YouSaid · the spoken record

Grant Sanderson

lines on the record: 142
first: 2020-08-23
most recent: 2020-08-23
sittings or episodes: 1
sources: podcast

Every line below is reproduced as it was said and linked to the record it came from. Nothing here is summarised or generated. Directory · Search · Corrections

“I think that is what when we use the word meaning to mean like you're sort of filled with a sense of happiness and energy to create more things like I have so much meaning taken from this like that yeah that's what fuels my pump at least”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, so but what I'm saying is, I don't understand the question what is the meaning of life in that I think people might be asking something very real. I don't understand what they're asking. Are they asking like, why does life exist? Like, how did it come about? What are the natural laws? Are they asking, as I'm making decisions day by day for what should I do? What is the guiding light that inspires like what should I do? I think that's what people are kind of asking.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“You'd be like the height of what you can't ask what is the height without an object, you can't ask what is the meaning of life without an intentful consciousness putting it, like I guess I'm revealing I'm not very religious.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I'm just joking. I don't think life has a meaning. I think I don't understand the question. I think meaning is something that's ascribed to stuff that's created with purpose. There's a meaning to this water bottle label in that someone created it with a purpose of conveying meaning. And there was like one consciousness that wanted to get its ideas into another consciousness. Most things don't have that property. It's a little bit like if I ask you what is the height?”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“And it's not a coincidence that the things I'm describing aren't like the most up to date progress on the Rahman hypothesis cousins or like there's context in which the analog of the Riemann hypothesis has been solved and like more discrete feeling finite settings that are more well behaved. I'm not describing that because it just takes a ton to get there and instead I think it'll be like productive to have an actual understanding of something that you can pack into 20 minutes.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“And to be clear, you should be interested in fundamental questions. I think that's a good habit to ask what the fundamentals of things are. But I think it takes a lot of steps to, like, certainly you shouldn't be trying to answer that unless you actually understand quantum field theory and you actually understand general relativity”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, I think I've heard a lot of the interest that people send me messages asking to explain Einstein's thing or asking to explain Wolfram's thing. One, I don't understand them, but more importantly. It's too big a You shouldn't be interested in those, right?”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“You should spend your brain cycles on problems that you will see resolved because then you're going to grow to see what it feels like for these things to be resolved rather than spending your brain cycles on something where it's not going to pan out. And the people who do make progress towards these things, like James Maynard is a great example here of like young creative mathematician who pushes in the direction of things like the twin prime conjecture rather than hitting that head on, just see all the interesting questions that are hard for similar reasons but become more tractable and let themselves really engage with those. So I think people should get in that habit. I think the popularization of physics should encourage that habit through things like the physics of simple everyday phenomena because it can get quite deep.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Things that make Fairmouth's less theme hard are actually quite deep. And so the cases that we can solve it for, it's like you can get these broad sweeps based on some hard but accessible bits of number theory. But before you can even understand why the general case is as hard as it is, you have to walk through those. And so any other attempt to describe it would just end up being like shallow and not really productive for the viewer's time. I think the same goes for most unsolved problem type things where I think, you know, as a kid, I was actually very inspired by the twin prime conjecture that totally sucked me in is this thing that was understandable. I kind of had this dream like, oh, maybe I'll be the one to prove the twin prime conjecture. And new math that I would learn would be like viewed through this lens of like, oh, maybe I can apply it to that in some way. But you sort of mature to a point where you realize.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Yes, actually. Barely. Mathology might be able to do a great job on this. He does a good job of taking stuff that's barely accessible and making it. But the... The core ideas of proving it for n equals b are hard, but they do get you real ideas about algebraic number theory. It involves looking at a number field that lives in the complex plane. It looks like a hexagonal lattice, and you start asking questions about factoring numbers in this hexagonal lattice. So it takes a while, but I've talked about this sort of like lattice arithmetic in other contexts. And you can get to it.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I have not yet. No. But if I did, do you know what I would do? I would talk about proving Fairmaz's last theorem in the specific case of n equals 3.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, I think that's super important. I mean, so you see this in math too with the big unsolved problems. So like the Clay Millennium problems, Riemann hypothesis.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“It's not like this novel theory of everything type thing, but to understand what's going on there really requires digging in depth to certain ideas. And if you let yourself think past what the video tells you about, what does circularly polarized light mean and things like that, it actually would get you to a pretty good appreciation of two-state states and quantum systems in a way that just trying to read about the hard parts about resolving quantum field theories with general relativity is never going to get you. So as far as popularizing science is concerned, like the audience should be less interested than they are in theories of everything. The popularizers should be less emphatic than they are about that for actual practicing physicists. It might be the case maybe more people should think about fundamental questions.