My coworkers continue to dump hundreds of lines of AI documentation in every PR and every other line of code has between one and ten lines of AI generated comments, talking about the real unlock and how things are byte for byte identical on the load bearing path or how the acceptance ladder is misleading.
Features are coming out and metrics are improving, but we’re basically in a post readability code base, with the occasional performative comment about a variable name.
I don’t really know how to address this situation or if it needs addressed. I certainly don’t read the long-winded AI comments or the AI documentation, but perhaps it’s useful for the AI on its next pass.
The part that astonishes me is that in the year of our common era two thousand twenty-six that it's not universally offensive and reviling to post an AI-generated response to another person.
If I'm reading something on the internet, I'm either reading it to learn, or I'm reading it to be persuaded. If I wanted the LLM to teach me (thank you, no), I would ask an LLM. I'm reading your website/newsletter/email because I want to hear from you.. If you can't be bothered to put your time into writing it and teaching me what you think, why should I be bothered to read it?
It's like when you type a search in Google and get a wall of SEO spam.
I do feel like there is some middle ground where AI helps, but there's this one guy who keeps emailing everyone with walls of dumbass text and does not integrate or understand the feedback he gets when we are telling him what he needs to do. He just runs it though the LLM and replies and then forgets half of it in the next email engagement which functions as /clear on his end apparently. Like... dude I have ChatGPT and Claude, too. Actually read your goddamn emails.
I went to a farm opening this weekend. A guy was there with fliers with about the barn renovation he completed, products they used, how the process went. 100% LLM output. How would you talk to your LLM about something like that?
It's an ego issue for most people. For one, part of it comes from genuine intimidation or fear, the notion that a machine can be smarter than you, or at least produce answers that are more knowledgeable than you could. But I think the more striking issue is what LLMs do to the perceived value of expertise. Especially in engineering there's this weird pecking order hierarchy, where someone should be listened to just "because they have the YOE/Experience/Other meaningless metrics" under their belt. Now all of a sudden, someone with relatively little (or no) background in a given subject is capable of just entering a prompt and within seconds obtaining an answer that is expert or near expert level. (You can see the results of this with engineers just hopping between different fields with ease)
The implication is that expertise is being devalued in real time, and in a sense it really is, but I think that most people really do need to face the reality of their emotions and where it comes from. For myself personally, I have no qualms about reading or even engaging with AI content, as long as the content itself is of quality and truthful.
I really like the point I read somewhere the other day; instead of sending me the AI output, just send me the prompt you used to generate it. That is the only part that contains only the information you are trying to convey. The rest is just guesses flowery language added, and it confuses the actual message being sent.
Many people use many prompts and revisions or prompts or combinations or human and AI writing. Its not a simple as just making a single prompt and copying the output as the blog post.
If some text is AI written as a response to a much shorter prompt, then the prompt and/or sources used to make the text should be published instead (or at the very least together with the text).
I had to do this with my boss. He sent an email saying he asked AI about an assignment he gave us, then sent us the reply from the LLM. It was verbose and lacked any and all awareness of the constraints we have within the company. The assignment was poorly defined from the outset, and I asked for the prompt he used, because I thought that would be far more useful to understand what he was looking for rather than what AI decided to spit back.
It turned out his prompt was equally uninspired... 1 or 2 sentences. That was the thought he put into it, which then had us spending hours trying to figure out the AI reply. He could have saved the whole team a full day by just sending the prompt... or nothing at all.
Funnily enough, for me AI writing is usually the summarizer rather than expander. I would usually give Claude 2-3x the amount of data in notes, context, braindump etc; when doing person-to-person communication, using AI to _expand_ the context rather than shrink it is doing everyone a disservice.
I wish I could do a "right-click -> view page source"-like "right click -> view prompt" or like the dark mode toggle (prompt <-> ai slop (expanded)), depending on my mood (although almost everything is in constant dark mode, so I believe I'd look mostly at the prompts)
The infinite monkey theorem[1] can be applied to show that AI is not incapable of producing quality output. However, expecting your audience to sort through the output from a very large quantity of monkeys is not respectful. I think the best approach would be:
"Here is the prompt and model I used, here is the portion of the output that stuck out to me as relevant and useful, and here, in my own words, is why I think that is so."
i.e. Highlight the good part of the output for me and, preferably, explain why you think it's good. Otherwise it's the same thing as posting slop, only with extra steps. You thought it was worth posting and I thought that made it worth looking at, but now I'm sorting through a bunch of monkey garbage looking for what made you think that. Being a different person, I may not find that hidden gem and I may become quite frustrated with you.
I think of a prompt more like a tweet. The medium does matter. If someone was only willing or able to put an abbreviated amount of thought and effort into something, it's not worth a disproportionate amount my time or attention.
Exactly. And I think the information theory matters here: it's worth putting in information proportional to the human-supplied input, not proportional to the AI-expanded output.
