Maya ☁️➡️🌸

5.9K posts

Maya ☁️➡️🌸 banner
Maya ☁️➡️🌸

Maya ☁️➡️🌸

@mayaofspring

on an observation arc | alt: @springofmaya | also on 🦋

Singapore (sometimes London) Katılım Haziran 2024
542 Takip Edilen813 Takipçiler
shellac
shellac@she_llac·
5% chance a genie grants you a small to medium sized wish 95% chance you severely embarrass yourself in front of everyone you know
English
13
0
26
816
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
Maya ☁️➡️🌸@mayaofspring

Given that being very excited about Anki (and AI for it) is my bit recently, this post feels targeted 😅 and I have thoughts. I basically agree as for the core analysis, but I draw a more optimistic conclusion. I view spaced repetition as an overpowered algorithm for a very specific problem: given a key-value mapping, import it into the neural network inside your head. The reason it's overpowered is that, once set up, it's very low in the amount of executive function it demands of you. It just works! But while your post focuses on the cases in which it doesn't work in, the primitive is more powerful than it sounds. Note that neural networks, when being trained, seek out patterns to compress and generalise. You can induce emergent capabilities in yourself via a simple process! So, if your domain is such that a meaningful chunk of it can be condensed into a key-value mapping, such that knowing the key-value mapping would make it a lot easier for you to make sense of new samples of the domain, and such that at some point if you want to master the domain you would have to implicitly know the mapping anyway... the most direct way to advance in the domain then is just using the overpowered technique for ingesting a key-value mapping. The canonical example of a domain like that is foreign language learning. But also biology, history, or other such stuff that is inherently more broad than deep. And maybe others? I don't have enough experience outside of it to say, but I do want to stay on a lookout for domains that admit that shape. Even though may sound silly at a glance, I want to test out putting drawing practice into Anki with cards like "draw an eye, sideways, in style X". Knowing versus not knowing the broad basics really makes the difference, and spaced repetition lets one leapfrog the broad basics. But if you're thinking of domains like maths, flashcards seem more questionable. It's less of a game of mass-ingesting training data in a systematised manner; it requires thinking and playing around with concepts until they click, and the concepts build on top of each other. And while I wouldn't think of trying to make "all unitary matrices are invertible" into a flashcard, I suppose the point of your post is to make that clear. I guess I was quite confused when my computer science coursemates were flashcarding for their exams. The main part I wanted to argue with is in the second-to-last paragraph: > In highly systematized domains, you don’t need AI in the first place, because there’s nothing for the AI to do except import a CSV into Anki. Is (say) language learning highly-systematised? If so, where are you going to get that CSV from? Someone had to create that CSV, and they had a specific thing in mind for it; what if you have a slightly different thing in mind? The Mandarin deck I initially downloaded has been quite good as far as decks go, but soon I noticed problems with it: it teaches Beijing Mandarin (as most teaching materials do), while I want the "generic" Mandarin. It has mistakes here and there, like showing the wrong sentence. It doesn't explain words and phrases it introduces. And for Japanese, I already know it to an intermediate but uneven level, so any off-the-shelf deck contains a lot of cards I consider trivial. > but the AI can’t look into your mind and know which facts are salient for you Well, I asked it to probe my Japanese so that it knows what level to aim for. And then I said "I'm going to read はたらく細胞, could you make me a deck that teaches me relevant vocabulary so that when I get to reading it will be easier for me?" And then I also said that it should attach little usage notes to the words and phrases and grammar it introduces in its card for better reinforcement. As a result I end up with decks that fit my vision of what I want out of them a lot better, when creating them manually to the same quality would have cost a lot of labour from me in noting down new words, looking up words in the dictionary, and researching for those little hints. And yes, I would learn in the process, but it wouldn't be low-executive-function anymore. And then I also think that people just haven't been good enough at making Anki decks, even for those domains that fit spaced-repetition well! A great deck tests and teaches at the same time. The cards should have no ambiguity to them, and answerable entirely by reflex. The decks should aim to reinforce connections and provide explanations for the answers it requests of you; all to reduce frustration and engage all the relevant brain channels throughout. The deck that only maps between dictionary entries is going to be tedious and error-prone; the decks I'm experimenting with are the opposite of that. So, there's low-hanging fruit still. And language models, with all their broad latent knowledge, are the perfect shape for this task. But I suppose this isn't what your post is about, as excited as I am about spaced repetition for the things I care about at the moment. Though, I do wonder, if there might exist something that you're looking for, that would be similar in its function to spaced repetition, but for more conceptual domains instead. As the genius behind spaced repetition is in noticing that broad domains are learnable via a systematised process with a well-defined loop. Maybe what one needs is more like an onslaught of exercise sheets exercising those concepts instead? Best of luck in finding the process you're looking for! x.com/mayaofspring/s…

