Golden Nuggets Podcast #14

What's up, everybody? It's another Golden Nuggets podcast about here with Zander. We talk about formulation today. Beautiful. Love that, dude. You know, that's a skill. I wonder how long it takes to get good at that.

What is it, like an auctioneer type of voice? Yeah, auctioneer or like a boxing announcer or something like that. Yeah, well, boxing, you're enunciating way more, but you're obviously putting some... putting some sauce in the pronunciation to make it fun for people but anyway all right all right so yeah basically yeah you describe it go ahead wait describe what no did you have something in mind what we're gonna do

No, no, no. Describe what we're going to do. Oh, we're doing the pod, dude. No, no. I just, I just mean like what we intend to do with the formulation review. I thought we were doing the pod, you know? Oh my God. This man pranked me, you guys. This man pranked me. Hang on. We've broken the fourth wall. We've explained how we do, how we make these now.

oh man oh okay all right well i i was totally confused about it all right well now we're yeah let's just no no no now we're in the nether zone we're between two different types of content so we really have no idea what's gonna happen it's like schrodinger's box or whatever We're in the DMZ now. Exactly. We're in the DMZ of learning. Fantastic. We're waiting for the wave function to collapse and pick one of the multiverses to sort of branch off into right now.

Yeah, I'll pretend I understand that reference. All right. Well, OK, let's go for a Golden Nuggets podcast. I've got some topics laid out for me for us right now. All right. All right. Been thinking a lot about machine learning generated. flashcards uh give me your immediate associations with what uh they're not good okay there you go okay well i've been thinking recently that they might actually have uh some utility my friend and uh

Here's the argument that I'm going to lay out for you, okay? Generating machine learning flashcards is very low cost because the machine literally does it for you, right? So what I'm imagining is you open your SuperMemo article.

You pick a paragraph that you want cards to be generated from. The machine learning algorithm generates them for you. So you literally don't put in any cognitive effort. And then you have the choice. You have the optionality. You have... the choice whether to keep or discard the items that it creates so it's not like you're being it's not like you're downloading a shared deck

all of these cards are given to you and perhaps you're also studying from Gazam, so you can't decide which ones you want to keep and which ones you want to discard. It's not like that. You actually have the choice which of those cards that it generates you want to keep. So you can only keep those ones that meet your threshold of...

quality or value or whatever and so that's what i'm thinking with the machine learning generator flashcards it seems like there would be very little downside because of the low cost of generating them and the ability to choose only those high value ones

Yeah, so I definitely see that side of it. And just to clarify my position, not all machine learning generated flashcards are bad. It's just that most of them are. And the primary reason... that most of them are bad is because, to me, the most interesting items that one could make are not representing information that is explicitly stated in the text or in the source material.

If it isn't in the source material, the machines just are not at a point where, first of all, it doesn't have all of the information in your brain. So even if it could make associations between... linguistically different but conceptually related concepts, it still would not be good because it doesn't have all your information.

all the information you have in your brain right so that's one component and another component is that it can't do that anyway so even if it did it doesn't matter and that lack of of really more high level items that are not just directly pulling from the source. That completely removes most of the utility of it for me. Now, of course, what you're stating there is like, oh, for the simpler stuff, we can just save time by doing this. And yeah, that's definitely true.

A couple of my concerns about it are, one, that it makes you kind of lazy in terms of the formulation. And it's sort of like... an atrophy of a muscle you know like you're sitting still for too long or you're not exercising enough or whatever it is and obviously it's the extreme case but if you don't try it's

It's just you lose that muscle. You lose it. Use it or you lose it. You know what I mean? So I think it's partially that. And I also think. Yeah, I don't know. It just it messes up your evaluation of things. And also there's benefit to writing. Items yourself even if they're sort of tedious like it's not clear to me

that if you use a bunch of machine learning generated items for the easy stuff that you would remember them just as well as if you had written them yourself. So those are my concerns with it primarily. And I think we are very far away. very far away from any sort of serious utility of this stuff at all. And trust me, I've seen as many of this type of thing with as many different approaches as people have to offer. As far as I'm aware,

