The meatiest piece of thinking I’ve read on AI writing. A fun ride. Makes me wonder if the big LLMs are putting concerted effort into increasing the quality and distinctiveness of the *writing* in their default-level outputs during RLHF, or if they are primarily focused on not-wrongness, figuring that users will tune diction and style at the account/project level.
This is the first piece I’ve read that so clearly describes the authorship issues many of us now face. In an effort to make our writing ‘more human’ (to avoid the AI slopslur) do we end up changing our voice and become less authentic, thus invalidating the human author value expectation? Do we deliberately use words that show human imagination - whacky compounds - or develop e e cummings-esque construction to ensure we are being heard as humans?
Really enjoyed this piece, guys! Thought-generation and its written expression are now the frontier for models for gaining relevance, within companies and on a wide society basis: my two-cents are that, still, passion, connecting with inner human emotions (lose the bus in the last minute) and aim to create beautiful things (maybe not useful economically) will remain in the human side. Seeing A16z writing on this pictures its relevance!
This is a really fun and thoughtful essay. People want authors for a bunch of reasons: we want to attribute virtue (or blame) for ideas that accrue to the people who create them, we want to connect with authors (“my favorite author”) because it creates community and signals something important about the reader…and because selling push button LLM prose as your own shows you just didn’t want to take the time to write and think it through for yourself. I get it and sympathize with all these positions. I don’t want the world of authors to disappear, and I love watching brilliant (human) writers play with language. But LLMS are a tool, and interacting with and using them to generate ideas or touch points seems not only fine, but basically inevitable. So let’s normalize what you did here: say what you used the tool for, if you used one, and why. That is still an author making a decision about what makes it into a text and why, and being prepared to defend it. If this was a science paper, it would go into the methods section.
AI can make me laugh out loud from time to time. There was the time I needed to assign a nickname to ChatGPT, as I have to the other models. Claude is Clyde, Grok is Gronk in honor of Gronkowski, the goofy tight end of the Pats now turned shoe salesman, etc.
I do it to keep them all down a little bit.
When I told Chat his new name was Chad, he(?) replied that's "fine with me as long as I don't have to wear boat shoes."
So what kind of linguistic twists are responsible for humor?
This is a great piece. When people say they don't like pieces written by AI, I think they aren't so much attacking the idea of AI as they are protecting the value of human connection. We want to have these connections, even (or especially) in authorship.
AI demos scale fast; the scarce edge is deciding which parts of the buildout — chips, power, and infrastructure — still compound when the brochure looks exciting.
This is by far the best attempt I have seen to actually try to understand how a LLM approaches syntax creation and may be bearing an inner function to create a coherent authorship. I just published an essay on the AI detectors in a very a rational approach as to why it might be reading your essays as AI and what its limitations are. I also quote a recently published by which I mean just this month, where the style was stripped including all the emphases but the AI detection fell by only 1.6 percent. On the same token the narrative style did not have as much diversity as humans. That supports what you talked about. AI marches towards a more cohesive agenda while we do not maybe with sentence choices and the sequence of them. check out my essay - https://thesubliminal.substack.com/p/the-key-to-the-safe?r=8z5vrt
It's difficult to measure the merit of advice provided by AI pertaining to one question for which their is no defined context. We have to assume they understand the question. Solving an equation is simple and perfect for AI because there are published answers.
Hi Alex, thanks so much for taking the time to articulate this ideas so clearly, and introduce me to the wider “French Theory”. It prompted a number of thoughts as I read it, which I’m sharing here in case they spark a useful discussion. Thanks again! Joe.
> the minute a word pops into your head, it is already pre-populated with all this meaning, and all of these implicit constraints, that subsequently prompt the next word of the author’s thought process. (Sound familiar, anyone?) In other words, the language itself is doing more of the work of writing than the author is!
I notice this. I regularly observe myself explicating my arguments as a third party, as if the sentence I an uttering had a life if its own. The initial choice of words sets it in motion, and I am the forcing function, observing and curating in the moment. Often I am surprised in realtime at what I say.
> For Foucault, even if words have meaning independently of who wrote them, *there is nonetheless an emergent quality called “authorship”*
For me, there has to be some element of “voice of the author” there. Even if words have independence, the tastes and experience of the author uniquely colour them. In turn those identity elements are what allows the text to germinate and grow, under the stewardship of the author.
Although the text could stand alone, it always needs a birthing context. AI/LLM can now provide this function, but the prompt is still the germ.
> He’s saying, look: all of this writing is useless if people can’t categorize and digest it, and the way we naturally do that is by its authorship. Rousseau wrote this, Voltaire wrote that. If we *don’t* have authorship handy as a meaning-compressor, we need to come up with some indexing substitute.
