompt because I'm pretty sure the prompting was a process with him he would prompt, get a bad LLM output, seeing the LLM wasn't understanding force him to clarify and we ended up with actually understandable content.

[nojs]

“tell my obnoxious boss to fuck off about the tps reports” isn’t a great career move for them though

[fcarraldo]

I would rather my reports tell me to fuck off than to generate something telling me to fuck off in polite but insincere terms full of emojis and em dashes. Honesty is valuable.

[throw283745]

I'm the opposite. If you're going to say something mean or obnoxious, then I would rather you say it politely with a lot of emojis.

[delusional]

I too would much rather that, because then I get to know them as unhelpful whining complainers and fire them. That's not a great outcome for that person though.

[LtWorf]

Lol you can't fire people for using AI when the corporate mandate lately is to use AI.

[LtWorf]

I think it has extra value in that it will be more unnerving to them than using profanities.

[afavour]

Generating AI responses to all your emails isn’t a great career move either, though

[LtWorf]

My former CTO was doing that all the time.

[wiseowise]

It isn’t a great career move regardless, because if you use on prom AI they most likely monitor your prompts.

[LtWorf]

You don't use the work AI for that :D

[laughinghan]

> You make subtle mistakes in how you perceive the world, the interlocutor makes similar mistakes

And the LLM makes subtle mistakes perceiving what you were trying to say, and I make subtle mistakes perceiving what the LLM generated.

It's interesting that you seem to think an LLM would be better than you at understanding what someone was trying to say. I have complete, 100% certainty that if a personal friend of mine was having trouble expressing themselves and was concerned about being misunderstood, I would understand what they were trying to say better than any existing LLM. (I suppose the exception would be if the friend was referencing some factual matter I'm unaware of but that the LLM has memorized, like a pop culture reference I didn't get or something.) Do you find that ChatGPT has more emotional intelligence than you?

[emodendroket]

That may be. But we often find ourselves communicating with people who are not close friends. And the concern may well be the opposite -- a desire to find and scrub ambiguity from one's own message.

[roenxi]

Emotional intelligence includes things like being able to perceive and manage your own emotions; it is hard to know what that would mean for an LLM. They don't have hormones or much of an ego. They also aren't usually in a position to perceive body language or tone.

But in terms of reading comprehension I haven't seen anyone who does a better job than an AI. It's a skill that technically caps out quite quickly, all you can really do is restate what someone already said and make inferences off that.

> I have complete, 100% certainty that if a personal friend of mine was having trouble expressing themselves and was concerned about being misunderstood, I would understand what they were trying to say better than any existing LLM.

Your personal circumstances are a mystery and close friends are typically people we're associating with because they think in fundamentally compatible ways. So sure, why not. But assuming a more-or-less normal person I expect an LLM would outperform them. LLMs are relentless about being reasonable and focusing on what was actually said in a way that most people usually need a bit of training to accomplish.

[filoeleven]

> But in terms of reading comprehension I haven't seen anyone who does a better job than an AI

You've clearly never seen a CoPilot summary of a Teams meeting.

[flexagoon]

I would absolutely prefer just getting the raw prompt. Emails are already ridden with enough completely unnecessary parts that are somehow still the cultural norm (eg. signatures). At least let the important part be to the point. I yearn for a world where people write emails the same way they write text messages.

[allarm]

> completely unnecessary parts that are somehow still the cultural norm (eg. signatures)

I don't even know where to start. Long story short: you're wrong.

[coldtea]

Nope, I would prefer getting the prompt.

I would also prefer anyone sending me the AI output and not either their ideas or their prompt, to not contant me ever again.

If I wanted the output of the AI, I'd ask it myself. You're just a useless intermediary if you send it to me.

[emodendroket]

Maybe you personally do; most people would react more poorly to "find a polite way to tell this guy to fuck off" than the output of that prompt.

[coldtea]

That's the most contrived counter-argument...

We've been writing business emails for 40+ years. Anyone with -2 stddev IQ and above would know to tweak a 10-word prompt to a more polite version before sending it

[emodendroket]

It's an extreme expression but "soften this message" is a common goal.

[conradludgate]

This might apply to the AI integrations in Gmail, but someone's own agent setup might have more references saved privately than yours. You can't ask your AI to search through my personal notes.