"please walk our observability stacks, our IaC code and config, our cloud account, and our application codebases to gather an RCA for for any errored out request responses we've been producing this week. include a tabulated breakdown."
~actual prompt i sent in to claude last week. doing what you propose would be basically impossible / nonsensical, and further undo the entire exercise, sending all the money spent on the tokens down the toilet.
put differently, if you use agents in any actually useful manner, the principle you describe erases whatever value they did end up providing. though since the whole underlying premise is that ai never provides any value, maybe that's the intended result and is as expected. in which case, i'd consider such an expression of desire more than a bit demagogue.
> Look, I get it. It’s Q3 2026, and we should expect that everyone is utilizing AI at SOME point in their process (sourcing ideas, creating outlines, refining prose, etc.).
Sounds like this isn't everyone, but rather, it's people who have nothing to say, and -- even after they've "sourced ideas" -- they don't know how to reason about it, nor how to communicate it.
This isn't everyone. This is people engaged in generating noise, for the sake of noise.
I left a comment on some of my interns, clearly AI generated code and response asking for him to please not respond with AI generated output — the response was talking about things I had known he was unfamiliar with because I had worked with him. The response itself had some particular formatting/content decisions that were generally a very odd way for anyone to write, but certainly not an intern. He responded that he did write this himself and that he was upset I would accuse him, etc.
I, of course, had no choice but to apologize. I hope I was right, but clearly there is some chance I was wrong.
We have a junior research student who does this on slack.
Before I think he was using AI and other tools to translate his messages since english isn't his first language. His own words, then AI translated it. Sounded kinda clunky, but I could sense the human behind the words and I gave grace since I can only imagine how hard it is to properly communicate your ideas when English isn't your first language.
But now it's gotten to a point where I can tell he's not using it for just translation. I'm gonna need to chat with him.
Redacted example below
```Hey @PERSON_WHO_ASKED_QUESTION Both good, and the retrieval one isn't written down anywhere. In order.
One corpus or four. My lean is one. Same chunks and embeddings tables, source_type column to tell them apart.
Values I'd propose, flat rather than nested: paper, dataset_description, dataset_readme, dataset_contributors, dataset_records, dataset_files. Description and readme split because their units already differ, one row versus one row per paragraph. A discriminator that can't separate those isn't doing much. The existing 440 rows would need backfilling to paper.
Chunk id in the same spirit: dataset doi, source type, ord. So EXAMPLE_DOI.
Worth checking before any of this matters: does chunks.doi carry a foreign key to papers.doi in 0001_init.sql? If it does, a metadata chunk with a dataset doi can't go in that table at all, and separate storage stops being a choice. Ten second read, I haven't done it. Shout if you get there first.
Retrieval is the one that's bigger than it looks. Search once across everything and nothing guarantees a paper chunk and a metadata chunk both land in the top k. The facts we want relate the two, and the generator can only write those if it sees both sides in the same window. So if one type systematically wins the ranking, that class of fact doesn't get worse. It becomes impossible.
Which way it goes I don't know. Two mechanisms pull opposite ways. Metadata chunks are short, tens of tokens against roughly 450 for a paper chunk, so they may just lose. But we embed context header plus text, and on a short chunk the header is most of the vector. The header is the dataset name, which is also most of the query. That points the other way.
Cheaper to measure than argue. Load one dataset's metadata, run a normal dataset level query, record the rank of the first chunk of each source type. Runnable as soon as any one of our four subtasks lands.
After that it's one pool, per source with quotas, or one pool with a floor per source type. I'd rather not pick before there's a measurement.
```
If a coworker sends me an AI generated report to help on my task, I am more likely to ask them to just send me the prompt and model they used. Let me own the session and steer it however I want.
I am pro AI, and also deeply share this sentiment.
I think sometimes people have a big long thread with Claude where they feel enthusiastic about the back and forth being very productive -- and then they genuinely want to share the summary of the conversation with their colleagues so that they can feel it too. It's not just typing a short prompt and copy/pasting the response.
But, Claude, especially Opus is so unnecessarily flowery, and smug -- and it's so identifiably characteristic.
It somehow rubs salt in the wound: "I didn't take the time to write this out in my own words, so now you get to listen to this petulant asshole mansplain it to you".
If it would get to the point, and with some humility, it would be a much different proposition.
> But if you’re my colleague and we’re in a Slack discussion and you post a wall of Claude output, then I’m afraid I received a different message than you intended.
i have what ill call a "condition" where i fear assuming someone did something stupid is insulting. i try not to patronize people because i also assume that would be insulting. this leads to issues where sometimes people really want to be lead to water rather than make an implied connection themselves
this is a roundabout way of saying, is there a genuinely polite way to tell colleagues posting ai copypasta makes them seem ignorant? i dont think ai:dr is workplace appropriate. id like to tell people gently and avoid patronizing if possible
I thought the exact same. There is also a growing problem with people using AI to summarize everything instead of reading the actual text. I thought this addressed that issue by having a section for the AI to help summarize or something similar.