English
1
0
0
147
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
People are too excited about AI supposedly greatly reducing the cost of code, and not excited nearly enough about AI greatly reducing the cost of Anki decks
English
4
0
31
802
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
Given that being very excited about Anki (and AI for it) is my bit recently, this post feels targeted 😅 and I have thoughts. I basically agree as for the core analysis, but I draw a more optimistic conclusion. I view spaced repetition as an overpowered algorithm for a very specific problem: given a key-value mapping, import it into the neural network inside your head. The reason it's overpowered is that, once set up, it's very low in the amount of executive function it demands of you. It just works! But while your post focuses on the cases in which it doesn't work in, the primitive is more powerful than it sounds. Note that neural networks, when being trained, seek out patterns to compress and generalise. You can induce emergent capabilities in yourself via a simple process! So, if your domain is such that a meaningful chunk of it can be condensed into a key-value mapping, such that knowing the key-value mapping would make it a lot easier for you to make sense of new samples of the domain, and such that at some point if you want to master the domain you would have to implicitly know the mapping anyway... the most direct way to advance in the domain then is just using the overpowered technique for ingesting a key-value mapping. The canonical example of a domain like that is foreign language learning. But also biology, history, or other such stuff that is inherently more broad than deep. And maybe others? I don't have enough experience outside of it to say, but I do want to stay on a lookout for domains that admit that shape. Even though may sound silly at a glance, I want to test out putting drawing practice into Anki with cards like "draw an eye, sideways, in style X". Knowing versus not knowing the broad basics really makes the difference, and spaced repetition lets one leapfrog the broad basics. But if you're thinking of domains like maths, flashcards seem more questionable. It's less of a game of mass-ingesting training data in a systematised manner; it requires thinking and playing around with concepts until they click, and the concepts build on top of each other. And while I wouldn't think of trying to make "all unitary matrices are invertible" into a flashcard, I suppose the point of your post is to make that clear. I guess I was quite confused when my computer science coursemates were flashcarding for their exams. The main part I wanted to argue with is in the second-to-last paragraph: > In highly systematized domains, you don’t need AI in the first place, because there’s nothing for the AI to do except import a CSV into Anki. Is (say) language learning highly-systematised? If so, where are you going to get that CSV from? Someone had to create that CSV, and they had a specific thing in mind for it; what if you have a slightly different thing in mind? The Mandarin deck I initially downloaded has been quite good as far as decks go, but soon I noticed problems with it: it teaches Beijing Mandarin (as most teaching materials do), while I want the "generic" Mandarin. It has mistakes here and there, like showing the wrong sentence. It doesn't explain words and phrases it introduces. And for Japanese, I already know it to an intermediate but uneven level, so any off-the-shelf deck contains a lot of cards I consider trivial. > but the AI can’t look into your mind and know which facts are salient for you Well, I asked it to probe my Japanese so that it knows what level to aim for. And then I said "I'm going to read はたらく細胞, could you make me a deck that teaches me relevant vocabulary so that when I get to reading it will be easier for me?" And then I also said that it should attach little usage notes to the words and phrases and grammar it introduces in its card for better reinforcement. As a result I end up with decks that fit my vision of what I want out of them a lot better, when creating them manually to the same quality would have cost a lot of labour from me in noting down new words, looking up words in the dictionary, and researching for those little hints. And yes, I would learn in the process, but it wouldn't be low-executive-function anymore. And then I also think that people just haven't been good enough at making Anki decks, even for those domains that fit spaced-repetition well! A great deck tests and teaches at the same time. The cards should have no ambiguity to them, and answerable entirely by reflex. The decks should aim to reinforce connections and provide explanations for the answers it requests of you; all to reduce frustration and engage all the relevant brain channels throughout. The deck that only maps between dictionary entries is going to be tedious and error-prone; the decks I'm experimenting with are the opposite of that. So, there's low-hanging fruit still. And language models, with all their broad latent knowledge, are the perfect shape for this task. But I suppose this isn't what your post is about, as excited as I am about spaced repetition for the things I care about at the moment. Though, I do wonder, if there might exist something that you're looking for, that would be similar in its function to spaced repetition, but for more conceptual domains instead. As the genius behind spaced repetition is in noticing that broad domains are learnable via a systematised process with a well-defined loop. Maybe what one needs is more like an onslaught of exercise sheets exercising those concepts instead? Best of luck in finding the process you're looking for! x.com/mayaofspring/s…
English
0
0
3
257
𝚓𝚒𝚖𝚖𝚢
𝚓𝚒𝚖𝚖𝚢@202accepted·
this sounds ridiculous until I am saying the same shit about japan “after much research, I’ve decided to go to shiojiri, which means salt ass. there is a cool recycling center there i want to see. feels local, like the real japan”
B1G_Ryan@B1G_Ryan