[coldtea]

The people that send that crap are the guys who one-off it, not the guys who have a carefully crafted context, skills, and local references on top of the prompt.

[8n4vidtmkvmk]

I was going to say... Sometimes I load up my context with a ton of data. The output is shorter than the input.

[floxy]

I've been going back and forth on this, and I think I agree with you about 95%. But for some cases, where I don't really have a choice about whether or not I'm going to hear from you, there are certain dullards that I'd rather have filtered through a chat bot. It really would be an improvement over whatever they had to say. Listening to their genuine human output has negative value, but filtered through an LLM might bring that value up to zero.

[jcgrillo]

> if your boss asks for honest feedback giving it to them with both barrels is a career limiting move

Or, from another point of view, expanding your career's event horizon. Imagine how much longer you might have languished in that stifling, stagnant job if you hadn't unloaded on the boss?

[zamadatix]

It really depends on the individual and the context they say it. Same for the bit about your boss.

[bigstrat2003]

No, I want the prompt. Slop from the clanker has zero value to me. At least the prompt was written by a human, even if it's otherwise not useful.

[delusional]

> I'd suggest the real ask from the friend is "put more time into communicating with me than a short prompt"

That's where the real truth is. If you're communicating in short prompt messages, did you really even have a thought worth communicating?

I sometimes abort writing an email when I can't think of a polite and constructive way to express my opinion. I have never once regretted aborting such email. I'll often come back to it later, with much better thoughts and an actual constructive point to write down, and then I never struggle with the "style" of the message.

[Jcampuzano2]

The problem is everything we send that was created from a prompt is not you.

If I wanted to read something that wasn't written by another person, you might as well swap out the from field for "Claude" or "gpt-5" or w/e and stop pretending you had any valuable input.

Sure there's something to be said about having an AI help. But I'd rather that be an attachment clearly labelled and for the content to be strictly reserved for a human.

This is already how I think PRs and such should be written. There should be a field or section of the description reserved for the AI generated content, with the rest being for the human to clearly describe their intention.

But instead we're living in a world of AIs masquerading as humans. And it's only getting worse.

[mrandish]

I've always assumed when I hear people talking about AI emails, they're talking about longer formal writing to groups, such as a proposal not personal 1:1 messages. Hearing this kind of surprises me but I don't know first-hand since I was able to retire just before LLMs became ubiquitous.

In my experience, 1:1 or 1:2 emails among coworkers, peers or friends tended to be as short and direct as possible anyway. Even pre-LLM, an email as long or structured as an LLM would create today would have been notably weird.

[nlawalker]

>Also I just find it a little insulting if someone sends me an AI response. I don’t know why, maybe because it feels not genuine.

It's because it's a pretty safe bet that the prompt either implicitly or explicitly says "I don't really care," and the act of using AI to generate the response is almost entirely an effort to conceal that.

[emodendroket]

I don't think that's a safe bet at all; the choice to put the message to an AI may express a desire, however misguided, to produce a "better" (clearer, less likely to cause offense, whatever) message than the author feels able to manage unaided.

[luckylion]

Of all the times someone used AI and did not disclose it, I've never found it to be a desire to be better, it was always a desire to do 5 minutes of work and still have a result that looks like you spent multiple hours. Except it only looks that way at a glance, it's usually bad because that type of person does not care, so their prompts are bad, and they don't care to even read the results.

In my experience, the people who actually wanted to improve the presentation of their messages are upfront about it and clearly say they're using AI to organize their thoughts or polish their english.

[nlawalker]

Yeah my comment was maybe a bit too snarky. I am personally a huge believer in assuming good faith and intent and giving people the benefit of the doubt, it’s just that when it comes to this, I’ve been in many situations where I have enough context that, even though I don’t have proof, I don’t have much doubt either.

[delichon]

To a friend or colleague, yes. But for quotidian business transactions LLMs are such a shortcut. I pointed one to three Amazon orders that I wanted to return with reasons for each, and asked it to look up the details and generate warranty service request emails for each. A quick review copy paste send, boom done. A 15 minute task done in 5. Repeat and it adds up. But you need to know when not to take the shortcut.

On a good day it's a 240 volt hair clipper for grooming yaks.

[pavel_lishin]

I think the difference there is that a warranty email isn't aimed at a person, but a company - which, I think cstross has pointed out, was humanity's first AI.