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, I think as far as popularized science is concerned, people are more interested in theories of everything than they should be. Because the problem is whether we're talking about trying to make sense of Weinstein's lectures or Wolfram's project, or let's just say listening to Witten talk about string theory, whatever proposed path to a theory of everything, you're not actually going to understand it. Some physicists will, but like you're just not actually going to understand the substance of what they're saying. What I think is way, way more productive is to let yourself get really interested in the phenomena that are still deep, but which you have a chance of understanding. Because the path to getting, to like even understanding what questions these theories of everything are trying to answer involves walking down that. I mean, I was watching a video before I came here about from Steve Mold talking about why sugar polarizes light in a certain way. So fascinating. Like really, really interesting.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, like if that's the actual case, that it's something we're consciously hearing people's disagreement, disagreeing with that disagreement and saying he wants to move forward anyway. That's an admirable aspect of leadership.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I mean, that's my knee-jerk reaction to the Walvis operators, I don't actually care that much either way. I'm not going to get viscally passionate. My initial reaction was like, yeah, this seems to make things more confusing to read. But then again, so does list comprehension until you're used to it. If there's a use for it, great. If not, great. But like, let's just all calm down about our spaces versus tabs debates here and like be chill.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“People like some parts of the benevolent dictator for life mantra, but once the dictator does things different than you once, suddenly dictatorship doesn't seem so great.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Is that actually true, or was it like there's a bunch of surrounding things that also was it actually the Walvre operator that?”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, so certainly Old Manum is way slower than it needs to be because of how it renders things on the backend is kind of absurd. I've rewritten things such that it's all done with like shaders in such a way that it should be just live and actually like interactive while you're coding it if you want to. You have like a 3D scene. You can move around. You can have elements respond to where your mouse is or things. That's not something that user of a video is going to get to experience because there's just a play button and a pause button. But while you're developing that can be nice. So it's gotten better in speed in that sense, but that's basically because the hard work is being done in a language that's not Python, but GLSL, right? But yeah, there are some times when it's like there's just a lot of data that goes into the object that I want to animate that then, just like Python is slow.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I mean, the biggest disadvantage is that it's slow. So when you're doing computationally intensive things, either you have to think about it more than you should, how to make it efficient, or it just takes long.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I mean, I love that it's like object oriented and functional, I guess, that you can kind of like get both of those benefits for how you structure things. So if you would just want to quickly whip something together, the functional aspects are nice.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“You're taking advantage of the fact that it's programmatic. You have loops, you have conditionals, you have abstraction. If any of those are like well fit for what you want to teach, to have a scene type that you tweak a little bit based on parameters or to have conditions so that things can go one way or another or loops so that you can create these things of like arbitrarily increasing complexity. That's the stuff that's meant to be animated programmatically. If it's just like writing some text on the screen or shifting around objects or something like that, things like that you should probably just use keynote would be a lot simpler. So try to find a workflow that distills down that which should be programmatic into Manum and that which doesn't need to be into like other domains. Again, do as I say, not as I do.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“It's amazing. You can get very, very far with it. And in a lot of ways, it would make more sense for some stuff that I do to just do in GeoGebra. But I kind of have this cycle of liking to try to improve Manum by doing videos and such. So do as I say, not as I do. The original thought I had in making Manum was that there's so many different ways of representing functions other than graphs. In particular, things like transformations, like use movement over time to communicate relationships between inputs and outputs instead of like extraction and y direction, or like vector fields or things like that. So I wanted something that was flexible enough that you didn't feel constrained into a graphical environment. By graphical, I mean like graphs with x coordinate y coordinate kind of stuff. But also make sure that”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I'll pretend like it isn't terrible for someone like Russ. I think step one is make sure that what you're animating should be done so programmatically because a lot of things maybe shouldn't. Like if you're just making a quick graph of something, if it's a graphical intuition that maybe has a little motion to it, use Desmos, use graph or use Geogebra, use Mathematica, certain things that are like really oriented around graph.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“That I can answer. So I always feel guilty if this comes up because I think of it like this scrappy tool that's like a math teacher who put together some code. People asked what it was, so they made it open source and they kept scrapping it together. And there's a lot of things about it that make it harder to work with than it needs to be that are a function of me not being a software engineer. I've put some work this year trying to like make it better and more flexible that is still just kind of like a work in process. One thing I would love to do is just get my act together about properly integrating with what the community wants to work with and what stuff I work on and making that not deviate and just like actually fostering that community in a way that I've been like shamefully neglectful of. So I'm just always guilty if it comes up.