I've seen all of the current instantiations of it, even the ones that are very unpopular, meaning not very known yet. So anyway, as far as I've seen, we're not even close. That's interesting. Yeah, like you... when you were talking about what kinds of items you find really valuable you talked about those ones that aren't that make connections that aren't

obviously present in the source text they sort of make a connection with something that's already stored in your prior knowledge right and so yeah there's this

interesting article by was when he's talking about formulation for programming items and he's talking about the different ways that you can decompose the value of an item into so for example firstly you have like the reference time how long does it take you to look up something on the internet and then by learning that item you sort of gain

that amount of time so that's one way you can uh sort of increase the value by studying flashcards is just by decreasing the amount of time you need to spend looking up things right and then i think what you're talking about um uh there's a much more difficult um measure of item value which is like the association uh bonus from learning items and that might be the one where it's like

where you actually gain the most value potentially, because obviously if you're making a good association, it could result in a business or new scientific breakthrough or something like that. And it's also the one that's most.

difficult to quantify as we've talked about before with the whole is it going to be a copper nugget or a golden nugget thing so right um yeah so yeah i thought that was interesting but one of the things that i um that uh that i did realize recently is that uh it is possible for example to train these machine learning models on your prior knowledge to a certain extent because you could provide machine learning model with the database of all of the prompts that you'd made before and

I don't know enough about machine learning to say whether that would be a good approach or whether that would work, especially in the current situation. I don't know whether the technology is there, whether it's at the level where...

it would be able to actually create valuable connections between sort of new information and your prior knowledge. But yeah, I guess there's potential there. I think I agree with you that... yeah it's probably they aren't probably going to be generating the most valuable items possible and you will always need to rely on your own item making ability but i still think it's interesting and yeah it's definitely something i want to experiment with in the future

I'm actually working on a little something to test it out which I'll be releasing very very soon so look out for that. Oh, that's exciting. Yeah, I mean, do you honestly think there's any utility in training it on your prior knowledge? I have not seen anything very good from that. Well, there's this project called Duo.

is being worked on by a guy called paul brickman mainly he's working on it i think it's like a collective um uh project but he's been doing basically all of the work so far and he's created this program called dual which I think it's an Obsidian plugin right now and it's basically GPT-2.

which is one of the natural language processing models. You might have heard of GPT-3, which is like the latest instantiation of it. But that one requires a huge amount of processing power in order to run. A ton. It's unbelievable.

So GPT-2 is like its little brother, but it's still pretty good. It does a really good job at generating text, basically. And what he's done is... uh you basically have gpt2 which is very good at generating text and then he he trains the model also on all of your obsidian notes so all of your writing what you've been writing about and then it creates this sort of chat bot

thing which he calls your dual and then you can type messages to your dual like natural language queries like uh tell me something about um i don't know dna or something like that and then the the chat bot basically because it has, it's sort of trained on all of your prior knowledge and all of your writing, it will give you a response. It will generate you some text that, you know, it contains.

it will contain references to all of the things that you've been writing about. So it does sort of have an understanding of your prior knowledge in that sense. And unfortunately I haven't had enough time to try it out yet, but it does look pretty cool. I was really surprised that something like that is actually possible, but it does seem to be becoming more and more feasible gradually. Yeah, it's...

The thing is, I think with the current instantiations, it's really cool. And I've said as much to Paul. And I really honestly believe that it's really cool. But it's just that as of right now, the results I've seen are not better. then suppose you're using like a Zettelkasten system where your links are, or your notes are linked, right? If you say, okay, I want to learn about DNA to give that same example.

then you just go to the dna tag or you search your notes for dna and then you just sort of traverse a few links and read a little bit and i think you'll end up in a better spot than whatever the output of the machine is so to me the question isn't can the machine generate anything useful it's is the content generated from the machine significantly better

or at least as good in a way that could save you time than what you could find by just quickly looking... Like, right now, I know that if I had interlinked notes about DNA, it would take me 30 or 45 seconds to... find my notes about that and activate the associations and rethink the same things that would bring me more value than something that GPT-2 could

provide me yeah that's just as of right now that's just a fact and that's not to nullify the work it's just to say it's fun but not useful not yet yeah what i really want to see is uh um someone do something like that sort of really depends on this ability if you know what i mean so is someone using this to actually generate really interesting ideas can they show that they use this thing to do that i think that's yeah it's a fair point um one of the things that i

don't necessarily like but you can also sort of see it as a benefit is that what GPT-2 is really good at is generating plausible text it's not doing human level reasoning and giving you a response that's based on uh like really deeply thinking about the query that you sent it so if you ask about dna it isn't giving you it isn't thinking if you know in the same sense that a human would it's it's generating

things that would be plausible in a dna context if you know what i mean and then stringing all of that together into english text and so that's i can see the benefit because what paul has told me about this is that it isn't necessarily supposed to be giving you strictly correct accurate responses it's mainly to be as like a sort of creativity booster it's not necessarily going to give you those crisp scientific academic responses

