- You can use AI to look through docs and do research.
- You can use AI to structure your argument.
- You can NOT use AI to actually write the prose.
- You can use AI to proofread, especially for how effectively you've communicated.
I don't know, man, that sounds a lot like using AI to write. Your writing probably says "load-bearing" less than Claude's outputs, but I'm not sure it's much less AI.
Reading through docs and doing research: of course. They're like much better search engines in that regard. It would be silly not to use them. Obviously, you need to review their search results against their sources.
Structuring arguments: bad idea. Here you're having the model do your thinking for you. The problems that occur with having it write text recur at a macro level; too-obvious patterns. Readers just want to know what the prompt was.
Writing the prose: obviously not.
Proofreading: I think this works pretty great, as long as you have a structure to how you're doing it. "How do I make this better or sharper": bad, you're having the model structure your arguments for you. But for a predefined rubric of stylistic things --- writing you know how to do, that's just tedious to check manually? Hard to see a downside.
Once upon a time the word processor was invented, and it made it so much easier to arrange your words on paper. iIf bits were wrong, you could just delete them, or copy and paste them across pages. It was actually pretty amazing what you could do; especially if you didn't mind dot-matrix type on fan-fold paper. Of course if you wanted to make a better impression, you'd type it on a typewriter, or better yet, hand-calligraph your writing with a fountain pen.
So LLMs allow you to work at the speed of reading, if not faster. So you can try a paragraph here, completely rewrite the tone. Start a chapter in 1PP and just have it rewritten to 3PP. And sofort and so on. The number of new transformations you can do is endless.
What do a lot of people do? Write a one line prompt and show everyone the oneshot result.
Ok fine, sometimes you don't need to write amazing prose. But you can also be like the early writers who adopted the word processor. If you put the same amount of elbow-grease into using an LLM as you did writing before, you can get some pretty cool results.
Well said, the issue being, the form factor of many AI products disincentivises this behaviour. To give an example, using a CLI tool with a frontier model to write a blog (actually fairly common), you have to re-open it in another editor, potentially reformat it or convert MD>DOCX, etc etc. There's a lot of noise and mess, and people naturally tend towards "not bother" and 1-shot some "amazing" prompt (because the value of their elbow grease is disproportionately low, unless they actually have something to say- many do not).
>Once upon a time the word processor was invented, and it made it so much easier to arrange your words on paper.
It’s easier to rearrange and rethink and tinker. Hard to imagine now but it was more efficient in the days when you would have to pencil out a memo and have someone type it up, you would make corrections and the final copy would be done in a couple passes. This has been studied, efficiency has actually gone down, in favor of immediacy and direct control.
What it frees you from is organizing your thoughts ahead of time, which takes practice and discipline. Using an LLM is just taking that further, you can avoid organizing your thoughts and ideas a bit more.
If someone uses LLM to help wording their idea in a way that's easier to understand, I am fine with it as long as they don't copy-paste it verbatim into their blogpost.
I am getting a little ranty here, but the usual LLM-isms like "not this, but that" or "that's the real footnote here" simply add zero value to the text, and I can't imagine anyone who has proofread the LLM-generated text even once will opt to keep those sloppy paragraphs. If I see them, I can't help but to think that the author had oneshotted the text. And it strongly suggests to me that the author has put zero effort into their work.
I disagree, the problem is the tool. Generating text instead of thinking about what text to use explicitly is not writing, and it's not worthy of someone else reading it.
Fundamental to exchanging ideas is discovering which ideas should be exchanged and how to express them. The process of writing itself is that exercise. If you use a tool to outsource the thinking then you're not thinking.
I definitely think that "not thinking about what you're writing" is a big danger with AI-generated text, but it's not a given. If you generate a block of text and then go through and edit it to better reflect your ideas and voice - that's the thinking process in action. It's also entirely possible to generate slop by manually typing every word - see any corporate communication ever.
You will NEVER be a good to great writer if you use AI to seed your writing and then edit it.
_Everything_ AI generates, is generalized slop. All of your ideas will be boring and derivative. Mediocre writers will be improved, but the best would only have their ability to think dulled.
Just think through it from the perspective that AI is generating art for you instead of writing. What would all art look like if good artists used it for ideas and "edited" it? It would look like generic, derivative shit. A perpetual crutch that the person can never abandon.
> Writing requires thinking, don't let an AI do the thinking for you.
Okay, but reading also requires substantial thinking. Moreover, we can say the same about, say, coding. And I am sure there are people who will say, yes, don't use AI to code. But I do, and the results are amazing, and I still think very deeply: in fact, I spend more time with AI coding (and thinking) than before as I do not get hamstrung on, how do I convert this Word doc to PDF in code again? I spend more time in higher order thinking. Same with when I read text written by an AI. I don't believe that writing is any more sacrosanct that coding. And reading and adjusting AI output requires substantial thinking.
I've been watching chess videos for many years but my own chess playing is just recreational.
