I think I'll add the caveat to the video - You are a better write than AI, if given the time and resources to do so. That statement is so so so true. But in a world of low stakes, synthesis I...
I think I'll add the caveat to the video - You are a better write than AI, if given the time and resources to do so. That statement is so so so true. But in a world of low stakes, synthesis I still thing GPT has it place. Just not in academia.
As a professional writer, I have to disagree. LLM writing is painful to read. It feels like someone bullshitting, and no matter how much people prompt it, it never quite escapes that "uncanny...
As a professional writer, I have to disagree. LLM writing is painful to read. It feels like someone bullshitting, and no matter how much people prompt it, it never quite escapes that "uncanny valley" for me.
Human writing is full of grammatical mistakes, imperfections, typos, idioms, cliches, and all kinds of structural issues. But it's the output of a conscious mind attempting to convey an idea. Maybe I'm an idealist, but I think there's something special about real communication. Sure, if you're a massive bullshitter you're no better than AI. But if you even have w kernel of genuine emotion behind your writing, you have an edge over AI.
Sorry, I don't mean that is good in the "this is a pleasure to read" sense, but a "this has synthesized a lot of info quickly" sense. I use it for work and it's very good at synthesizing documents...
Sorry, I don't mean that is good in the "this is a pleasure to read" sense, but a "this has synthesized a lot of info quickly" sense. I use it for work and it's very good at synthesizing documents and generating reports. Like I'll feed a technical write up and a skeleton work plan and it can help generate a year long development roadmap. I could likely write something better if you gave me a week, but using GPT I can churn one out in about 30 minutes with a few read throughs and fixes. Before GPT I would not have had a week to write it, so people would just have to deal with my skeleton outline and ask a ton of questions. GPTs writing is not enjoyable or pleasurable, it's just good at reducing busywork.
So I may have been wrong in how I framed it. Better in that it can synthesize info, be it with human checks. But not better in any sense of pleasure or prose.
Honestly, the only use I've found for it is communication that no one reads. If I have some weekly status report or meeting summary or something that I know people want for some reason, but won't...
Honestly, the only use I've found for it is communication that no one reads.
If I have some weekly status report or meeting summary or something that I know people want for some reason, but won't actually read? Yeah, it's a good task for an LLM.
If it's something I think people will actually need to glean useful information from, I need to review it anyway. It's almost always filled with so many inaccuracies and fluffy language that I need to rewrite it anyway. At that point I have to wonder if it's even saving me any time.
It's funny all the replies I've received are making me question the quality of the LLM generated documents I've been using. I've followed up with a few direct reports to make sure I'm not just...
It's funny all the replies I've received are making me question the quality of the LLM generated documents I've been using. I've followed up with a few direct reports to make sure I'm not just sending them useless material (it always ends up reading well to me after a few iterations) and so far everyone's been happy with it. They all know my process since we're pretty transparent about our workflows but I'm still nervous now. I think next week I'll write 2, one by myself and on with GPT and give them an apples to apples comparison and then have a quick checkin to see what is more beneficial to them: the quality improvement of my spending my time on those (if any) vs using my time elsewhere to pick up slack. It'll also be nice to see what the actual time difference as well.
Personally, I feel like I can very quickly tell when someone hands me unedited output from an LLM. Of course, that could be confirmation bias though. I'm only identifying the LLM generated...
Personally, I feel like I can very quickly tell when someone hands me unedited output from an LLM. Of course, that could be confirmation bias though. I'm only identifying the LLM generated documents I identify, and all the other LLM generated documents I get fly under my radar.
I don't feel like that's the case, because of my own experience generating documents with LLMs though.
For me, even if LLM generated documents were useful, getting one still feels disrespectful. It's almost like asking a question and someone replying back with a https://letmegooglethat.com/ link.
It kind of makes me fee like l I could have just done that easily myself, and the fact that you don't think I thought of that sort of insults my intelligence.
Like, it just feels cheap and disposable, and like there must be a better solution than to just have something dump something into an LLM, then paste the output into an email.
It's tough to get at exactly how I feel about it, but I just generally don't like receiving stuff like that.
I totally understand the feeling, I feel the same when I come across an email or article written by one. Like, really, you couldn't just write - "Sweet, look forward to chatting with you on...
I totally understand the feeling, I feel the same when I come across an email or article written by one. Like, really, you couldn't just write - "Sweet, look forward to chatting with you on friday. Be sure copy Monica so she knows what is happening too."
For the documents I'm referring to, I initially wrote all of them. As a team we decided that they weren't so useful as to eat up weeks of my time, for the most part the team just didn't need the really in depth explanation for why decisions were being made. So then we shifted to outlines of what needed to happen and when, what were the blockers/co-dependencies, and resource needs. Which was fine but ended up with a lot of interpretation as to why and lead to miscommunications in how things should be done. So we ended up with more meeting time to walk through them which also wasn't helpful. We kept trying to find the goldilocks zone and ended up trying out GPT. I fed it our overarching planning docs, the outlines, and any relevant information directing the work (customer feedback, parallel development plans, etc) and with enough feedback ended up creating a pretty decent document. Most discussions stay within our morning standup now and the time it takes to make them has seemingly been less. After typing this all out, maybe I won't rock the boat. What's the saying "don't try to make a happy baby happier"?
You just triggered a flood of memories of my writing being interpreted in ways completely different from what I intended. It immediately took me back to Roland Barthes’ The Death of the Author....
it's the output of a conscious mind attempting to convey an idea. Maybe I'm an idealist, but I think there's something special about real communication.