[majormajor]

Last time I returned anything to amazon it was a two-minute-or-less task that didn't involve anything like a warranty service request email?

[delichon]

I buy sous vide cookers that die in about 3 months (all brands I've found do when run 24x7). That's outside of the normal return window and requires a warranty claim to get a replacement. For that you go through the maker, not Amazon. This basically doubles the useful life of the purchase, but I'd rather find one that lasts.

[emodendroket]

I mean, yeah, I consider myself a decent writer, but I don't think that skill was best or more meaningfully expressed in writing Jira tickets so I'm pretty happy to outsource that.

[nickvec]

Regarding your last point, I personally find it insulting when I receive an AI response because of the subtle deception involved. When I am talking with another human via Slack/email/etc., I am under the impression that the messages I am receiving are written by the coworker/person in question. When instead I receive an LLM's output, that trust is broken, especially if it is not made transparent that AI was used to generate the message.

[reaperducer]

Also I just find it a little insulting if someone sends me an AI response.

If you can't be bothered to write it, I can't be bothered to read it.

[singingtoday]

I'm dyslexic. I put every comment through the LLM, or other tools. Including this comment.

I understand where you're coming from but please believe me when I tell you that if I write comments myself nobody will understand them and it just turns into an argument where people claim I say things that I didn't say.

By filtering my comments through an llm, I have reduced this issue significantly.

[coldtea]

>I'm dyslexic. I put every comment through the LLM, or other tools. Including this comment.

What we said here applies to the general population, not such special cases.

Of course, if a dyslexic lets it do more than correct typos and grammar/syntax, the "don't send me LLM crap" also applies to them!

And even for non-native speakers, I'd prefer to get their actual output, not the LLM version.

[frizlab]

There’s a difference between fixing a few incorrect words in a text and having an essay (or email or whatever) written from a few words. I don’t think the parent comment would object w/ your use of LLMs.

[singingtoday]

Thank you for seeing it my way!

[filoeleven]

I was going to ask: what does an LLM get you that a regular spell checker does not? Then I recalled all of the comments I've read over the years about "loosing to much money do to a rouge actor." The capability of an LLM proofing correct words and not just correct spelling is genuinely helpful here. Seems like a good fit for a small local model.

[what]

> nobody will understand them and it just turns into an argument where people claim I say things that I didn't say

I don’t think that’s dyslexia.

[galangalalgol]

That is a fantastic use and would likely benefit asd as well. Could you share your strategies for creating something that is still you and concise but comprehensible to neurotypicals?

[singingtoday]

https://i.imgur.com/FyTfcyx.png

[galleywest200]

How can I trust that the sender actually understands the words they want me to read if it is all just AI output?

[Yokohiii]

You don't. The sender is just pushing responsibility back to you.

Probably the best response would be "No." The least anticipated one, creating a chance of actual communication.

[hxtk]

When I read what someone has written, I learn things beyond its literal text from the fact that I’m reading it, which implies it was worth effort for them to put into so many words and send to me through the medium they chose and put the level of care that they chose into their wording.

The LLM erases those choices and erases the cost of verbosity so there’s much less for me to learn from a message, and much less I stand to learn about a person though repeated exchanges to help me better contextualize future messages I receive from them.

[throw283745]

I use LLM to summarize a lot of emails, so it doesn't really matter how it's written.

I wonder if there could be just some type of open agreement we're both using LLM and then they can communicate with one another more effectively.

[throwaway_19sz]

I disagree with the sentiment. The prompt is not what you meant to say. It is part of the private act of thinking, of choosing your words. The thing you send is what you meant to say. There has always been a vital distinction between what you think and what you say. The notion that you should ever have direct access to private thoughts that someone chose not to send to you is wrongheaded.

It’s like receiving a letter and then demanding to see someone’s private notebook where they jot down their thoughts, so you can see what they ‘really meant to say’. Most communication involves a degree of negotiation and persuasion. Thoughts are necessarily private.

There is a problem with AI generated emails: they change the balance of time/effort required to write them vs. to read them. I think this is a valid concern.

[Planktonne]

> The prompt is not what you meant to say. It is part of the private act of thinking

The private act of thinking does not involve a round trip to a corporation's servers; that would make the whole term meaningless.

[dataflow]

> The private act of thinking does not involve a round trip to a corporation's servers; that would make the whole term meaningless.

