Totally agree. A benefit to a contrarian is, the greater the consensus the greater the edge. LLMs are like only being allowed to view the first result on Google, which puts consensus on steroids. There's going to be some serious alpha in places where LLMs give back consensus opinions but miss profitable (or beautiful) edges.
It made me think of LLM's as an instrument and the music we make with them as another analogy with similar meanings.
We are all playing the same instrument (more or less), we bring our context to tuning the ideas it produces, but the instrument only has so many notes. Not surprising we play similar tunes.
I like this because music demonstrates the rich generative potential of combinatorial composition, especially when in harmony with collaborators.
It also tells us it would be great if we could design some alternative instruments, attuned to and trained on different frequencies and bodies of knowledge.
ooh I like the instrument analogy, that feels more productive/less depressing than the one I went with. LLM ideas as chords and harmonies + some X factor brought by the user
The source for the Fuller is from a biography "Buckminster Fuller's Universe" and it is a paraphrase of his ideas rather than a quote, but this is my memory speaking 25 years after reading the book, so I could very possibly be fabricating the whole thing!
> So in an era where both distribution and ideation are commoditized, the valuable complement becomes establishing provenance.
This may very well be true, but I'm unsure (at the moment) that ideation is actually commoditized. It depends on how much LLMs saturate the space of potential worthwhile ideas. If it's a small percentage, then the ideas that aren't in LLM-space are still not commoditized (and may paradoxically rise in value because execution is also _somewhat_ commoditized by LLMs).
There are so many people in this world but majority of the internet is American with some European. What are the rest of the world thinking? I don’t think they are arriving at the Oklahoma example. Would a researcher from Macedonia have similar ideas to one in Nigeria? I think LLMS pre-commodify ideas only for markets where ideas were commodities in the first place. So for mature markets recursion has run itself into a less generative state but for everyone else it might only be the beginning.
Distribution has two parts. One is getting the thing out there. That's, as you say, solved. The other part is getting the thing to its intended recipients. That is still very, very expensive. Attention still costs time or money, usually both. You pointed this out yourself when you said advertising isn't going away.
Loved this piece and put into words something that I'd been noticing as well. Funny enough, I also got pushed toward stigmergy by the LLMs on a project I was working on a few months ago and it's been wild seeing it all over the place the last few weeks.
I think there's also something "sooner" adjacent (or between sooner and boomer?) that involves discovering new models' capabilities first and capitalizing on them in the short period before they're commodified. For example, if you were one of the first people publishing LLM-writing with all the em-dashes and "it's not X it's Y"'s, you can gain a bit until it's everywhere and becomes an obvious tell.
re: taking advantage of capabilities first, reminds me of people posting about new things they did with a new model but it usually seems to last for three days on twitter and seems to have a spectacle like vibe to it
Oh yeah, there's definitely the spectacle side of things. But I think something like discovering Opus 4.6's long form writing capabilities and starting to publish with it early is in a bit of a different category.
I had a post do pretty well on Hacker News back in February, and the comments section had a ton of people who didn't realize it was written with AI (https://news.ycombinator.com/item?id=47431237). There were a few who spotted it, but I think it was early enough from the release of the model that most people hadn't run into 4.6's tics as much yet.
for sure, I came to know of the idea through Venkat + James Scott I think was primarily more popular in a leftist political setting prior to Venkat’s piece
Alpha for being first feels important for understanding the shift from the pre-printing press provenance first culture to the modern publicity-first approach which I have been thinking about since reading Ibn Khaldun: An intellectual biography last year.
(There is plenty more to say, but like most of your pieces, the ideas this sparked are bigger than the time I have to work things out.)
While a minor part of the book, Ibn Khaldun’s account of his teachers and his teachers’ teachers establishes the provenance of his training in a way that feels unavailable to us now.
After the printing press, phrases like “standing on the shoulders of giants” persist as acknowledgement, but they can no longer trace the lineage of ideas back through living transmission.