But it's going to give you something that might trigger something in your brain, might cause you to create a connection between something that you didn't expect or something like that. But the connection... Excuse me. The connection seems like it could have been made, like I say, by just manually reading. I understand that it's cool to ask a computer, but are we doing this because it's cool or are we doing it because it's useful? Right, right, right.

Yeah, it's always got to compete with the human and also all of these just more primitive search technologies, as you mentioned before as well. Right. And as of right now, there.

If you're talking particularly in the context of a zettelkast note system, almost all of the value of this sort of semantic searching or semantic summarization or whatever it is that this... GPT-2 would allow is nullified because you can already find that with the zettelcast and that's one of the primary benefits a conversation with the notes that's how Luhmann described it and so I just

While I think it's very, very, very cool and I support this work and I hope it continues, it's just not at a point where it's very good. Where I think it might be useful is if...

you're using it on text that you didn't write. And so it doesn't have a bunch of links. Like I tried to do it with the incremental reading guide. I tried to give GPT-3 because I have access to that, the whole incremental reading guide. Now, of course there's limits. So I had to... do sections, but I tried to give it that and then ask it questions about incremental reading based on the guide that weren't explicitly answered in there, but they were implicitly.

And I actually got some pretty good results. And so the utility of that is that the incremental reading guide is not densely linked. You didn't write it. Odds are you don't remember very much of it word for word unless you...

attempted to do that so there's value in that kind of thing and i imagine if you could import a whole book or something and then ask questions of the book that would be useful as well what if i could ask questions of your uh jewel what if i could uh what if you could sell access to your sort of api endpoint for your um zettelcasting system and i could uh you know ask you some questions or interview you through via your uh ai friend um

Yeah, I don't know. It just seems likely to be mostly nonsense. One, because the technology is not so good. And two, because my personal notes are not... that coherent if you're not me right if you don't have my prior knowledge yeah i don't know yeah that makes sense all right well do you see any potential um are there any technologies out there that you could think

that you think could be sort of meaningfully combined with what we already have to create sort of big improvements. Yeah. Yeah, it's hard to say because... You need a high hit rate. That's the problem. Even if you could find a low-level application of it, you need a really high hit rate in order for it to be... And by hit rate, I just mean success rate. I'm probably using the wrong term. But I just mean...

The percentage of the time when you... I don't know why I'm coughing, dude. The percentage of the time when you ask the machine for something... How often does it get it right? That's what I'm concerned with. And if it's really high, even if the application is pretty simplistic, that's valuable. But it's hard to find a domain where...

It's high and of any utility at all. But somewhere it's good are like reading PDFs, for example. Amazon has... an api for this machine learning api that's very good and that kind of thing so those applications i think are good but for more complex stuff probably not yeah yeah it's a shame yeah i wish there was um i don't know i'm just always looking out for uh technologies that i feel like they could be combined with what we already have to

I don't know, create some big changes or something. I feel like one of the things that's really missing is perhaps social collaboration. I guess that's something that Rome's looking into. You're more of a Roman. than i am in fact i recently converted roman uh so ah rom is the best have you tried out the uh the social collaboration side of it do you think it's valuable um

As of right now, it's not valuable because almost everybody sucks at it. Okay. I don't mean at the social collaboration aspect. I mean most everybody sucks at using... Rome or Zettelkasten. It's almost as if no one understands what the point of it is. But, you know, that's a story for another day, I guess. I'm just supremely disappointed in the type of stuff that people use Rome for.

Unbelievable. But amen to each their own. When are you going to be bringing out your $1,000 course so you can teach us how to use Rome? See, that's the thing.

I don't, I actively don't want to teach a course on anything that I like because then I suddenly create a conflict where if something better comes along, I'm stuck. I'm locked in accidentally because If a large portion of your income is from teaching a course about some piece of software that you like, and then a better piece of software comes along, you have to switch.