Firstly, this is just wrong. People can and do search stuff online as they're writing something up.

Second, your distinction is the one that's meaningless. The LLM could be running locally on a private machine...

[Planktonne]

1. Yes, but the research isn't part of the 'private act of thinking'; it's an input to it.

2. It could; it basically never is though, so let's deal with the general case.

[throwaway_19sz]

You’re missing the point. OK, it’s not ‘thinking’ in a pure sense; the point is it’s private notes. If someone says something to you, they are allowed to privately prepare it first (thinking, jotted notes, LLM rewriting, running it past lawyers, whatever) and you don’t get to see all that. That’s how it has always been.

Just because someone says/writes something to you doesn’t mean you have any moral right to see all the things they considered and then decided not to say.

[brailsafe]

> Also I just find it a little insulting if someone sends me an AI response. I don’t know why, maybe because it feels not genuine.

I'm yet to experience this, but it reminds me of an interaction I'd regularly have with a younger friend I no longer spend much time with, who would take my in-real-life casual conversation prompts that attempted to engage him in inconsequential curious dialogue, and simply google the answers in front of me.

He'd sort of deliver that response in such a way that it seemed like he assumed I either didn't know the definition of some common term or fact and he was unlocking new information on my behalf, or that I thought he didn't know and revealed a possible weakness in his knowledge of trivia. In reality I was embarrassed by proxy because I had unintentionally provoked a smug, emotional response from someone who wasn't able to contend well with ambiguity, revealed how dependent they were on their phone, and who's first instinct was deeply uncharitable and lacking in curiosity. It led to some awkward situations where I had to explain that I wasn't trying to test them in a battle of intellect, but rather ask what they personally thought or how they used a term who's actual dictionary definition is rarely relevant to it's usage. I tried to find a way to adapt to this incompatibility by thinking about different ways I could communicate with this friend that would indicate my intention to basically just get them yapping more clearly, but never found a way that clicked.

I fear that dependence on various LLM chat tools will lead to more of this sort of problematic thinking and introduce increasingly challenging communication conflicts across otherwise not so distinct age groups. Banter that's intended to be playful will continue to devolve into two sides having no clue what the other means or if they're joking or exaggerating or what, linguistic flourishes, colloquialisms, or regional differences, will become more difficult to execute upon despite there being no intrinsic reason for an incompatibility to exist.

Less akin to knowledge of memes or slang, and moreso to the prevalence of new-irony among people under the age of 23ish who don't know any other type of irony and vice-versa with older -> younger. Things could get weird

[esseph]

> Also I just find it a little insulting if someone sends me an AI response. I don’t know why, maybe because it feels not genuine.

Ironically the people I have talked to that do this are actually trying to give you their very best at communicating, at least better than they think they can do. Which may or may not be true. I know being a bit on the spectrum that my comms don't always convey the "tone" that I mean, despite what I'm thinking when writing it.

[RobotToaster]

> I’d much rather you just send me the prompt

I'm sure "please analyse this email and explain why this complete moron is wrong, write as if addressing someone with the comprehension of a 5 year old", would be happily received by my boss.

[patates]

Usually my back and forth with the llm is much longer than the result. Add the skills, the context, access to my calendar... I'd probably need to send a zip file instead of 1 short paragraph of text. Maybe letting agents representing people, with access to non public information will be secure in the future. For now, that part is manual. I need to control information on the way out.

You probably should find it insulting when someone obviously didn't take the time needed to form a proper response. Prompting "write something apologetic and kind for not attending their wedding" and sending the response is not okay. Like this example, it's easy to recognize when being a human counts, we should also be able to justify automating stuff when it really is just boring information exchange.

[blks]

Do you really need the agents, skills and them having access to your calendar, plus multiple back and forth, to create an email about scheduling something - and anyway you will have to check it manually?

[patates]

Yes because it's not a single calendar, and I sometimes need to be reminded of potential plans (not saved on a calendar), history with that specific person ("oh I already made a demo to this person 2 years ago and from my notes, it seems that I didn't get a lot of feedback, perhaps prioritize someone else?"), and other context. I offer the LLM context, and I in return get a distilled context from which it's easier to make decisions.

Writing important emails used to be a 1-2 hour occasion, nowadays only 15 minutes, but surely not 10 seconds.