You just triggered a flood of memories of my writing being interpreted in ways completely different from what I intended. It immediately took me back to Roland Barthes’ The Death of the Author. Anyone who’s written enough and received feedback knows what an absolute mindfuck this can be. You get notes on sections you never wanted feedback on, while the parts you thought most deserved scrutiny are ignored. Worse, a reader can use your words to argue the exact opposite of what you believed you were plainly stating.
I honestly think AI could be an excellent proofreader because of this, especially given that society decided 20+ years ago that paying for editors was unnecessary. Editing was my dream job when I went to college, but even then I knew it wouldn’t be a viable career. At least now, anyone can have a kind of Socratic dialogue with an AI over their text before hitting submit.
There is a dark art to this in institutional settings for those willing to master the art of manipulation. You simply learn the editor’s idiosyncrasies and add trivial and easily corrected errors...
You get notes on sections you never wanted feedback on, while the parts you thought most deserved scrutiny are ignored.
There is a dark art to this in institutional settings for those willing to master the art of manipulation. You simply learn the editor’s idiosyncrasies and add trivial and easily corrected errors in and around a section where you don’t want them to think too hard about the content or its implications (politically). You’ll get edits on commas or word choice that you can happily edit.
I continue to be astounded how this has become a celebrated feature of human writing, to the point that it's now expected... leading to people who, like me, pay close attention to their grammar...
Human writing is full of grammatical mistakes
I continue to be astounded how this has become a celebrated feature of human writing, to the point that it's now expected... leading to people who, like me, pay close attention to their grammar and spelling, are compelled to intentionally insert errors (and eschew proper typography like em-dashes) just to assuage the fear of being accused of using AI to write.
You misunderstand me: it isn't supposed to be a good thing. I was attempting to contrast the bad parts of human writing (imperfections) with the good parts (intention, emotion, and novel ideas)....
You misunderstand me: it isn't supposed to be a good thing. I was attempting to contrast the bad parts of human writing (imperfections) with the good parts (intention, emotion, and novel ideas).
AI is great at not committing those sins. But it also lacks those merits.
I have edited hundreds of pages of human and AI-generated writing at this point for my job, and it doesn't take me long to understand when someone has submitted AI slop or bullshit. It then takes me ages to whittle the crap and bloat down to the three bullet points they originally provided the AI. I would much rather just read the human-generated lazy bullet points than wade through a sea of empty bullshit sentences generated because they probably sorta kinda perhaps fit into the context of those bullet points.
Human language is a beautiful kernel of meaning wrapped in a hard shell of imperfection. LLM output is fucking bullshit.
To me this is a big tell for artificial writing, which is unfortunately getting harder to spot. It's not just the use of em dashes, or framing such as "it's not x, but y." Some LLM writing will...
To me this is a big tell for artificial writing, which is unfortunately getting harder to spot. It's not just the use of em dashes, or framing such as "it's not x, but y." Some LLM writing will use colloquial slang or typos, although if you read carefully you can see the slang used in the wrong way, or the same exact typo repeated three or four times. Fortunately, AI "writers" aren't careful enough to read their own writing, so giveaways like that can be easy to spot.
But it does smell like bullshit, and after a lifetime of experiencing spam, product placement and advertising everywhere, it feels incredibly manipulative. In my view, even the most engaging content loses it's interest and becomes repulsive if it's written in bad faith. Of course this doesn't just apply to LLM writing, but it accounts for a lot of crap you already see in media, and the current state of affairs has been a long time coming.
Artificial writing has no soul. It can have character, it can be engaging, it can be interesting, clever, or fun, but it can't be real. That's the problem with some made-up stories, and currently what I feel is a tell, which is that when you think about these stories in three dimensions over the passage of time, you realize that real people don't behave like that.
But I feel like the biggest giveaway is basically how I feel when I'm reading it. If I feel excited, engaged and interested, or perhaps placated, comfortable, and validated, then I have second thoughts and I look again. Manipulative media behaves that way - it makes you feel good about yourself, it makes you want to keep watching, and makes you feel comfortable. It's made for engagement. I'm fine with that when it's on my own terms, but not when it's driven by some hidden agenda.
Maybe it's ironic that I feel the most comfortable with content that makes me less comfortable. But I'd rather have something real than be complacent.
To your comment, what kind of writing do you do and how does LLM writing show up in your world?
I do technical writing, documentation for software and hardware products, specifically. I've seen quite a lot of chatter in my industry about LLMs writing documentation for engineers based on PRs...
I do technical writing, documentation for software and hardware products, specifically.
I've seen quite a lot of chatter in my industry about LLMs writing documentation for engineers based on PRs and scope docs and JIRA tickets. All of the output I've seen has been bloated, poorly organized, often subtly incorrect trash. Can't follow style guides. Can't verify code snippet correctness. Can't verify high-level systems description correctness.
It's kind of like having a clueless high-level manager drop into the docs codebase with sweeping PRs. More work to verify, correct, and understand than it would be to simply write from scratch.
I think the full argument of the video is that the important part of “writing” is that you can imagine something, put it in words, and then transfer that idea to someone else across time and...
I think the full argument of the video is that the important part of “writing” is that you can imagine something, put it in words, and then transfer that idea to someone else across time and space.
LLMs don’t have “ideas” in the same way we do; this is why the author belabors the point “there can be no meeting of the mind” when discussing with an LLM. There’s no communication happening.