Because or you can't switch because you have to keep talking about this software. So it's Yeah, I don't want to I don't want to say here's a better way to use Rome. I would rather say here's a better way to use notes but even i don't know myself but that's the reason why i'm not going around telling people that i know but oh well

It is what it is, man. That kind of thing happens a lot where people who don't know very much about anything related to the thing they're talking about, they just go, oh, whatever, there's a market, so I'll sell it. That's very true. That's very true. All right, guys.

Yeah, we haven't got very much time today, so we're going to keep it nugget dense. No, let's talk a little bit more. Okay, okay. Oh, yeah. All right. Yeah, I'm down. All right, cool. So, yeah, I want to ask, though, are you using a Zettelkasten system?

me right now yes uh well i've been using um obsidian but the thing is i don't think i use it like uh like a zettelkasten i think there's a big difference when you're like everything i write in that uh zettelkasten is to be viewed um by other people so it all gets um synced with git and that

goes to my wiki so that other people can read it and that i think that vastly changes the sorts of uh for example the links that you make like i think i mentioned this last time you know i link between things that are very obvious to me already because I want other people to understand the argument, but if I was just using it just for myself, then there are some links that I...

probably wouldn't make because they wouldn't want to bring down the average link quality i think we talked about this last time you did you perhaps disagreed but yeah i think you can definitely see a difference when for example you read uh andy matushak's notes he puts them up um not necessarily for other people to read like they're mainly for him i feel like so when you read through them

yeah it's like this probably requires some of some of these notes require andy's prior knowledge in order to understand them because they haven't necessarily been written with uh with the intent to have someone else read them and 100 understand the the flow and like there are other ways you can see that for example in this in the way that he doesn't have like an index or anything so you can't tell what's been updated or you know what's been added recently so yeah yeah but yeah yeah i mean

I think you should certainly have a section I think maybe you could do this with Probably if you just create a folder in obsidian, right? That's gonna make a subfolder in the vault and then in your git

you can just do git ignore of that folder, probably. That's true, that's true. And then that way, you can write private stuff in there, and then, you know, it'll have a different quality to a different character, and that might be... really interesting thing to try so I would definitely advise doing that because there's a lot of interesting use cases of this that wouldn't they wouldn't make sense to show other people because they're about your life or they're just sort of very random stuff

Like, for example, I this was a very recent example of value, direct value that I got. Very simple search in Rome. So I've been writing. lot every single day in Rome for weeks like way more than I ever imagined that I went thousands of words a day Wow okay and and I had noted that When I feel very sleep deprived, obviously I feel tired, but...

I also find it easier to get into a flow state. And so I noted that. And I just said, oh, I wonder if there's a link between sleep deprivation and the flow state. feelings of it when you're very very tired and you're pushing past tiredness there must be some sort of shutting off of certain parts of the brain and again I'm using that very loosely here because I'm not a neuroscientist but just go with me on it

So I assume there must be something or I thought there must be something similar. And so I link that and then I just go, I'll just search my sleep deprivation tag because whenever I feel sleep deprived, I tag it with that just so I can. find what happened to me on days where I was sleep deprived. And so I looked at that and then I searched, I limited the references to the flow on that page. And I realized I had made that link another time, like way earlier on.

So I made the same link accidentally twice, but it wasn't accidental. I had made the same association sort of roughly differently. And it just made me think, oh, now that I've thought of this twice, I wonder if there's something to it. You know what I mean? At least for me. It wasn't like a one-off experience. And so that's direct value. And I've gotten way more value than that. I really feel like the way that I've been using it.

is the way that it should be used at least if you're trying to get the utility out of it that i'm trying to get out of it so yeah i've been getting a ton of value out of that kind of thing but almost none of it is documenting stuff that i've learned and then you know, hoping that I can get value out of it that way. Almost none of my notes are notes about other things. It's notes about my life or notes about ideas I've had or ways that I've solved problems.