Lmao

English
3
1
24
1.2K
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
@ben_r_hoffman @MunicipleOrrery I think you're basically arguing that Mao pre-1955 was good, and I'm arguing that Mao post-1955 was very, very, very bad? hmm I wonder if there is a famous quote that works as the synthesis here
Maya ☁️➡️🌸 tweet media
English
1
0
1
14
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
As I learn more, I realise that the reason why not everyone thinks of Mao as as bad as Hitler is down to, well, Deng's pragmatism again. He declared Mao "70% good, 30% bad" as a compromise for institutional continuity, but he surely did not believe anything close to that!
Jack@tracewoodgrains

The Anatomy of Ideological Capture: How Wikipedia Whitewashes Mao Recently, I posted a passing aside making fun of how Wikipedia frames Mao's legacy, assuming that what I saw was self-evident. I got predictable pushback from Maoists and tankies, which didn't surprise me. What surprised me was the number of generally good-faith left-leaning people in my circles who treated my assertion as absurd, asserted that the article was fine and balanced, and accused me of just wanting propaganda for my side. Now, I should be clear—brace yourself for controversy—I am no fan of Mao. I toss him in a bin alongside Hitler and Stalin as one of the three most catastrophic leaders of the twentieth century, one who had such an extraordinary combination of malice and will to power that he killed more people than perhaps any other one individual in history. As far as I'm concerned, his name is mud, and the good that has come to China should be recognized as a result of Deng Xiaoping, a man he purged twice, doing everything possible to reverse his policy short of undermining his own claim to rule. But I digress. That's not what I'm objecting to. I'm not asking Wikipedia to make a prosecutor's case against the man; I can do that myself. I'm upset because the section looks precisely how I would approach a statement were I Mao Zedong's defense attorney. First: start with glowing praise, every word technically defensible. Lead with all your good facts, looking for every convenient data point or stock line. Phrase them in ways that most everyone reading will instinctively parse as good. He's important, influential. He's a political intellect, a theorist, a military strategist, a poet, a visionary. He drove imperialism out of China, he unified China, he ended civil war (don't press too hard on the details of that war!). Find reforms you can claim for him, find a sympathetic survey or two, note that he reduced poverty. Spend a whole paragraph laying out nothing but praise for him. But people know he killed people! What do you do with that? Well, any lawyer whose client has some bad facts will tell you precisely what you do with it. You don't hide it—that just lets the other side bring it up. Makes you look dishonest. Be upfront about it, but massage it a bit. Tell the story from your protagonist's view. Make it land smoothly. You start by sandwiching it between good facts, naturally. Everyone's just had a paragraph about how great this guy is. Now you're ready to slide in that tens of millions of people died. But wait! Mostly, you can add, it was starvation (probably unintentional!), but also mumble mumble mumble executions etc. But he didn't usually give direct orders to kill! And according to one sympathetic writer, most deaths were unintentional, and the rest were "necessary victims in the struggle to transform China." Use his voice! Then, yes, yes, it's been described as autocratic and totalitarian, and people called him a tyrant. Yada yada yada, we know this. Anyway, he was compared to the first emperor of a unified China. Isn't that neat? Finally, tie it off with a neat bow: Forget about the deaths, the population grew! His strategies continue to be used; his ideology is popular and influential today! It's a picture-perfect defense. Would it be made stronger by omitting the killings? No! You've given people just enough to say that you're being honest, presenting a nuanced, thorough picture of a complicated man. Enough about Mao. People objected to my Hitler comparison because we're supposed to treat mass murderers who win and whose ideas remain popular as fundamentally different to mass murderers who lose. Very well. Commenters proposed Franco. Let's see what happens when you have a mix of defense and prosecution on a case, with the prosecution winning out. How do you start out this time? He's controversial. He ruled for a long time, he suppressed opposition, he ran propaganda campaigns. Hard to evaluate in a detached way—and look, his citizens were subjected to constant messages that he was good. You can't trust their objectivity! When you praise him, note that he's "significant"—who can deny that! but it's not Good, per se—and a successful counter-revolutionary—good if you hate revolution! None of the glowing praise to start things off. None of the fawning. Mao ran propaganda campaigns as well, Mao suppressed opposition as well—but it only merits mention with Franco. Onward! Note again that he's controversial and divisive. Present the supporter's case, making sure to frame it in ideological terms rather than the absolute-good terms used for Mao's positives. Good if you like anti-communism and nationalism, good if you hate socialism. And supporters credit those ideological stances for Spain's economic success! Add a bit about who praises and supports him and who opposes him. Next, find someone readers will have particularly divided opinions about, and be sure to contextualize him. While Philip Short is just Philip Short, William F. Buckley, Jr. is an American Conservative Commentator. Be sure to note that he praised Franco in explicitly divisive ideological terms, and recontextualize his statement: Franco wrested government "from the democratically elected government of the country." Then present the critics' case unsparingly and directly, using examples everyone will agree are bad things: thousands of deaths,political repression, complicity in Axis crimes. (The legacy section continues for many more paragraphs of minutia, most of it negative.) --- Do you see the difference? Do you see the shape of each? Franco is presented unsparingly, his crimes understood, with most praise presented in divisive ideological terms and criticism presented in universal terms. Mao's entry is practically a coronation speech for a paragraph, followed by carefully mitigated bad facts before ending strong. Maybe it's obsessive or neurotic or what-have-you to write all of this, but—to use the internet's erstwhile favorite term of abuse—I genuinely feel gaslit. You guys are reading the same article as I am, aren't you? You're seeing the same paragraphs I am. It's propaganda! It's clearly propaganda! You're not reading a thoughtful, nuanced, balanced take on a complex individual, you're reading propaganda for a mass murderer and then telling me I'm being silly and ideologically captured when I point out it's a bit weird. Propaganda does not stop being propaganda because it acknowledges bad facts. A defense attorney does not stop being a defense attorney when they let some criticism slip in. Glowing praise followed by a concession to reality does not a balanced portrait of a mass murderer make.