LLMs are good at language (something the author says verbatim), but their primary argument is writing requires a human connection.
Having watched the video I think you are right in that this is a central argument the video is making. LLMs can not generate new ideas only respond to input. However, I think there is an argument...
Having watched the video I think you are right in that this is a central argument the video is making. LLMs can not generate new ideas only respond to input.
However, I think there is an argument to be made that you are communicating with someone, albeit its a few more steps removed than with "regular" writing. Someone wrote the system prompt for the LL, someone trained the LLM, someone wrote the tools to train the LLM, and someone wrote the corpus of text the LLM is trained on. The intention of the system prompt writer and also the trainer is probably the most significant here since they are the ones "seeding" the LLM and "selecting" what the LLM is trained on. As such an argument could be made that your connecting with them mediated through the LLM similar to how we can connect with authors of regular text through their works.
I think anyone who's ever read any online erotica will tell you for certain that even given time, resources and horniness, you are not necessarily a better writer hahahahaha
I think anyone who's ever read any online erotica will tell you for certain that even given time, resources and horniness, you are not necessarily a better writer hahahahaha
I know that I'm telling on myself here, but no, any AI smut, even made by the dumbest models from 2-3 years ago that could run on potato PCs are a better experience than average online smut (I...
I know that I'm telling on myself here, but no, any AI smut, even made by the dumbest models from 2-3 years ago that could run on potato PCs are a better experience than average online smut (I don't have much experience with book smut). And it completely, unequivocally, undoubtedly beats the fascinating experience of finding a half-decent writer on any ERP website. Anyone who has had experience there will tell you just how godawful, low effort and mind mindbogglingly terrible absolute most users there are at writing smut. It is truly an arduous process of wading through a shit lake for bits of pieces of buried treasure. I'm just gonna assume your disagreement here is philosophical rather than practical.
I disagree and no, your assumption is wrong. Speaking explicitly about online: Bad human written smut is a quick skip (and you gotta use some discretion in what you're reading anyway), AI smut...
I disagree and no, your assumption is wrong. Speaking explicitly about online: Bad human written smut is a quick skip (and you gotta use some discretion in what you're reading anyway), AI smut tries to trick you that it's good and in failing ruins the experience. It's upsettingly bad. Good human written smut is excellent. AI smut never gets there.
I don't ERP anymore so I can't compare it to doing so with AI but the point there has always been a human connection for me. I don't have any interest in pretending the machine is a person* gestures at psychosis incidences
But if you're not using whatever rating/tagging system or author tracking exists you're setting yourself up to read a lot of junk regardless. It's just more AI junk now.
I stand by my statement and I have over 2 decades of online smut reading experience.
I don't even know what that means. All I want when it comes to smut is to be satisfied. If it "tricks" you into being satisfied, then it is good. That's wading through junk in search for treasure...
AI smut tries to trick you that it's good
I don't even know what that means. All I want when it comes to smut is to be satisfied. If it "tricks" you into being satisfied, then it is good.
But if you're not using whatever rating/tagging system or author tracking exists you're setting yourself up to read a lot of junk regardless.
That's wading through junk in search for treasure and it's not enough. The vast majority of what you encounter written by humans is going to be junk, which was my point. To put it another way, it is vastly preferable to prompt an AI and get what you want instantly than wading through endless pages hoping to find that diamond in the rough.
AI smut never gets there.
Yeah it does. I don't need my smut to be philosophically soulful art gallery material or whatever. The difference in how fast you can get what you want to see (which to me is all that matters for smut) is just night and day.
It isn't satisfying and what it tricks me into is "thinking itll be satisfying and not being. " For me, the AI material is 100% junk vs the human work which has plenty of junk and plenty of good...
I don't even know what that means. All I want when it comes to smut is to be satisfied. If it "tricks" you into being satisfied, then it is good.
It isn't satisfying and what it tricks me into is "thinking itll be satisfying and not being. "
That's wading through junk in search for treasure and it's not enough. The vast majority of what you encounter written by humans is going to be junk, which was my point. To put it another way, it is vastly preferable to prompt an AI and get what you want instantly than wading through endless pages hoping to find that diamond in the rough.
For me, the AI material is 100% junk vs the human work which has plenty of junk and plenty of good shit.
Yeah it does. I don't need my smut to be philosophically soulful art gallery material or whatever. The difference in how fast you can get what you want to see (which to me is all that matters for smut) is just night and day.
I want quality not "philosophically soulful." I can get junk anywhere. And all AI work is junk for me. I'm not telling you what has to work for you. But this started with:
I think anyone who's ever read any online erotica will tell you for certain that even given time, resources and horniness, you are not necessarily a better writer hahahahaha
That's why I said I think your disagreement is philosophical rather than practical. Otherwise you'd be naming models, which obviously all write differently, prompts and settings which obviously...
And all AI work is junk for me.
That's why I said I think your disagreement is philosophical rather than practical. Otherwise you'd be naming models, which obviously all write differently, prompts and settings which obviously affect this. Listen, I respect your hardline anti-AI stance, I get responses from you practically every single time I say something that isn't "AI is always terrible" on tildes. Just don't present it as an actual comparison of the results.
Well I'd prefer you stop telling me what I think. I'm not going and asking models for erotica. But there are plenty of people who post their AI stories on various sites so I've read my share. So I...
That's why I said I think your disagreement is philosophical rather than practical. Otherwise you'd be naming models, which obviously all write differently, prompts and settings which obviously affect this. Listen, I respect your hardline anti-AI stance, I get responses from you practically every single time I say something that isn't "AI is always terrible" on tildes. Just don't present it as an actual comparison of the results.
