Have you thought about how AI interfaces with this? The LLMs are pretty good at reading now. They're not *amazing* at reading carefully (https://www.lesswrong.com/posts/s58hDHX2GkFDbpGKD/linch-s-shortform?commentId=i62nwrwLhSxpe2vCo), but imo they're very good at reading things quickly and giving you a sense of the important bits for stuff like legal briefs and earnings reports.
Might respond to this with a longer post later, but shooting from the hip here:
There's a lot already said about "known unknowns" vs "unknown unknowns," but I think that there's probably another level to it. At level 1 (known unknowns), you have questions like "What's the GDP of Baltimore?" At level 2, (unknown unknowns), you have questions like "what are some surprising and counter-intuitive things about Baltimore's economy, the things that most people wouldn't think to ask?" And then, at level 3 (unknown^3), you have all of the information that you wouldn't get from the level 2 query but would still like to know.
Because the world is so high-fidelity, most of the world's knowledge exists at level 3, which is basically beyond what you can capture from an LLM query. Like, the way you find underpriced stocks is not by going to an LLM and saying "what stocks are underpriced right now?" That's where the value of Doing The Reading comes in.
I think that several of the "media audit" episodes I described during my time with Theorist actually do a pretty good job of illustrating some of the things that current LLMs are bad at.
For example, take the Food Theory Crystal Pepsi episode. I was basically starting with the same level of sophistication as most people whose knowledge just came from reading the viral Reddit headline:
"TIL Coke countered Crystal Pepsi with Tab Clear. The idea was that Tab was perceived as an inferior, boring diet type drink, and equating it as competition with Crystal Pepsi would damage its image. Coke succeeded and both products were dead within five months"
I went in and read Sergio Zyman's version of the story expecting to write an episode *praising* this guy for his brilliant product strategy, just like all of the other credulous writers, and it was only upon reading his own words that I realized "this timeline makes no sense, this reeks of post-hoc confabulation from a guy who is already infamous for his other failed product launches, the viral tweets and reddit posts and bloggers just took this guy at his word!"
So let's imagine a counter-factual scenario, where instead of "doing the reading," I had just gone to ChatGPT and said "this seems like an interesting story, tell me more about it!" or even a more skeptical query like "Did this actually happen?" And the ChatGPT reasoning model will dutifully research the internet, find all of the low-effort articles that are just credulously repeating the thing that one Coca-Cola exec said, and references to books that repeated the narrative. It will also add some context, it will give you firm numbers and maybe even hedge a little and say "well Zyman might be overstating his case because Crystal Pepsi probably would have declined even absent Tab Clear," but it will not uncover all of the inconsistencies in Zyman's story.
Another good example of someone who Did The Reading and discovered the sort of things that current LLMs would miss is hbomberguy's video essay "ROBLOX_OOF.mp3" where he basically discovered that a guy named Tommy Tallarico had spent an entire career lying and then getting those lies repeated by credible-sounding sources that naive researchers were happy to parrot. (ChatGPT will tell you the truth about Tommy Tallarico *now*, because hbomberguy's essay inspired a lot of "expose" posts that have made his discoveries more legible to the LLM, but someone had to be the first to look closely at the official narrative and say "it's hard for me to reconcile several 'facts' that seem incompatible with each other."
I will also say that in the "unknown" abstraction ladder, oftentimes what you do by Doing The Reading is descending the abstraction ladder.
Like, if you ask ChatGPT "what's an investment I should make if I want to make a lot of money," there's probably no alpha there, because what ChatGPT is going to give you is summary of the current public narrative. But an investor with a specific thesis could say, "Okay, I want to use the SEC Company Facts API to pull the stats for every publicly-traded company in a given sector, calculate their gross cap ratio and net cash ratio, and then look for interesting businesses where the market is valuing the operating business low after accounting for cash and debt." Or a less-informed person could make this same query and not know which qualifiers to include and learn things like, "huh, it seems like all of the 'cheapest' companies by this metric are SPACs, or firms with restricted cash," and thereby learn something new.
(This also counts as Doing The Reading. It leads you to interesting discoveries, like "wait, how is it possible that this therapeutics company has $500mm on their balance sheet, but they're only trading at $300mm, how is that even possible? It seems like I'm going into a store and spending $300 on a piggy bank with $500 cash inside of it, what's the catch? And this cellar therapy company is also trading below cash, and so is this other one, and wait, they're all biotech companies..." And then you've learned something new about the world, and have the ability to ask better and more precise questions.)
A version that comes up in a lot of my conversations with young job-seekers: Read LinkedIn. If you want to get a job, look at what the people who had that job were doing before they got it. Then, if you can, do that.
You're 100% right. A moist critical video doesn't take a week to edit. No doubt at least 2.