Writing/programming as a specialization of just going through the damn thing is actually sacrosanct IMO. Spectatorship as a generalization of reading is second-class. You do become a worse programmer as you move to "higher order thinking", it just matters less in a world where LLMs exist.
> And reading and adjusting AI output requires substantial thinking.
It does, but less of that thinking translates to skills transferred. What amount of the creativity and magic of Tolkien transferred to his editor? I'm sure it took substantial thinking to edit Tolkien's work.
Perhaps not 1:1, but, personally, I can read for much longer than I can write. And so I'd argue that this duration increase matches any deeper thought that comes with shorter writing bursts.
Just this morning I used Claude to build an "omit needless words" app, which lets me paste in text and then marks up that text with strikes through the words that I can leave out.
I used it to edit a document I'm working on, where I accepted about a third of its suggestions. I feel pretty good about it.
If you have a Claude account you should be able to access the artifact I built here: https://claude.ai/artifact/RK7XfnaoxVV1Uji84ETaLr - it uses "Claudeception" - the feature of Claude artifacts that lets them use Claude prompts - so it will be billed against your Claude account (at a cost of probably a fraction of a cent per run.)
While I do agree with the premise of the article, if find this way of thinking applicable for writing code as well. Your code is meant to not only be read by machines, but also other fellow humans. Writing code requires substantial amount of thinking as well. Shouldn't we write code without AI as well?
I'm on version 6 of a writing an agent voice, style, tone, and orientation guide I have been developing over the course of the past year. Every time the agent writes something using the guide, I edit it and then update the guide where needed. Slowly, the agent is beginning to sound eerily close to my natural writing voice. No, I didn't use it for this comment :)
I use it to write in German, because my German sucks.
I first write the thing myself, then feed it through Deepl Write to see how the AI would write it. It's been very helpful.
I write everything else myself. However I don't know how I feel about spending my time on Earth typing personal responses to AI-generated messages. Do those people deserve my time and energy?
AI for proofreading/spelling correction and reordering works quite well. My words, but perhaps structured in a way which makes more sense, with sentiment replaced by synonyms.
Basically an automatic "second-pass" which I would've caught had I re-read it, but automatically.
Also, don't use AI to translate into a language you can't read back. It glosses over assumptions you didn't know you made, and the output might not mean what you intended.
I said this once in my recent "Trustworthy AI" class that... "I think writing with AI is one of the most detrimental things you can do".
The reaction was pretty much silence and most people probably thought I was some angry, arrogant, anti-ai purist... which isn't the case. It can be acceptable to use a language model in code generation but requires new forms of discipline and tooling.
But for for just regular old writing... Should never, ever be generated. I'll likely never change that stance.
That's not the argument in the article. Not only that but coding is deterministic up to a certain point. Once the project becomes complex, it becomes harder and harder to account for all edge cases. Sometimes it requires heuristics that are not deterministic, etc. There are infinite ways to implement the same requirements.
That argument is like saying that weather is deterministic. It is if you know every single variable, but that's impossible.
I'm not sure where people are going with using LLMs for research on a topic.
LLMs construct documents like a middle manager trying to get you fired, so it's a bad idea to get a system to try to research something for you. Anytime I've used the ol' panopticon-homunculus for something applied and niche I always follow the source linked to verify it. And the source is vague and generally unverifiable. Humans can actually link other documents that are actually readable to a greater degree.
I think a large part of the blame is the dataset for large scale LLMs being thoroughly and disproportionality poisoned with popular media sites like Reddit.
- You can use AI to look through docs and do research.
- You can use AI to structure your argument.
- You can NOT use AI to actually write the prose.
- You can use AI to proofread, especially for how effectively you've communicated.
I don't know, man, that sounds a lot like using AI to write. Your writing probably says "load-bearing" less than Claude's outputs, but I'm not sure it's much less AI.
Structuring arguments: bad idea. Here you're having the model do your thinking for you. The problems that occur with having it write text recur at a macro level; too-obvious patterns. Readers just want to know what the prompt was.
Writing the prose: obviously not.
Proofreading: I think this works pretty great, as long as you have a structure to how you're doing it. "How do I make this better or sharper": bad, you're having the model structure your arguments for you. But for a predefined rubric of stylistic things --- writing you know how to do, that's just tedious to check manually? Hard to see a downside.
So LLMs allow you to work at the speed of reading, if not faster. So you can try a paragraph here, completely rewrite the tone. Start a chapter in 1PP and just have it rewritten to 3PP. And sofort and so on. The number of new transformations you can do is endless.
What do a lot of people do? Write a one line prompt and show everyone the oneshot result.
Ok fine, sometimes you don't need to write amazing prose. But you can also be like the early writers who adopted the word processor. If you put the same amount of elbow-grease into using an LLM as you did writing before, you can get some pretty cool results.
It’s easier to rearrange and rethink and tinker. Hard to imagine now but it was more efficient in the days when you would have to pencil out a memo and have someone type it up, you would make corrections and the final copy would be done in a couple passes. This has been studied, efficiency has actually gone down, in favor of immediacy and direct control.