Learning from books rather than teachers obviously changes what an individual inheritor carries forward — but it may also relate to the density of ruins relative to new construction.
The pre-printing press world was saturated with the physical remains of Roman and Egyptian civilization, making it easy to feel that the original work was superior to its newest instantiation, that the present was a diminished copy of something greater.
Is the return of provenance a symptom of a digital world in decline? A world where the slop invasion has destroyed our ability to fight bit rot, leaving ruins made from our once-mighty online edifices?
Interesting ideas, but I am not convinced by your conclusion. There is competition in the llm world, unbiased data makes better llms, better llms are smarter, smarter llms can tell when data is biased
Buckminster Fuller also observed in the 1960's about how ideas "whose time [had] come" would seem to be independently discovered within a very short span of time. The implicit question I'm left with after your essay is: So what now? Is the next frontier of model training some combination of architecture shift and training data curation? Can this be accomplished with fine-tuning alone?
Dont really know the answer to those questions but it does seem like there is an artisanal LLM culture thats slowly developing so smaller, local LLMs which are trained on different context are interesting to watch. Where did Buckminster Fuller observe this? Any additional context?
Amazing! I like to think of this as "everyone knows the idea but nobody knows the others know it" haha. ESPECIALLY if you work with LLMs and don't understand their patterns / how it works.
But the idea being common is one thing - the common knowledge never forming is even worse, because you get dozens of people each privately convinced they have the "key" to something, each building on the assumption of a head start that doesn't exist...
Feels like an updated version of “If you read what everyone else is reading, you will think what everyone else is thinking” (Murakami)
Another way to fight it is to read widely based on your own interests.
Totally agree. A benefit to a contrarian is, the greater the consensus the greater the edge. LLMs are like only being allowed to view the first result on Google, which puts consensus on steroids. There's going to be some serious alpha in places where LLMs give back consensus opinions but miss profitable (or beautiful) edges.
Nice synthesis!
It made me think of LLM's as an instrument and the music we make with them as another analogy with similar meanings.
We are all playing the same instrument (more or less), we bring our context to tuning the ideas it produces, but the instrument only has so many notes. Not surprising we play similar tunes.
I like this because music demonstrates the rich generative potential of combinatorial composition, especially when in harmony with collaborators.
It also tells us it would be great if we could design some alternative instruments, attuned to and trained on different frequencies and bodies of knowledge.
Great fact about Oklahoma lol!
ooh I like the instrument analogy, that feels more productive/less depressing than the one I went with. LLM ideas as chords and harmonies + some X factor brought by the user
Haha totally!
I also like that even to the same music we can all dance to our own rhythm. Some of us are better at it than others
The source for the Fuller is from a biography "Buckminster Fuller's Universe" and it is a paraphrase of his ideas rather than a quote, but this is my memory speaking 25 years after reading the book, so I could very possibly be fabricating the whole thing!
> So in an era where both distribution and ideation are commoditized, the valuable complement becomes establishing provenance.
This may very well be true, but I'm unsure (at the moment) that ideation is actually commoditized. It depends on how much LLMs saturate the space of potential worthwhile ideas. If it's a small percentage, then the ideas that aren't in LLM-space are still not commoditized (and may paradoxically rise in value because execution is also _somewhat_ commoditized by LLMs).
There are so many people in this world but majority of the internet is American with some European. What are the rest of the world thinking? I don’t think they are arriving at the Oklahoma example. Would a researcher from Macedonia have similar ideas to one in Nigeria? I think LLMS pre-commodify ideas only for markets where ideas were commodities in the first place. So for mature markets recursion has run itself into a less generative state but for everyone else it might only be the beginning.
Distribution has two parts. One is getting the thing out there. That's, as you say, solved. The other part is getting the thing to its intended recipients. That is still very, very expensive. Attention still costs time or money, usually both. You pointed this out yourself when you said advertising isn't going away.
glad to see you getting your flowers for your smart essays my dude
Ah thank you so much man!