So much fruit, hanging so low.
Fantastic post!
Came for the Costco "secret" and stayed for the reading.
Have you thought about how AI interfaces with this? The LLMs are pretty good at reading now. They're not *amazing* at reading carefully (https://www.lesswrong.com/posts/s58hDHX2GkFDbpGKD/linch-s-shortform?commentId=i62nwrwLhSxpe2vCo), but imo they're very good at reading things quickly and giving you a sense of the important bits for stuff like legal briefs and earnings reports.
Might respond to this with a longer post later, but shooting from the hip here:
There's a lot already said about "known unknowns" vs "unknown unknowns," but I think that there's probably another level to it. At level 1 (known unknowns), you have questions like "What's the GDP of Baltimore?" At level 2, (unknown unknowns), you have questions like "what are some surprising and counter-intuitive things about Baltimore's economy, the things that most people wouldn't think to ask?" And then, at level 3 (unknown^3), you have all of the information that you wouldn't get from the level 2 query but would still like to know.
Because the world is so high-fidelity, most of the world's knowledge exists at level 3, which is basically beyond what you can capture from an LLM query. Like, the way you find underpriced stocks is not by going to an LLM and saying "what stocks are underpriced right now?" That's where the value of Doing The Reading comes in.
I think that several of the "media audit" episodes I described during my time with Theorist actually do a pretty good job of illustrating some of the things that current LLMs are bad at.
For example, take the Food Theory Crystal Pepsi episode. I was basically starting with the same level of sophistication as most people whose knowledge just came from reading the viral Reddit headline:
"TIL Coke countered Crystal Pepsi with Tab Clear. The idea was that Tab was perceived as an inferior, boring diet type drink, and equating it as competition with Crystal Pepsi would damage its image. Coke succeeded and both products were dead within five months"
I went in and read Sergio Zyman's version of the story expecting to write an episode *praising* this guy for his brilliant product strategy, just like all of the other credulous writers, and it was only upon reading his own words that I realized "this timeline makes no sense, this reeks of post-hoc confabulation from a guy who is already infamous for his other failed product launches, the viral tweets and reddit posts and bloggers just took this guy at his word!"
So let's imagine a counter-factual scenario, where instead of "doing the reading," I had just gone to ChatGPT and said "this seems like an interesting story, tell me more about it!" or even a more skeptical query like "Did this actually happen?" And the ChatGPT reasoning model will dutifully research the internet, find all of the low-effort articles that are just credulously repeating the thing that one Coca-Cola exec said, and references to books that repeated the narrative. It will also add some context, it will give you firm numbers and maybe even hedge a little and say "well Zyman might be overstating his case because Crystal Pepsi probably would have declined even absent Tab Clear," but it will not uncover all of the inconsistencies in Zyman's story.
Another good example of someone who Did The Reading and discovered the sort of things that current LLMs would miss is hbomberguy's video essay "ROBLOX_OOF.mp3" where he basically discovered that a guy named Tommy Tallarico had spent an entire career lying and then getting those lies repeated by credible-sounding sources that naive researchers were happy to parrot. (ChatGPT will tell you the truth about Tommy Tallarico *now*, because hbomberguy's essay inspired a lot of "expose" posts that have made his discoveries more legible to the LLM, but someone had to be the first to look closely at the official narrative and say "it's hard for me to reconcile several 'facts' that seem incompatible with each other."
I will also say that in the "unknown" abstraction ladder, oftentimes what you do by Doing The Reading is descending the abstraction ladder.
Like, if you ask ChatGPT "what's an investment I should make if I want to make a lot of money," there's probably no alpha there, because what ChatGPT is going to give you is summary of the current public narrative. But an investor with a specific thesis could say, "Okay, I want to use the SEC Company Facts API to pull the stats for every publicly-traded company in a given sector, calculate their gross cap ratio and net cash ratio, and then look for interesting businesses where the market is valuing the operating business low after accounting for cash and debt." Or a less-informed person could make this same query and not know which qualifiers to include and learn things like, "huh, it seems like all of the 'cheapest' companies by this metric are SPACs, or firms with restricted cash," and thereby learn something new.
(This also counts as Doing The Reading. It leads you to interesting discoveries, like "wait, how is it possible that this therapeutics company has $500mm on their balance sheet, but they're only trading at $300mm, how is that even possible? It seems like I'm going into a store and spending $300 on a piggy bank with $500 cash inside of it, what's the catch? And this cellar therapy company is also trading below cash, and so is this other one, and wait, they're all biotech companies..." And then you've learned something new about the world, and have the ability to ask better and more precise questions.)
A version that comes up in a lot of my conversations with young job-seekers: Read LinkedIn. If you want to get a job, look at what the people who had that job were doing before they got it. Then, if you can, do that.