What it frees you from is organizing your thoughts ahead of time, which takes practice and discipline. Using an LLM is just taking that further, you can avoid organizing your thoughts and ideas a bit more.
If someone uses LLM to help wording their idea in a way that's easier to understand, I am fine with it as long as they don't copy-paste it verbatim into their blogpost.
I am getting a little ranty here, but the usual LLM-isms like "not this, but that" or "that's the real footnote here" simply add zero value to the text, and I can't imagine anyone who has proofread the LLM-generated text even once will opt to keep those sloppy paragraphs. If I see them, I can't help but to think that the author had oneshotted the text. And it strongly suggests to me that the author has put zero effort into their work.
Fundamental to exchanging ideas is discovering which ideas should be exchanged and how to express them. The process of writing itself is that exercise. If you use a tool to outsource the thinking then you're not thinking.
_Everything_ AI generates, is generalized slop. All of your ideas will be boring and derivative. Mediocre writers will be improved, but the best would only have their ability to think dulled.
Just think through it from the perspective that AI is generating art for you instead of writing. What would all art look like if good artists used it for ideas and "edited" it? It would look like generic, derivative shit. A perpetual crutch that the person can never abandon.
Why do people feel the need to invent the past that never existed?
Okay, but reading also requires substantial thinking. Moreover, we can say the same about, say, coding. And I am sure there are people who will say, yes, don't use AI to code. But I do, and the results are amazing, and I still think very deeply: in fact, I spend more time with AI coding (and thinking) than before as I do not get hamstrung on, how do I convert this Word doc to PDF in code again? I spend more time in higher order thinking. Same with when I read text written by an AI. I don't believe that writing is any more sacrosanct that coding. And reading and adjusting AI output requires substantial thinking.
Writing/programming as a specialization of just going through the damn thing is actually sacrosanct IMO. Spectatorship as a generalization of reading is second-class. You do become a worse programmer as you move to "higher order thinking", it just matters less in a world where LLMs exist.
> And reading and adjusting AI output requires substantial thinking.
It does, but less of that thinking translates to skills transferred. What amount of the creativity and magic of Tolkien transferred to his editor? I'm sure it took substantial thinking to edit Tolkien's work.
I don't think anyone would argue reading a book takes as much thinking as writing one…
https://www.scientificamerican.com/article/why-writing-by-ha...
I used it to edit a document I'm working on, where I accepted about a third of its suggestions. I feel pretty good about it.
If you have a Claude account you should be able to access the artifact I built here: https://claude.ai/artifact/RK7XfnaoxVV1Uji84ETaLr - it uses "Claudeception" - the feature of Claude artifacts that lets them use Claude prompts - so it will be billed against your Claude account (at a cost of probably a fraction of a cent per run.)
I don't write code for my colleague's pleasure, I do it to solve a problem as fast and efficiently as possible.
He’s five years older, a bit more tired than me, more going on in his life. Easily annoyed by bad code.
Weirdly enough he has exactly the same name and birthday as me.
I first write the thing myself, then feed it through Deepl Write to see how the AI would write it. It's been very helpful.
I write everything else myself. However I don't know how I feel about spending my time on Earth typing personal responses to AI-generated messages. Do those people deserve my time and energy?
Basically an automatic "second-pass" which I would've caught had I re-read it, but automatically.
https://markgreville.ie/2024/12/10/generative-ai-is-digital-...
The reaction was pretty much silence and most people probably thought I was some angry, arrogant, anti-ai purist... which isn't the case. It can be acceptable to use a language model in code generation but requires new forms of discipline and tooling.
But for for just regular old writing... Should never, ever be generated. I'll likely never change that stance.
> Writing requires thinking, don't let an AI do the thinking for you.
And then the author goes on to admit that they use AI for coding. Does coding not require thinking or what?
I agree that we shouldn't be delegating thinking to AI, but stopping at writing seems disingenuous.
Code is deterministic, and language is not. If I write:
total = 0 for x in [1, 2, 3]: total += x
…nobody is going to argue with me that “total” is anything but 7.
But language is different. What you perceive and what I perceive from the written word can be completely different (e.g https://youtube.com/watch?v=naleynXS7yo&ra=m )
That argument is like saying that weather is deterministic. It is if you know every single variable, but that's impossible.
- sure its good to ask AI feedback about your module or code review or identify vulnerabilities or better ways of doing your existing code
- but maybe stop asking AI to generate entire projects from scratch?
LLMs construct documents like a middle manager trying to get you fired, so it's a bad idea to get a system to try to research something for you. Anytime I've used the ol' panopticon-homunculus for something applied and niche I always follow the source linked to verify it. And the source is vague and generally unverifiable. Humans can actually link other documents that are actually readable to a greater degree.
I think a large part of the blame is the dataset for large scale LLMs being thoroughly and disproportionality poisoned with popular media sites like Reddit.