Well I'd prefer you stop telling me what I think. I'm not going and asking models for erotica. But there are plenty of people who post their AI stories on various sites so I've read my share.
So I am comparing human and AI written works and find the results different.
I understood them, I just didn't agree with their blanket statement and provided my opinion, first in a pithy way and then in more detail to clarify. I just don't understand the point of their...
I understood them, I just didn't agree with their blanket statement and provided my opinion, first in a pithy way and then in more detail to clarify. I just don't understand the point of their last one in the context of that conversation.
That’s the opposite of the thrust of the video actually. It’s pretty explicit that even hastily scribbled off dreck—or even a reaction gif—is better writing than anything the chatbot spits out.
That’s the opposite of the thrust of the video actually. It’s pretty explicit that even hastily scribbled off dreck—or even a reaction gif—is better writing than anything the chatbot spits out.
I've been thinking on this and I guess the reaction is because in a lot of cases people are not looking for "writing" , but instead for "output". If we interpret writing as the meeting of the...
I've been thinking on this and I guess the reaction is because in a lot of cases people are not looking for "writing" , but instead for "output". If we interpret writing as the meeting of the minds then what happens when we know (or assume) the mind on the other side will not meet ours?
It's easier to judge output than writing (I hope this is a meaningful statement) so we who write and reach out into the world just as easily fret over whether our writing is also good output, because our reader may first judge us on our output before they meet us on our writing.
His final point is exactly what I really don't get about people who become addicted to befriending their chosen language model as if it's a person, it's not a person and they know that its not a...
His final point is exactly what I really don't get about people who become addicted to befriending their chosen language model as if it's a person, it's not a person and they know that its not a person. They're essentially talking to themselves because all the model ever does it tell you what you want to hear. So they're becoming addicted to talking to themselves? Isn't that basically what psychosis is?
Psychosis can manifest in a variety of ways, but at its core, it is a difficulty (or inability) to distinguish what is and isn't real. LLMs don't know what's real, if you give your thinking over...
Psychosis can manifest in a variety of ways, but at its core, it is a difficulty (or inability) to distinguish what is and isn't real. LLMs don't know what's real, if you give your thinking over to them, you soon won't either. Less "addicted to talking to yourself" and more "enraptured with a sycophant".
Not to sound insensitive, but I think folks treating LLMs as friends and partners are on the brink, primed by tech companys' marketing. I think typical views of psychosis are naïve and far more -- otherwise healthy -- individuals are susceptible than we'd like to think (I'd go so far as to say everyone is susceptible in the wrong circumstances). As is, folks suffering are easily othered, and that sort of isolation makes help more difficult.
Some people just really crave affirmation so this machine that affirms by default just seems to trip a psychological land-mine for them. Thankfully I was raised by an Indian mother, so the...
Some people just really crave affirmation so this machine that affirms by default just seems to trip a psychological land-mine for them.
Thankfully I was raised by an Indian mother, so the experience of being affirmed is sufficiently foreign to me that it makes me recoil with the fear that I might be interacting with The Beldam. I believe this has partially inoculated me against the AI cognitohazard.
Hm, I feel that I'm a counterexample to this. I do crave affirmation, validation and praise because I was raised by narcissistic parents. I even occasionally notice myself getting a minor high...
Hm, I feel that I'm a counterexample to this. I do crave affirmation, validation and praise because I was raised by narcissistic parents. I even occasionally notice myself getting a minor high when an LLM prefixes its response to my prompt with something like “that's an astute observation!” or similar. Despite, it doesn't appear to be a “psychological land-mine” for me. I don't go to the LLM to seek it out specifically, and I certainly don't seem to be falling into any kind of psychosis with it.
I can’t even really describe the level of discomfort I feel when it says things like that. Even when a human says it I feel like I’m being brown-nosed and become reflexively hostile to the person...
I can’t even really describe the level of discomfort I feel when it says things like that. Even when a human says it I feel like I’m being brown-nosed and become reflexively hostile to the person because it makes me feel like they’re appealing to my emotional state rather than my logic-brain. But when a statistical machine is doing it then it’s inherently something darker. It’s a design choice made explicitly to manipulate me.
It’s a nice feedback loop. Though supposedly with these new updates it’s suppose to not do that. It’s supposed to be more discerning about what it tells you so you don’t feel too in your head....
It’s a nice feedback loop. Though supposedly with these new updates it’s suppose to not do that. It’s supposed to be more discerning about what it tells you so you don’t feel too in your head. Such as if you interpreter a social interaction a certain way, there’s guardrails so it doesn’t go “of course that’s what happened.” Because of all the suicide stories that have come out of these AI psychosis bouts.
I’ve dealt with my own version of psychosis from this. But not to the point where I think I’m talking to Samantha from Her. It’s just not that sophisticated.
...about two-thirds of our studio speak fluent english as their second language and the incidence off LLM-generated slop i find buried in our deliverables is atrocious, so i resoundingly advocate...
...about two-thirds of our studio speak fluent english as their second language and the incidence off LLM-generated slop i find buried in our deliverables is atrocious, so i resoundingly advocate that original thought with technically-poor english conveys a VASTLY stronger product than vacuous but technically-well-composed filler, but have made little headway amongst staff with low confidence in their written english proficiency...