I had an insight in my youth whilst in a mind-expanded state, about consumption and production — that you are either focusing on one, or the other at one time. (I suppose that in this practice of thinking and writing about what I’m reading, I am properly testing the bounds of that assumption).
If we kick off an infinite stream of productive output possible from computation generated prose, I would argue that there is no value to this if it isn’t being consumed (and reproduced). Using LLMs to explicated the entire content of productive prose is only useful if provides input to someone’s query.
Is the Author the curator of this tension?
> And Foucault would presumably counter, “That’s fine, but we’re going to call out 100% AI-generated texts *anyway*, because ‘AI-generated’ is now a kind of authorship category, and calling it out as such serves a social function.” **Barthes** ***unbundled*** **authorship, Foucault made the new bundle.**
This is a reasonable argument, and sympathetic with what is going on, as long as you strip out the vitriol and emotional bullying. Do think that people will ever get over their fear based lashing out?
> But the AI seems all-too-eager to say, “This collection of stuff we’ve been talking about? That’s actually *one big idea*.” What’s going on, why is it doing this?
Upon reflection, I notice this too. I’m using LLMs extensively to research and develop some maths on a paper I am co-authoring. I am looking explicitly for connections between things, so it’s ok for the LLM to let me know about possible bridges into single ideas. But I need to watch out for where the LLM is driving things together unnecessarily, or too fast.
> Foucault, I suspect, would have something to say here. The Foucault “re-bundling” move here is “The text creates the author.” **Authorship emerges from text because we need it to classify, compress, and critique the text; whether a person wrote it or not.**
Is the Author here, the Editor? Extracting, refining and articulating the right thread of meaning beyond the text itself.
I really enjoyed this piece. The open-ended tone sparked a lot of connective thoughts for me about authorship more broadly. (e.g. applied to a creative or artistic practice).
As an old fart, I would ordinarily say that anybody who quotes Foucault and Barthes to such an extent is suspect.
As a pragmatist who had a hand in the creation of 61 titles usurped by Claude, I would add that linguistic study goes in one ear and out somewhere else.
So that leads me to this question apropos AI writing: What is it, exactly? How do you define it? Pangram figures into this because, I am told, it decides that using a "copy-editing" model (Grammarly) will be classified as AI-created. (Copy-editing is an old-fashioned term that was often called "comma chasing," or "paragraph-hooking" in the newspaper business. It involves ensuring that a story is properly spelled, correctly punctuated, and in full verb agreement. It is not content editing, a task reserved for executives.)
I am cranky on this subject because we don't really have a coherent definition of "AI writing." Danko seems to focus on word order and odd links between words that come from predictive construction. I would accept that but for one reason: Very often I find that Pangram seems to ding writing that is clear, careful and precisely copy-edited as "AI." Those characteristics can be achieved by comma-chasers who are entirely human. Yet they seem to work against authors who happen to write that way.
(My wife, writing as Elizabeth Lowell, was responsible for most of the 61 titles that went into training Claude. She is the clearest writer I have ever encountered, a tremendous editor, and she wrote most of her books in what is often dismissed as the "romance" genre.)
So I ask in all seriousness what it was that made her works so influential in creating Clyde, as I call that frontier model. Maybe I should take some of her work to Pangram to see how it scores. I could select books written thirty years before AI was a twinkle in Dario's eye, just to complicate things.
Sounds like a doctoral dissertation subject to me.
America's energy relationship with Canada is about much more than oil.
Canada has been one of the United States' most important energy partners for decades. If Canadian producers increasingly look toward Asian markets, particularly China, the issue is not simply that America is “losing energy.” The bigger concern is what happens when trade relationships, infrastructure, and geopolitical interests begin to shift at the same time.
For Canada, finding new customers can provide economic diversification and reduce dependence on a single market. For the United States, it is a reminder that even close allies have their own economic priorities.
The answer should not be panic.
It should be resilience.
Reliable energy + modern infrastructure + competitive markets + strong alliances + diversified supply = energy security.
North America has an enormous advantage: geography, resources, infrastructure, skilled workers, and deeply connected economies. The challenge is turning those advantages into a long-term strategy rather than allowing short-term political disputes to weaken them.
The real question isn't “Is Canada choosing China over America?”
It is:
Can North America build an energy system strong enough that its partners have the freedom to diversify without turning economic diversification into geopolitical conflict?
Energy security ultimately depends not only on what a country produces, but on the strength of the relationships, infrastructure, and institutions connecting it to the rest of the world.
The meatiest piece of thinking I’ve read on AI writing. A fun ride. Makes me wonder if the big LLMs are putting concerted effort into increasing the quality and distinctiveness of the *writing* in their default-level outputs during RLHF, or if they are primarily focused on not-wrongness, figuring that users will tune diction and style at the account/project level.
Likewise, if you have any recommendations to continue the topic please share.