English
1
0
18
728
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
@MunicipleOrrery @ben_r_hoffman This is one of those cases when one can either discuss how to attribute the opaque trends as seen from the outside, or just peek inside and see that at all steps Deng&co were the ones trying their best to run the country, while Mao for his last 20 years only brought destruction
English
1
0
1
44
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
@she_llac ok maybe depends on the application! (the current one I'm looking at is "learn a foreign language", where slop-brain risk seems minimal)
English
1
0
4
88
Maya ☁️➡️🌸
Maya ☁️➡️🌸@mayaofspring·
@xav_moss maybe less advanced for me but I'm restarting my Japanese learning by setting up thematic Anki decks (this one is for reading Cells at Work)
Maya ☁️➡️🌸 tweet mediaMaya ☁️➡️🌸 tweet mediaMaya ☁️➡️🌸 tweet media
English
1
0
0
140
zephyr
zephyr@ShadowyZephyr·
@mayaofspring i almost accidentally typed "~" instead of "!" which would have changed the meaning quite a bit
English
1
0
5
38
zephyr
zephyr@ShadowyZephyr·
"woke left" "woke right" actually i just woke up
English
1
0
27
302
unconed 🛸💫👻
The Solaris analogy of incomprehensible alien motivations and thoughts is a good one, but I think the reality is more dull than that. The alien-ness of AI isn't due to there being something different from us on the inside, it's that there's nothing on the inside. Evolution never needed us to distinguish a person from a rock that pretends to think, so we are bad at it.
Maya ☁️➡️🌸@mayaofspring