...i'm grateful to see this video convey that message, but fourty minutes' attention is too great an ask of peoples' time, which i suppose brings the issue full circle...
"IF ANYONE CAN GENERATE THIS SLOP IN TWO MINUTES THEN WHY IS OUR CLIENT PAYING US?"
It seems like there ought to be a way to use an LLM to help people with mediocre English skills that improves grammar and usage without making other things worse? But getting that to happen might...
It seems like there ought to be a way to use an LLM to help people with mediocre English skills that improves grammar and usage without making other things worse? But getting that to happen might require someone to write new software to provide the right scaffolding.
Maybe it will get better at some companies as best practices become better known, but it will take time to develop those.
...our CEO has thrived despite imperfect english for thirty years (example one in how ideas matter in our profession, not technical expression) and he's recently taken to writing whatever he...
...our CEO has thrived despite imperfect english for thirty years (example one in how ideas matter in our profession, not technical expression) and he's recently taken to writing whatever he intends to express, then running it through chatGPT for technical cleanup, then making a final edit pass to make sure his original ideas survive intact...
...that's one of the better examples i've seen of using the tool for its strengths rather than a BS generator...
I didn't know BFI has been doing annual Sight and Sound polls for best video essays. This was apparently one of the top nominees last year. I think I broadly agree with the conclusion and the...
I think I broadly agree with the conclusion and the rationale, but this is one of the most disorienting videos I've watched in a long time so I'm not even confident in how well I've examined the argument. Following the thread of Stein's lyricism through Gen Z memes is hard enough; I think the desktop DJ format, as one YouTube commenter put it, and timing onscreen just made it tougher.
Oh I actually really liked the format. So much so, in fact, that I might adopt it the next time I have to do a work presentation instead of relying on PowerPoint.
Oh I actually really liked the format. So much so, in fact, that I might adopt it the next time I have to do a work presentation instead of relying on PowerPoint.
Idk, these always feel like... cope, to me. In this video and in many cases, the argument proceeds by first redefining "good writing" into something such that only humans can be good, or even...
Idk, these always feel like... cope, to me. In this video and in many cases, the argument proceeds by first redefining "good writing" into something such that only humans can be good, or even applicable, in the new definition, and then arguing how that is the case.
But the danger there is, is that the definition other people use? Is it the definition that matters to you? Like it or not, the reality is that most writers desire for other people to like and enjoy their writing. Does it matter if AI text is always inferior when defined about the amount of intention a sentient human is trying to transmit, if in practice AI produced text is perfectly attractive to human readers who then proceed to spend their time reading the AI text instead of yours?
I feel like it sets people up with unrealistic expectations of the future.
The overall concept also gets to things like death of the author. If writing is about the transmission of ideas, what if you gain meaning from a piece of writing that the author didn't intend at all?
Or to go to art, is a painting of a landscape superior to viewing it yourself, since the latter is the creation of inhuman natural forces with no concern for design or aesthetic principle?
IMO, regardless of the capability of our current systems to do, it is fundamentally possible for machines to create text that is
Appealing to humans
Meaningful to humans
if nothing else in a monkey-typewriter-shakespeare way. And that may still be inferior if we value the "soul" of the writing imbued by the writer, but also, does that matter?
I think I'll add the caveat to the video - You are a better write than AI, if given the time and resources to do so. That statement is so so so true. But in a world of low stakes, synthesis I still thing GPT has it place. Just not in academia.
As a professional writer, I have to disagree. LLM writing is painful to read. It feels like someone bullshitting, and no matter how much people prompt it, it never quite escapes that "uncanny valley" for me.
Human writing is full of grammatical mistakes, imperfections, typos, idioms, cliches, and all kinds of structural issues. But it's the output of a conscious mind attempting to convey an idea. Maybe I'm an idealist, but I think there's something special about real communication. Sure, if you're a massive bullshitter you're no better than AI. But if you even have w kernel of genuine emotion behind your writing, you have an edge over AI.
Sorry, I don't mean that is good in the "this is a pleasure to read" sense, but a "this has synthesized a lot of info quickly" sense. I use it for work and it's very good at synthesizing documents and generating reports. Like I'll feed a technical write up and a skeleton work plan and it can help generate a year long development roadmap. I could likely write something better if you gave me a week, but using GPT I can churn one out in about 30 minutes with a few read throughs and fixes. Before GPT I would not have had a week to write it, so people would just have to deal with my skeleton outline and ask a ton of questions. GPTs writing is not enjoyable or pleasurable, it's just good at reducing busywork.
So I may have been wrong in how I framed it. Better in that it can synthesize info, be it with human checks. But not better in any sense of pleasure or prose.
Honestly, the only use I've found for it is communication that no one reads.
If I have some weekly status report or meeting summary or something that I know people want for some reason, but won't actually read? Yeah, it's a good task for an LLM.
If it's something I think people will actually need to glean useful information from, I need to review it anyway. It's almost always filled with so many inaccuracies and fluffy language that I need to rewrite it anyway. At that point I have to wonder if it's even saving me any time.
It usually isn't.
It's funny all the replies I've received are making me question the quality of the LLM generated documents I've been using. I've followed up with a few direct reports to make sure I'm not just sending them useless material (it always ends up reading well to me after a few iterations) and so far everyone's been happy with it. They all know my process since we're pretty transparent about our workflows but I'm still nervous now. I think next week I'll write 2, one by myself and on with GPT and give them an apples to apples comparison and then have a quick checkin to see what is more beneficial to them: the quality improvement of my spending my time on those (if any) vs using my time elsewhere to pick up slack. It'll also be nice to see what the actual time difference as well.