[fantasizr]

I feel bad for this generation who will have to listen to prompted eulogies and wedding toasts. And worse for the genuine writers who now will be accused of having done so.

[coldtea]

Or to have to live all their lives with slop thrown at them left and right...

[strken]

I'm actually a bit offended, or at least wary, when a human writes an email with the length and structure of an LLM's output.

Too many words per idea shows me that they're trying to hide something, manage or manipulate me, or play some other game.

[pash]

The obvious solution is to run things in reverse, inputting the AI-generated output to recover the prompt that generated it.

Most generative models can be run in reverse by algorithms that already exist [0], but you have to have the model weights. For closed-weight models, or for a process that can handle unknown models, you’d have to do some engineering.

But do we have the technology to build models that back out the prompt from suspected AI output? Yes.

0. I don’t mean that most neural networks are invertible functions. They’re not. But you can do backprop in reverse, from output to input, to train a model to generate an input to the original model that best predicts its output.

[noddybear]

Most of the functions that LLMs perform aren’t bijective, though.

What prompt constructs the output ‘The answer is 3’ or ‘Yes that’s a great idea’?

[pash]

Right, that’s why I wrote, “I don’t mean that most neural networks are invertible functions.”

For a neural network that is not bijective, you can obtain an input that maps to a desired output by the following algorithm.

1. Start with a trained neural network. (The weights will not change throughout this procedure.)

2. Pick a random input.

3. Given an output for which you want to compute an associated input, feed the input into the network to compute the output.

4. Compute the loss of the computed output relative to the target output (e.g., mean-square error). If the loss is sufficiently small, you’ve found an input that maps to an output close to your target output and you’re done.

5. Otherwise, compute the gradient of the loss with respect to the input (e.g., by backprop).

6. Update the input according to a gradient-update rule. And go back to Step 3.

In theory, you can recover a “representative” prompt for the output of an LLM in this manner. For outputs that could have been generated by a large set of disparate prompts, obviously this won’t work well.

[bentcorner]

> Also I just find it a little insulting if someone sends me an AI response

I run into this with AI-generated PR comments. I think where I work we are still grappling with the "right point", because LLMs can certainly provide valuable feedback, but they are not at the point where they can do so unsupervised, and to do so just feels unprofessional.

And there's another layer where it is even worse when a colleague spends the time critiquing code and someone (or something) replies to the comment with mostly useless filler. It's like being handed a small hand-crafted gift and then throwing it into the garbage in front of them.

[rogerrogerr]

I often use LLMs to write or review code, but never ever copy/paste from them into any text box attached to my name.

If my name is posting something, I understand it. If I can read an LLM response and reword it, I understand it.

I really want copying LLM prose into anywhere without a block quote to be a firable offense.

[LtWorf]

> Also I just find it a little insulting if someone sends me an AI response. I don’t know why, maybe because it feels not genuine.

It is certainly insulting, I see it more of a malicious compliance way of working.

[gcanyon]

I invite you to consider the Seinfeld finale, where Elaine uses a cell phone to call a friend to ask about her ill father. The call has technical issues, and Jerry lectures Elaine about mistreating her friend that way, when she should have waited to call from home.

That was less than 30 years ago. Cell service is more reliable today than then, but it's not perfect, and not only don't we think twice about making <whatever> calls on a cell phone, many/most of us don't even have land lines anymore.

Sending the prompt is the difference between showing up with flowers, and saying, "I feel strongly enough about you to bring flowers."

https://en.wikipedia.org/wiki/The_Finale_(Seinfeld)

[globular-toast]

I wrote this almost a year ago: https://blog.gpkb.org/posts/just-send-me-the-prompt/

[m463]

> if someone sends me an AI response.

I wouldn't say "not genuine", I would instead say "not mindful" or "not meaningful".

maybe not quite "mindless" and "meaningless", but maybe 10%.

[madrox]

While I tend to agree, if that email were about arranging a D&D session, I don't think I would care. Especially if the alternative is we never find a time to play D&D.

[amfarrell617]

What if there were multiple prompts and a simulated-conversation back and forth as I tried to figure out my own thoughts?

[anal_reactor]

Claude please write a comment explaining why "just saying things" is a retarded idea because most people are dumb fucks and the only way to cohabitate with them is to tell them what they want to hear, not what is true.