AI expression, human expression In Solaris, a 1961 sci-fi novel, humans discover an alien intelligence in the form of an ocean planet. It manifests in various complex phenomena, including being capable of creating very fine simulacra, apparently sourced from the researchers' memories. It's unclear what goals the ocean has in recreating the protagonist's past lover, or if it has any goals at all; the novel revolves around the sheer alienness and incomprehensibility of it. (I appreciate the Polish education system for having made me read it) Whenever I see diffusion models' output, I think of the Solaris ocean. The process is unlike nearly any other that came before it; it's not a physical tool, and not a mechanical algorithm either. It's an alien mind on its own, that "knows" of human concepts and how to represent them, yet the array of pixels it outputs is a product of inhuman intent that we've only scratched the surface of. And then people put "make me a poster" in the input field of the alien mind and print out the result. It's difficult for me to not get quite reactionary about the existence of AI image generation. Socially, it enabled new forms of deception and lowered the barrier to entry for them. Aesthetically, the deluge of 'AI slop' made human environments, both offline and online, less pleasant to explore; I believe this is the complaint that @zetalyrae made. I think what grates is that alienness, such that, when the human uses the image verbatim, at 'face value', it feels off, as if something inhuman is in the room; and the models getting better doesn't quite wash that feeling off, if anything it makes people more paranoid. But anyway, you're in a place when you need some image. what do? 1. No image A perfectly acceptable option, if a little bland. Bad image can be more unpleasant than the absence of an image. You can invest your aesthetic points elsewhere, like the choice of typeface. 2. Clip art The traditional slop option. The problem with AI images isn't really with the thoughtful users, it's with the careless ones. Now, this isn't quite directly actionable anymore, as typing a prompt is inherently less effortful than assembling clip art, but I think it's interesting to note that often the aesthetic experience of walking around is based on whatever default people know to reach for in the area. Unfortunately, Microsoft Word largely failed at its potential of silently uplifting people's aesthetic experiences, or not making them worse. In Japan, the situation is somewhat better thanks to Irasutoya. Essentially, there is a singular go-to website for clip-art that everyone knows about; the Irasutoya illustrations are of decent quality and maintain a cohesive style, though maybe a bit too cutesy for the Western taste. Personally, the careless clip art in Singapore MRT makes me annoyed, but not as much as an equivalent AI image would. In Japan I find the Irasutoya collages quite lovable, and I would feel rather sad if AI images were to displace them (which they did in one of the hotels I checked into, unfortunately) 2.5 Stock photos Also the traditional slop option; I don't have anything clear to say on them yet. 3. AI image The modern slop option. And this is where the difference betwen careless and thoughtful use gets more stark. The model output is going to reek of inhuman intent; can you select and frame the image as to turn it artistic instead? I mostly use the quoted discourse to let me gather my recent thoughts and ramble a bit about this topic once I realised that the reply I started writing got too long to fit in tweet size, so I don't quite feel like litigating @tracewoodgrains's blog post cover images in detail; I think some of them are used well, some less so. The ones that are good tend to show intent as to the specific image choice, but also, in some sense, utilise the inhumanity positively, for example by inducing a dreamlike atmosphere in the use of the image. There is something to be said about the 'use'/'mention' distinction, but in the realm of visual language instead. Usage I don't like tends to use the inhuman intent as if it came from the author. Usage I like treats it more as a 'mention', a quoted thing. I probably can't explain more specifically than that. Overall I would like to lean on the norm of not using AI images though; most I've seen make me uneasy, in a bad way. (and it's one of the things I like less about "TPOT" at large) 4. Make images yourself. Confession: I don't know how to draw. I've got some intuitions as for composition, and I can move around nodes in Inkscape, but the ability to create a visual representation of an object on a page based on my imagination is one I presently lack. And I'm dissatisfied with that. It's a language that I don't currently know how to produce anything in, and it feels like this blocks an important avenue of my self-expression. Fortunately, I also think that things are fundamentally trainable with intention and repetition. I don't suppose I would become a great visual artist, as aptitude matters, but, just as I'm currently spending an hour a day on Mandarin and making steep progress, I believe that I could get similar effects with drawing practice, if I only make sure to proceed consistently and methodically this time; my prior attempts were akin to painstakingly translating one sentence over 2 hours to make it perfect, which is clearly not what effecting practice is like. Essentially what I want is to be able to navigate more domains; current me, fluent in Polish, fluent in English, halfway there in Japanese, able to write computer programs, able to spot a good photo opportunity, able to decorate a place, etc., has a vastly richer experience of the world than the 10-year-old me who could do only the very first thing. (One thing that struck me about visiting Japan again last week was just how alive it feels in the little illustrations everywhere; the Japanese seem to be very widely trained in expressing their thoughts via drawing, which I'm also told comes from often treating art as a communal thing. As a result I made the resolution with my travel companion for both of us to lock in on drawing this year, which I'm meaning to start soon once I settle down a bit more) 5. Establish a relationship with an illustrator There is a social component to art; it's not just us floating in a soup of disembodied texts and images. If you can team up for a joint vision, you and your readers are going to find that considerably more meaningful. ---------------------------------------------------- I don't think if I've got a clear prescription. I will not use AI images myself. There is a sense in which perfect is the enemy of the good. But also AI images are not going to count as good to lots of people. Expanding the toolkit of one's expression is a valuable thing to do, and it's clear that some cultures do that more than the others. I would prefer the culture where people near-universally can draw to the one where people delegate it away to an AI model specialised in direct mimicry.

English
1
1
7
861