Brilliant 👌
Thank you Bhuvan
Loved this piece and put into words something that I'd been noticing as well. Funny enough, I also got pushed toward stigmergy by the LLMs on a project I was working on a few months ago and it's been wild seeing it all over the place the last few weeks.
I think there's also something "sooner" adjacent (or between sooner and boomer?) that involves discovering new models' capabilities first and capitalizing on them in the short period before they're commodified. For example, if you were one of the first people publishing LLM-writing with all the em-dashes and "it's not X it's Y"'s, you can gain a bit until it's everywhere and becomes an obvious tell.
re: taking advantage of capabilities first, reminds me of people posting about new things they did with a new model but it usually seems to last for three days on twitter and seems to have a spectacle like vibe to it
Oh yeah, there's definitely the spectacle side of things. But I think something like discovering Opus 4.6's long form writing capabilities and starting to publish with it early is in a bit of a different category.
I had a post do pretty well on Hacker News back in February, and the comments section had a ton of people who didn't realize it was written with AI (https://news.ycombinator.com/item?id=47431237). There were a few who spotted it, but I think it was early enough from the release of the model that most people hadn't run into 4.6's tics as much yet.
No disrespect to Venkatesh though James c Scott deserves the og claim to legibility no?
for sure, I came to know of the idea through Venkat + James Scott I think was primarily more popular in a leftist political setting prior to Venkat’s piece
Alpha for being first feels important for understanding the shift from the pre-printing press provenance first culture to the modern publicity-first approach which I have been thinking about since reading Ibn Khaldun: An intellectual biography last year.
(There is plenty more to say, but like most of your pieces, the ideas this sparked are bigger than the time I have to work things out.)
whats the Ibn Khaldun connection? I have not read that book
While a minor part of the book, Ibn Khaldun’s account of his teachers and his teachers’ teachers establishes the provenance of his training in a way that feels unavailable to us now.
After the printing press, phrases like “standing on the shoulders of giants” persist as acknowledgement, but they can no longer trace the lineage of ideas back through living transmission.
Learning from books rather than teachers obviously changes what an individual inheritor carries forward — but it may also relate to the density of ruins relative to new construction.
The pre-printing press world was saturated with the physical remains of Roman and Egyptian civilization, making it easy to feel that the original work was superior to its newest instantiation, that the present was a diminished copy of something greater.
Is the return of provenance a symptom of a digital world in decline? A world where the slop invasion has destroyed our ability to fight bit rot, leaving ruins made from our once-mighty online edifices?
Interesting ideas, but I am not convinced by your conclusion. There is competition in the llm world, unbiased data makes better llms, better llms are smarter, smarter llms can tell when data is biased
The edge is moving from ideation to judgment.
LLMs can surface the same analogy, same frame, same “contrarian” take for thousands of people at once.
So the scarce layer is no longer the first draft of the idea.
It is taste, provenance, synthesis, and the ability to know which “insight” is just latent-space consensus wearing a fake mustache.
Buckminster Fuller also observed in the 1960's about how ideas "whose time [had] come" would seem to be independently discovered within a very short span of time. The implicit question I'm left with after your essay is: So what now? Is the next frontier of model training some combination of architecture shift and training data curation? Can this be accomplished with fine-tuning alone?
Dont really know the answer to those questions but it does seem like there is an artisanal LLM culture thats slowly developing so smaller, local LLMs which are trained on different context are interesting to watch. Where did Buckminster Fuller observe this? Any additional context?
Amazing! I like to think of this as "everyone knows the idea but nobody knows the others know it" haha. ESPECIALLY if you work with LLMs and don't understand their patterns / how it works.
But the idea being common is one thing - the common knowledge never forming is even worse, because you get dozens of people each privately convinced they have the "key" to something, each building on the assumption of a head start that doesn't exist...
Amazing write up