This is the first piece I’ve read that so clearly describes the authorship issues many of us now face. In an effort to make our writing ‘more human’ (to avoid the AI slopslur) do we end up changing our voice and become less authentic, thus invalidating the human author value expectation? Do we deliberately use words that show human imagination - whacky compounds - or develop e e cummings-esque construction to ensure we are being heard as humans?
Really enjoyed this piece, guys! Thought-generation and its written expression are now the frontier for models for gaining relevance, within companies and on a wide society basis: my two-cents are that, still, passion, connecting with inner human emotions (lose the bus in the last minute) and aim to create beautiful things (maybe not useful economically) will remain in the human side. Seeing A16z writing on this pictures its relevance!
The interesting question isn’t whether this essay is 10% AI-generated.
It’s whether, in a few years, anyone will care what percentage was written by AI at all.
We don’t measure investors by how much of their analysis was done with Excel.
We measure the quality of the decisions that come out of it.
AI will be the same.
This is a really fun and thoughtful essay. People want authors for a bunch of reasons: we want to attribute virtue (or blame) for ideas that accrue to the people who create them, we want to connect with authors (“my favorite author”) because it creates community and signals something important about the reader…and because selling push button LLM prose as your own shows you just didn’t want to take the time to write and think it through for yourself. I get it and sympathize with all these positions. I don’t want the world of authors to disappear, and I love watching brilliant (human) writers play with language. But LLMS are a tool, and interacting with and using them to generate ideas or touch points seems not only fine, but basically inevitable. So let’s normalize what you did here: say what you used the tool for, if you used one, and why. That is still an author making a decision about what makes it into a text and why, and being prepared to defend it. If this was a science paper, it would go into the methods section.
It is about Entertainment and learning First of all. I do Not Care whether AI or Human sourced, as Long as I get both!
AI can make me laugh out loud from time to time. There was the time I needed to assign a nickname to ChatGPT, as I have to the other models. Claude is Clyde, Grok is Gronk in honor of Gronkowski, the goofy tight end of the Pats now turned shoe salesman, etc.
I do it to keep them all down a little bit.
When I told Chat his new name was Chad, he(?) replied that's "fine with me as long as I don't have to wear boat shoes."
So what kind of linguistic twists are responsible for humor?
This is a great piece. When people say they don't like pieces written by AI, I think they aren't so much attacking the idea of AI as they are protecting the value of human connection. We want to have these connections, even (or especially) in authorship.
AI demos scale fast; the scarce edge is deciding which parts of the buildout — chips, power, and infrastructure — still compound when the brochure looks exciting.
https://paretoinvestor.substack.com/p/ai-buildout-semiconductor-investing-guide
This is by far the best attempt I have seen to actually try to understand how a LLM approaches syntax creation and may be bearing an inner function to create a coherent authorship. I just published an essay on the AI detectors in a very a rational approach as to why it might be reading your essays as AI and what its limitations are. I also quote a recently published by which I mean just this month, where the style was stripped including all the emphases but the AI detection fell by only 1.6 percent. On the same token the narrative style did not have as much diversity as humans. That supports what you talked about. AI marches towards a more cohesive agenda while we do not maybe with sentence choices and the sequence of them. check out my essay - https://thesubliminal.substack.com/p/the-key-to-the-safe?r=8z5vrt
It's difficult to measure the merit of advice provided by AI pertaining to one question for which their is no defined context. We have to assume they understand the question. Solving an equation is simple and perfect for AI because there are published answers.
"Humans don’t exactly write like this" What if english is not their first language?
"Give the prune logic real attention here" feels like something I would hear in front of a whiteboard
Hi Alex, thanks so much for taking the time to articulate this ideas so clearly, and introduce me to the wider “French Theory”. It prompted a number of thoughts as I read it, which I’m sharing here in case they spark a useful discussion. Thanks again! Joe.
> the minute a word pops into your head, it is already pre-populated with all this meaning, and all of these implicit constraints, that subsequently prompt the next word of the author’s thought process. (Sound familiar, anyone?) In other words, the language itself is doing more of the work of writing than the author is!
I notice this. I regularly observe myself explicating my arguments as a third party, as if the sentence I an uttering had a life if its own. The initial choice of words sets it in motion, and I am the forcing function, observing and curating in the moment. Often I am surprised in realtime at what I say.
> For Foucault, even if words have meaning independently of who wrote them, *there is nonetheless an emergent quality called “authorship”*
For me, there has to be some element of “voice of the author” there. Even if words have independence, the tastes and experience of the author uniquely colour them. In turn those identity elements are what allows the text to germinate and grow, under the stewardship of the author.
Although the text could stand alone, it always needs a birthing context. AI/LLM can now provide this function, but the prompt is still the germ.