And so I'm not being hypocritical here by saying, don't use Rome as a second brain, and then I'm doing it myself. I am actually not taking notes on almost anything. Okay, that's interesting.

yeah i've always been a little bit sketched out uh you know writing about like um really personal stuff like stuff that happens in my life i don't think i've ever done that in uh super memo i've never done that with Rome or Obsidian rather or anything else just because I'd be like freaked out like I guess it's just because I've always wanted at some point just to be able to

put my entire sm collection up on the internet or something like that and just have anyone look through it or anything like that and i i'd feel uh really sketched out doing that if i was putting for example other people's information about other people up there as well i'd uh

I definitely feel a little bit more nervous about doing that. Oh, 100%. Yeah, and that's why, you know, there's a lot of coolness about, whatever, working with the garage door up or things like this. But also, it's limiting, man. Oh, yeah. You even Andy there's a lot of private notes like I know for a fact that his daily notes are not visible on me on the live version of his site which is exactly how i think it should be there should be some private stuff in everybody's life

Not everybody is entitled to know everything that you think or everything that you write. And if you're limiting yourself to only writing about stuff that's public, you're losing a lot of value.

I don't think that's worth it. There's a lot more value to be gained by this stuff. But anyway, I'm just... I'm recently very interested in this kind of thing, and I'm very interested in using it in ways that... using it in ways that feel natural to me and feel like I'm pulled towards them instead of out of some weird sense of duty or oh I have to take notes about this I never write anything because of that I only write because I think oh

That's really cool. I would like to save that and think more about it and this kind of thing. And if you're doing it out of some sense of duty, at least for me, it totally ruins the experience. It's not sustainable at all. And yeah, it's just very boring. And who wants to do boring stuff, man? I'm so sick of boring stuff. So many people, they talk boring. They say boring stuff. They do boring stuff. Just don't be boring, please. And if I'm boring you.

I apologize, but I can assure you I'm trying not to be boring. Please, just... I'm trying not to. Turn us off immediately if you find us boring. You don't have to sit through this. Yes, go do something else with your life.

If you find this boring, stop listening ASAP. Go do something interesting, please. You should turn us off immediately. Don't be boring. And flame us on Twitter as well. Say how bad we are. I'm terrible. Make fun of me, dude. Make fun of me. Shame me for being... come on bring it on

So when are we going to be seeing, you know, a fresh new wave of articles from you? Are we going to be seeing that imminently? Oh, I've been writing, like, almost all of my recent tweets and stuff like that are results of... stuff that i've written in rome like oh that would be good and it's almost all writing that i've just done in rome and just

Post it on Twitter. Obviously, I don't tweet very much. So that's not saying much but there like I said there are thousands and thousands and thousands of words that I've written in Rome over the past few weeks. It might be It might be a hundred pages. I really don't know. Like it's, it's a lot of stuff. So yeah. And that's just, that's only over the last few weeks. Obviously that's not enough time to determine the stickiness of the habit or.

whether it's a long-term thing that I'm going to do. But yeah, as of right now, I'm liking a lot. And it doesn't have to be Rome, but I will say I like Rome better than Obsidian. And... It's hard for me to articulate why that is I like something about the blocks. I think that's very good And whether people like the design of Rome or not, I think this is another case of over-exaggerating the badness of a user interface. I don't know why people do this. I don't understand it.

I understand that Rome is not like a polished up app that looks like it was made for mass market consumption. I understand that. But what I don't understand is literally say, oh, it looks terrible. It's awful. It's the worst user. Dude, come on. You know that's not true. You know that's not true. Come on. I don't know why people exaggerate this stuff. They do it with SuperMemo as well. Is SuperMemo's user interface great? No.

But it's not awful. I've seen some awful user interface. You know what has an awful interface and an awful user experience? PayPal. PayPal has an awful user experience. Even fast actions in PayPal feel slow.

Because of the way they've done the loading animations and the button focus states and the button active states and all this kind of thing So there are certain stuff that feels awful, but I just don't think Rome is one of those and obviously everything i say on here i'm not sticking by these opinions my whole life okay and a lot of this stuff a lot of this stuff i'm just saying because whatever i just feel like saying it in the moment and maybe i won't

Won't stand by it even tomorrow. So who knows? You can shame us for being boring. Shame us for being hypocrites. Nothing we say matters. Yeah, it's all nothing, man. It's all nothing. We're just living life here. Don't worry about it, all right? Just take a couple breaths. Relax yourself.

You know, drink some water, stay hydrated, be nice to yourself, get some sleep, get some exercise. Don't worry about it. It's all good. All right. There you go. All right. All right. Well, with that, should we call it? I mean, that's pretty good life advice. Yeah. Yeah. That was good. That was very nice. All right, guys. We'll see you in the next one. All right. See ya.