"Look, I get it. It’s Q3 2026, and we should expect that everyone is utilizing AI at SOME point in their process (sourcing ideas, creating outlines, refining prose, etc.)."
Hell no. Sourcing ideas? From AI? Are you kidding? Essentially guaranteeing they won't be original? Refining prose? Making it much more likely it'll be generic as hell. Call me old fashioned, but while I can certainly understand asking LLMs concrete questions - how do I fix this, where can I unsubscribe from that - using AI to help communicate seems bass-ackwards. LLMs have no idea what you're talking about - they're free associating probabilistic sentences! What on earth value could that have in communicating. They self contradict, hallucinate, speak most frequently in a weird smug corporatese. How is this an attractive use case of the technology?
There is a light at the end of the tunnel for all the minds burned out by AI slop.
I already saw someone getting fired for only producing AI text as part of their entire output, after being unable to explain what they “wrote” in multiple situations.
And recently my company enacted a mandate that text made for humans should not be AI generated.
Sooner or later every company will start realizing that this is just people too lazy to actually work and coasting on a paycheck, at the expense of every other worker.
This makes me happy to hear that someone got fired for that, but I just cannot picture that happening at my organization. Perhaps standards are too low. What sort of thing did they get fired for? Was there some warning etc?
A similar metaphor (I think I came up with but can't remember) is the slop sandwich. You can deliver slop to me but only when wrapped in a thick layer of human intervention (which usually also means removing 50% of the slop)
This hurts particularly much when it comes to software changelogs/release notes, most recently for me was the solidjs 2 beta, reading the announcement post was painful, even the headers are LLM:isms.
> TL;DR (too long; didn’t read) was the solution for social media.
> AI;DR (AI; didn’t read) is the solution for AI slop.
Maybe I'm the odd one out here and didn't plumb the worse-depths of social media, but I feel I need to defend good old TLDR as a slightly different, and less-hostile animal.
It can be a moral condemnation of the poster, where they're disrespecting everyone else's time by posting something big and vague... But it usually isn't. At least half the time TLDR is someone providing a summary because they think it'd be useful. (Or as hostile critique.)
tl;dr as the sole reply to someone IS a condemnation of what they wrote. tl;dr at the front of your message with a short summary, is an admission that the long version may not be for everyone and there's a core point for quick consumption.
I'd say that sounds about the same for ai;dr. I could see myself posting a prompt like that in front of the generated report.
The purpose of writing is to communicate. The person communicating is an important part of the message.
Consider the sentence "I'm waiting for you at home." It really matters if it's said by your loving wife or an evil clown. You can extrapolate from this.
So how much time do you deem a minimum amount for someone to spend on some unit of work before you're willing to read it? And what technologies are allowed -- are word processors allowed or does it have to be done with a writing instrument like a pencil? What about spell check? Do you care if they had done their research with a Wikipedia or do you prefer card catalog for the really important work?
In one case, I could tell because the tone of the piece sounded exactly like the kind of stuff I would read after prompting an LLM and I had never heard of the author before, but I knew that they were like an intern or a new college grad and for some reason their piece was trying to persuade employees in a certain field about how they had to use AI itself. It just rubbed me the wrong way because if they did use AI, then they didn't even bother to rewrite the repetitive instances of "it's not X; it's y", and also "This isn't X".
Imagine a future where 99% of text is written by 1 unpassionate person. Reading the same patterns, the same tropes, the same flow across paragraphs, over and over, would be unpleasant. It already is unpleasant.
Eventually, AI will be a far better writer than a human is, and then nobody will want to read the human generated subpar writing. It's already approaching that.
>If you’re not bothered enough to review and edit it...
>...then I’m not going to bother reading it.
Just playing Devil's Advocate here, but we've had copywriters and editors for generations now. Not reading something because AI did the post-writing work feels kinda weak. Skip stuff started from AI, fine, but this is just because it's AI, not because the author didn't do the task. And I do get the point is that if an AI did /most/ of the work, ok to that, too, but not post-write processing alone.
The problem with copywriters is that they wrote "copy" - which turns out to be writing that I don't want to read. And the problem with AI is that much of what it trained on is copy, so it trained to write stuff that I don't want to read.
So in the end, I don't care if a human wrote it or an AI, I don't want to read it either way.
What's wrong with copyrighters? They write like corporations, not like humans (see the Cluetrain Manifesto for more on this). They sound like nobody - literally inhuman. I don't want to read that. I want to hear a human voice. If I can't hear that in your writing, I don't need to read it.
Features are coming out and metrics are improving, but we’re basically in a post readability code base, with the occasional performative comment about a variable name.
I don’t really know how to address this situation or if it needs addressed. I certainly don’t read the long-winded AI comments or the AI documentation, but perhaps it’s useful for the AI on its next pass.