Personally, I feel like I can very quickly tell when someone hands me unedited output from an LLM. Of course, that could be confirmation bias though. I'm only identifying the LLM generated documents I identify, and all the other LLM generated documents I get fly under my radar.
I don't feel like that's the case, because of my own experience generating documents with LLMs though.
For me, even if LLM generated documents were useful, getting one still feels disrespectful. It's almost like asking a question and someone replying back with a https://letmegooglethat.com/ link.
It kind of makes me fee like l I could have just done that easily myself, and the fact that you don't think I thought of that sort of insults my intelligence.
Like, it just feels cheap and disposable, and like there must be a better solution than to just have something dump something into an LLM, then paste the output into an email.
It's tough to get at exactly how I feel about it, but I just generally don't like receiving stuff like that.
I totally understand the feeling, I feel the same when I come across an email or article written by one. Like, really, you couldn't just write - "Sweet, look forward to chatting with you on friday. Be sure copy Monica so she knows what is happening too."
For the documents I'm referring to, I initially wrote all of them. As a team we decided that they weren't so useful as to eat up weeks of my time, for the most part the team just didn't need the really in depth explanation for why decisions were being made. So then we shifted to outlines of what needed to happen and when, what were the blockers/co-dependencies, and resource needs. Which was fine but ended up with a lot of interpretation as to why and lead to miscommunications in how things should be done. So we ended up with more meeting time to walk through them which also wasn't helpful. We kept trying to find the goldilocks zone and ended up trying out GPT. I fed it our overarching planning docs, the outlines, and any relevant information directing the work (customer feedback, parallel development plans, etc) and with enough feedback ended up creating a pretty decent document. Most discussions stay within our morning standup now and the time it takes to make them has seemingly been less. After typing this all out, maybe I won't rock the boat. What's the saying "don't try to make a happy baby happier"?
You just triggered a flood of memories of my writing being interpreted in ways completely different from what I intended. It immediately took me back to Roland Barthes’ The Death of the Author. Anyone who’s written enough and received feedback knows what an absolute mindfuck this can be. You get notes on sections you never wanted feedback on, while the parts you thought most deserved scrutiny are ignored. Worse, a reader can use your words to argue the exact opposite of what you believed you were plainly stating.
I honestly think AI could be an excellent proofreader because of this, especially given that society decided 20+ years ago that paying for editors was unnecessary. Editing was my dream job when I went to college, but even then I knew it wouldn’t be a viable career. At least now, anyone can have a kind of Socratic dialogue with an AI over their text before hitting submit.
There is a dark art to this in institutional settings for those willing to master the art of manipulation. You simply learn the editor’s idiosyncrasies and add trivial and easily corrected errors in and around a section where you don’t want them to think too hard about the content or its implications (politically). You’ll get edits on commas or word choice that you can happily edit.
I continue to be astounded how this has become a celebrated feature of human writing, to the point that it's now expected... leading to people who, like me, pay close attention to their grammar and spelling, are compelled to intentionally insert errors (and eschew proper typography like em-dashes) just to assuage the fear of being accused of using AI to write.
i dont’ get how that's supposed be a good think.
You misunderstand me: it isn't supposed to be a good thing. I was attempting to contrast the bad parts of human writing (imperfections) with the good parts (intention, emotion, and novel ideas).
AI is great at not committing those sins. But it also lacks those merits.
I have edited hundreds of pages of human and AI-generated writing at this point for my job, and it doesn't take me long to understand when someone has submitted AI slop or bullshit. It then takes me ages to whittle the crap and bloat down to the three bullet points they originally provided the AI. I would much rather just read the human-generated lazy bullet points than wade through a sea of empty bullshit sentences generated because they probably sorta kinda perhaps fit into the context of those bullet points.
Human language is a beautiful kernel of meaning wrapped in a hard shell of imperfection. LLM output is fucking bullshit.
To me this is a big tell for artificial writing, which is unfortunately getting harder to spot. It's not just the use of em dashes, or framing such as "it's not x, but y." Some LLM writing will use colloquial slang or typos, although if you read carefully you can see the slang used in the wrong way, or the same exact typo repeated three or four times. Fortunately, AI "writers" aren't careful enough to read their own writing, so giveaways like that can be easy to spot.
But it does smell like bullshit, and after a lifetime of experiencing spam, product placement and advertising everywhere, it feels incredibly manipulative. In my view, even the most engaging content loses it's interest and becomes repulsive if it's written in bad faith. Of course this doesn't just apply to LLM writing, but it accounts for a lot of crap you already see in media, and the current state of affairs has been a long time coming.
Artificial writing has no soul. It can have character, it can be engaging, it can be interesting, clever, or fun, but it can't be real. That's the problem with some made-up stories, and currently what I feel is a tell, which is that when you think about these stories in three dimensions over the passage of time, you realize that real people don't behave like that.
But I feel like the biggest giveaway is basically how I feel when I'm reading it. If I feel excited, engaged and interested, or perhaps placated, comfortable, and validated, then I have second thoughts and I look again. Manipulative media behaves that way - it makes you feel good about yourself, it makes you want to keep watching, and makes you feel comfortable. It's made for engagement. I'm fine with that when it's on my own terms, but not when it's driven by some hidden agenda.
Maybe it's ironic that I feel the most comfortable with content that makes me less comfortable. But I'd rather have something real than be complacent.