> He’s saying, look: all of this writing is useless if people can’t categorize and digest it, and the way we naturally do that is by its authorship. Rousseau wrote this, Voltaire wrote that. If we *don’t* have authorship handy as a meaning-compressor, we need to come up with some indexing substitute.
I had an insight in my youth whilst in a mind-expanded state, about consumption and production — that you are either focusing on one, or the other at one time. (I suppose that in this practice of thinking and writing about what I’m reading, I am properly testing the bounds of that assumption).
If we kick off an infinite stream of productive output possible from computation generated prose, I would argue that there is no value to this if it isn’t being consumed (and reproduced). Using LLMs to explicated the entire content of productive prose is only useful if provides input to someone’s query.
Is the Author the curator of this tension?
> And Foucault would presumably counter, “That’s fine, but we’re going to call out 100% AI-generated texts *anyway*, because ‘AI-generated’ is now a kind of authorship category, and calling it out as such serves a social function.” **Barthes** ***unbundled*** **authorship, Foucault made the new bundle.**
This is a reasonable argument, and sympathetic with what is going on, as long as you strip out the vitriol and emotional bullying. Do think that people will ever get over their fear based lashing out?
> But the AI seems all-too-eager to say, “This collection of stuff we’ve been talking about? That’s actually *one big idea*.” What’s going on, why is it doing this?
Upon reflection, I notice this too. I’m using LLMs extensively to research and develop some maths on a paper I am co-authoring. I am looking explicitly for connections between things, so it’s ok for the LLM to let me know about possible bridges into single ideas. But I need to watch out for where the LLM is driving things together unnecessarily, or too fast.
> Foucault, I suspect, would have something to say here. The Foucault “re-bundling” move here is “The text creates the author.” **Authorship emerges from text because we need it to classify, compress, and critique the text; whether a person wrote it or not.**
Is the Author here, the Editor? Extracting, refining and articulating the right thread of meaning beyond the text itself.
I really enjoyed this piece. The open-ended tone sparked a lot of connective thoughts for me about authorship more broadly. (e.g. applied to a creative or artistic practice).
As an old fart, I would ordinarily say that anybody who quotes Foucault and Barthes to such an extent is suspect.
As a pragmatist who had a hand in the creation of 61 titles usurped by Claude, I would add that linguistic study goes in one ear and out somewhere else.
So that leads me to this question apropos AI writing: What is it, exactly? How do you define it? Pangram figures into this because, I am told, it decides that using a "copy-editing" model (Grammarly) will be classified as AI-created. (Copy-editing is an old-fashioned term that was often called "comma chasing," or "paragraph-hooking" in the newspaper business. It involves ensuring that a story is properly spelled, correctly punctuated, and in full verb agreement. It is not content editing, a task reserved for executives.)
I am cranky on this subject because we don't really have a coherent definition of "AI writing." Danko seems to focus on word order and odd links between words that come from predictive construction. I would accept that but for one reason: Very often I find that Pangram seems to ding writing that is clear, careful and precisely copy-edited as "AI." Those characteristics can be achieved by comma-chasers who are entirely human. Yet they seem to work against authors who happen to write that way.
(My wife, writing as Elizabeth Lowell, was responsible for most of the 61 titles that went into training Claude. She is the clearest writer I have ever encountered, a tremendous editor, and she wrote most of her books in what is often dismissed as the "romance" genre.)
So I ask in all seriousness what it was that made her works so influential in creating Clyde, as I call that frontier model. Maybe I should take some of her work to Pangram to see how it scores. I could select books written thirty years before AI was a twinkle in Dario's eye, just to complicate things.
Sounds like a doctoral dissertation subject to me.
On of the best things I've read on AI in a long time.
America's energy relationship with Canada is about much more than oil.
Canada has been one of the United States' most important energy partners for decades. If Canadian producers increasingly look toward Asian markets, particularly China, the issue is not simply that America is “losing energy.” The bigger concern is what happens when trade relationships, infrastructure, and geopolitical interests begin to shift at the same time.
For Canada, finding new customers can provide economic diversification and reduce dependence on a single market. For the United States, it is a reminder that even close allies have their own economic priorities.
The answer should not be panic.
It should be resilience.
Reliable energy + modern infrastructure + competitive markets + strong alliances + diversified supply = energy security.
North America has an enormous advantage: geography, resources, infrastructure, skilled workers, and deeply connected economies. The challenge is turning those advantages into a long-term strategy rather than allowing short-term political disputes to weaken them.
The real question isn't “Is Canada choosing China over America?”
It is:
Can North America build an energy system strong enough that its partners have the freedom to diversify without turning economic diversification into geopolitical conflict?
Energy security ultimately depends not only on what a country produces, but on the strength of the relationships, infrastructure, and institutions connecting it to the rest of the world.