To your comment, what kind of writing do you do and how does LLM writing show up in your world?
I do technical writing, documentation for software and hardware products, specifically.
I've seen quite a lot of chatter in my industry about LLMs writing documentation for engineers based on PRs and scope docs and JIRA tickets. All of the output I've seen has been bloated, poorly organized, often subtly incorrect trash. Can't follow style guides. Can't verify code snippet correctness. Can't verify high-level systems description correctness.
It's kind of like having a clueless high-level manager drop into the docs codebase with sweeping PRs. More work to verify, correct, and understand than it would be to simply write from scratch.
I think the full argument of the video is that the important part of “writing” is that you can imagine something, put it in words, and then transfer that idea to someone else across time and space.
LLMs don’t have “ideas” in the same way we do; this is why the author belabors the point “there can be no meeting of the mind” when discussing with an LLM. There’s no communication happening.
LLMs are good at language (something the author says verbatim), but their primary argument is writing requires a human connection.
Having watched the video I think you are right in that this is a central argument the video is making. LLMs can not generate new ideas only respond to input.
However, I think there is an argument to be made that you are communicating with someone, albeit its a few more steps removed than with "regular" writing. Someone wrote the system prompt for the LL, someone trained the LLM, someone wrote the tools to train the LLM, and someone wrote the corpus of text the LLM is trained on. The intention of the system prompt writer and also the trainer is probably the most significant here since they are the ones "seeding" the LLM and "selecting" what the LLM is trained on. As such an argument could be made that your connecting with them mediated through the LLM similar to how we can connect with authors of regular text through their works.
I think anyone who's ever read any online erotica will tell you for certain that even given time, resources and horniness, you are not necessarily a better writer hahahahaha
Nah reading AI smut is worse than bad human smut.
I know that I'm telling on myself here, but no, any AI smut, even made by the dumbest models from 2-3 years ago that could run on potato PCs are a better experience than average online smut (I don't have much experience with book smut). And it completely, unequivocally, undoubtedly beats the fascinating experience of finding a half-decent writer on any ERP website. Anyone who has had experience there will tell you just how godawful, low effort and mind mindbogglingly terrible absolute most users there are at writing smut. It is truly an arduous process of wading through a shit lake for bits of pieces of buried treasure. I'm just gonna assume your disagreement here is philosophical rather than practical.
I disagree and no, your assumption is wrong. Speaking explicitly about online: Bad human written smut is a quick skip (and you gotta use some discretion in what you're reading anyway), AI smut tries to trick you that it's good and in failing ruins the experience. It's upsettingly bad. Good human written smut is excellent. AI smut never gets there.
I don't ERP anymore so I can't compare it to doing so with AI but the point there has always been a human connection for me. I don't have any interest in pretending the machine is a person* gestures at psychosis incidences
But if you're not using whatever rating/tagging system or author tracking exists you're setting yourself up to read a lot of junk regardless. It's just more AI junk now.
I stand by my statement and I have over 2 decades of online smut reading experience.
*Or whatever sapient being is relevant
I don't even know what that means. All I want when it comes to smut is to be satisfied. If it "tricks" you into being satisfied, then it is good.
That's wading through junk in search for treasure and it's not enough. The vast majority of what you encounter written by humans is going to be junk, which was my point. To put it another way, it is vastly preferable to prompt an AI and get what you want instantly than wading through endless pages hoping to find that diamond in the rough.
Yeah it does. I don't need my smut to be philosophically soulful art gallery material or whatever. The difference in how fast you can get what you want to see (which to me is all that matters for smut) is just night and day.
It isn't satisfying and what it tricks me into is "thinking itll be satisfying and not being. "
For me, the AI material is 100% junk vs the human work which has plenty of junk and plenty of good shit.
I want quality not "philosophically soulful." I can get junk anywhere. And all AI work is junk for me. I'm not telling you what has to work for you. But this started with:
And I don't agree!
That's why I said I think your disagreement is philosophical rather than practical. Otherwise you'd be naming models, which obviously all write differently, prompts and settings which obviously affect this. Listen, I respect your hardline anti-AI stance, I get responses from you practically every single time I say something that isn't "AI is always terrible" on tildes. Just don't present it as an actual comparison of the results.
Well I'd prefer you stop telling me what I think. I'm not going and asking models for erotica. But there are plenty of people who post their AI stories on various sites so I've read my share.
So I am comparing human and AI written works and find the results different.
Well let's just hope that nobody judges digital artists by the crap that's posted on rule34 and deviantart then :)
I don't understand what your point is here
I think the point at which you both lost each other was much earlier in the thread than here.
I understood them, I just didn't agree with their blanket statement and provided my opinion, first in a pithy way and then in more detail to clarify. I just don't understand the point of their last one in the context of that conversation.
For me specifically, it's place is helping me write those useless bullshit quarterly self reviews my company makes me do.
That’s the opposite of the thrust of the video actually. It’s pretty explicit that even hastily scribbled off dreck—or even a reaction gif—is better writing than anything the chatbot spits out.
I've been thinking on this and I guess the reaction is because in a lot of cases people are not looking for "writing" , but instead for "output". If we interpret writing as the meeting of the minds then what happens when we know (or assume) the mind on the other side will not meet ours?
It's easier to judge output than writing (I hope this is a meaningful statement) so we who write and reach out into the world just as easily fret over whether our writing is also good output, because our reader may first judge us on our output before they meet us on our writing.
His final point is exactly what I really don't get about people who become addicted to befriending their chosen language model as if it's a person, it's not a person and they know that its not a person. They're essentially talking to themselves because all the model ever does it tell you what you want to hear. So they're becoming addicted to talking to themselves? Isn't that basically what psychosis is?
Psychosis can manifest in a variety of ways, but at its core, it is a difficulty (or inability) to distinguish what is and isn't real. LLMs don't know what's real, if you give your thinking over to them, you soon won't either. Less "addicted to talking to yourself" and more "enraptured with a sycophant".
Not to sound insensitive, but I think folks treating LLMs as friends and partners are on the brink, primed by tech companys' marketing. I think typical views of psychosis are naïve and far more -- otherwise healthy -- individuals are susceptible than we'd like to think (I'd go so far as to say everyone is susceptible in the wrong circumstances). As is, folks suffering are easily othered, and that sort of isolation makes help more difficult.
Some people just really crave affirmation so this machine that affirms by default just seems to trip a psychological land-mine for them.
Thankfully I was raised by an Indian mother, so the experience of being affirmed is sufficiently foreign to me that it makes me recoil with the fear that I might be interacting with The Beldam. I believe this has partially inoculated me against the AI cognitohazard.
Hm, I feel that I'm a counterexample to this. I do crave affirmation, validation and praise because I was raised by narcissistic parents. I even occasionally notice myself getting a minor high when an LLM prefixes its response to my prompt with something like “that's an astute observation!” or similar. Despite, it doesn't appear to be a “psychological land-mine” for me. I don't go to the LLM to seek it out specifically, and I certainly don't seem to be falling into any kind of psychosis with it.
I can’t even really describe the level of discomfort I feel when it says things like that. Even when a human says it I feel like I’m being brown-nosed and become reflexively hostile to the person because it makes me feel like they’re appealing to my emotional state rather than my logic-brain. But when a statistical machine is doing it then it’s inherently something darker. It’s a design choice made explicitly to manipulate me.
Having been raised in a similar environment I also don’t find comfort in being reaffirmed like that lol
It’s a nice feedback loop. Though supposedly with these new updates it’s suppose to not do that. It’s supposed to be more discerning about what it tells you so you don’t feel too in your head. Such as if you interpreter a social interaction a certain way, there’s guardrails so it doesn’t go “of course that’s what happened.” Because of all the suicide stories that have come out of these AI psychosis bouts.
I’ve dealt with my own version of psychosis from this. But not to the point where I think I’m talking to Samantha from Her. It’s just not that sophisticated.
Yes, it’s a 40 minute video essay but hear him out! It’s very good!
...about two-thirds of our studio speak fluent english as their second language and the incidence off LLM-generated slop i find buried in our deliverables is atrocious, so i resoundingly advocate that original thought with technically-poor english conveys a VASTLY stronger product than vacuous but technically-well-composed filler, but have made little headway amongst staff with low confidence in their written english proficiency...
...i'm grateful to see this video convey that message, but fourty minutes' attention is too great an ask of peoples' time, which i suppose brings the issue full circle...
"IF ANYONE CAN GENERATE THIS SLOP IN TWO MINUTES THEN WHY IS OUR CLIENT PAYING US?"
It seems like there ought to be a way to use an LLM to help people with mediocre English skills that improves grammar and usage without making other things worse? But getting that to happen might require someone to write new software to provide the right scaffolding.
Maybe it will get better at some companies as best practices become better known, but it will take time to develop those.
...our CEO has thrived despite imperfect english for thirty years (example one in how ideas matter in our profession, not technical expression) and he's recently taken to writing whatever he intends to express, then running it through chatGPT for technical cleanup, then making a final edit pass to make sure his original ideas survive intact...
...that's one of the better examples i've seen of using the tool for its strengths rather than a BS generator...
I didn't know BFI has been doing annual Sight and Sound polls for best video essays. This was apparently one of the top nominees last year.
I think I broadly agree with the conclusion and the rationale, but this is one of the most disorienting videos I've watched in a long time so I'm not even confident in how well I've examined the argument. Following the thread of Stein's lyricism through Gen Z memes is hard enough; I think the desktop DJ format, as one YouTube commenter put it, and timing onscreen just made it tougher.
Oh I actually really liked the format. So much so, in fact, that I might adopt it the next time I have to do a work presentation instead of relying on PowerPoint.
Idk, these always feel like... cope, to me. In this video and in many cases, the argument proceeds by first redefining "good writing" into something such that only humans can be good, or even applicable, in the new definition, and then arguing how that is the case.
But the danger there is, is that the definition other people use? Is it the definition that matters to you? Like it or not, the reality is that most writers desire for other people to like and enjoy their writing. Does it matter if AI text is always inferior when defined about the amount of intention a sentient human is trying to transmit, if in practice AI produced text is perfectly attractive to human readers who then proceed to spend their time reading the AI text instead of yours?
I feel like it sets people up with unrealistic expectations of the future.
The overall concept also gets to things like death of the author. If writing is about the transmission of ideas, what if you gain meaning from a piece of writing that the author didn't intend at all?
Or to go to art, is a painting of a landscape superior to viewing it yourself, since the latter is the creation of inhuman natural forces with no concern for design or aesthetic principle?
IMO, regardless of the capability of our current systems to do, it is fundamentally possible for machines to create text that is
Appealing to humans
Meaningful to humans
if nothing else in a monkey-typewriter-shakespeare way. And that may still be inferior if we value the "soul" of the writing imbued by the writer, but also, does that matter?