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“We'll get there. Yeah, I'm not saying I'd be impressed but surprised I'll be impressed, but I think we'll get there on algorithms doing math like that.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Would have only been so very subtly. And I think the difference between a math student making the mistake and a mathematician who's experienced seeing that kind of pattern is that they'll have a sense from what the problem itself is whether the pattern that they're observing is reasonable and how to test it. And I would just be very impressed if there was any algorithm that was actively accomplishing that goal.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, I mean, there's all sorts of wonderful examples of false patterns in math where one of the earliest videos I put on the channel was talking about you're kind of dividing a circle up using these chords and you see this pattern of 1, 2, 4, 8, 16. I was like, okay, pretty easy to see what that pattern is. It's powers of two. You've seen it a million times. But it's not powers of two. The next term is 31. And so it's like almost a power of two, but it's a little bit shy. And there's actually a very good explanation for what's going on. But I think it's a good test of whether you're thinking clearly about mechanistic explanations of things, how quickly you jump to thinking it must be powers of two. Because the problem itself, there's really no good way to, I mean, there can't be a good way to think about it as doubling a set because ultimately it doesn't. But even before it starts to, it's not something that screams out as being a doubling phenomenon. So at best, if it did turn out to be powers of two,”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, I think a lot of understanding is like, I don't mean to denigrate pattern recognition. Pattern recognition is most of understanding, and it's super important and it's super hard. And so when it's demonstrating this kind of real understanding, compressing down some data, that might be pattern recognition at its finest. My only point would be that what differentiates math, I think, to a large extent is that the pattern recognition isn't sufficient and that the kind of patterns that you're recognizing are not like the end goals, but instead they are the little bits and paths that get you to the end goal.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Expertise and exposure to lots of things. It's why it's good to learn from as many examples as you can, rather than just from the definitions. It's to get that level of intuition. But to actually concretize it into a piece of math, you do need to test your hypotheses and if not prove it, have an actual explanation for what's going on, not just a pattern that you've seen.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I honestly haven't dug into what's going on within it in a way that I can speak intelligently to. I guess it doesn't surprise me that it's bad at numerical patterns because maybe I should be more impressed with it, but that requires having a weird combination of intuitive and formulaic worldview. So you're not just going off of intuition when you see Fibonacci numbers. You're not saying like intuitively, what do I think will follow the 13? Like, I've seen patterns a lot where like 13s are followed by 21s. Instead, it's that the way you're starting to see a shape of things is by knowing what hypotheses to test where you're saying, oh, maybe it's generated based on the previous terms, or maybe it's generated based on like multiplying by a constant or whatever it is. You like have a bunch of different hypotheses and your intuitions are around those hypotheses, but you still need to actively test it. And it seems like GPT-3 is extremely good at that sort of pattern matching recognition that usually is very hard for computers, that is what humans get good at through”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Which of course it would because it's drained on like the internet data stories. But what you see unfolding is a description of COVID-19 if it were a zombie apocalypse. And the early aspects of it are kind of shockingly in line with what's reasonable. And then it gets out of so quickly.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“My absolute favorite. Yeah. So Tim Blay, who runs a channel called A Capella Science, who's like tweeting a bunch about playing with it. And so GPT-3 was trained on the internet from before COVID. So in a sense, it doesn't know about the coronavirus. So what he seeded it with was just a short description about a novel virus emerges in Wuhan, China and starts to spread around the globe. What follows is a month-by-month description of what happens. January colon. That's what he seeds it with. So then, what GPT 2 generates is like January, then a paragraph of description, February, and such. And it's the funniest thing you'll ever read because it predicts a zombie apocalypse.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Yeah, that's huge value. I mean, think of the writing process. Sometimes a big part of it is just getting a bunch of stuff on the page, and then you can decide what to whittle down to. So if it can be used in like a man-machine symbiosis where it's just giving you a spew of potential ideas that then you can refine down, like it's serving as the generator and then the human serves as the refiner, that seems like a pretty powerful dynamic.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“You know, I think I got an email a day or two ago about someone who wanted to try to use GPT-3 with Manum where you would like give it a high-level description of something, and then it'll automatically create the mathematical animation. Like, trying to put me out of a job here.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Chaos theory is another good instance there where it's like, yeah, a ton of things are hard to describe, but how do you have ones that have a simple set of governing equations that remain like arbitrarily hard to describe?”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“That tells us is that sometimes Possible programs are interesting in some way. It's like, yeah, tons of them. But the much, much more delicate question is when you can have a low information description of something that still becomes interesting.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Like one fun example of this, you know, Shannon's noisy coding in theorem, noisy coding theorem information theory that basically says if I want to send some bits to you, maybe some of them are going to get flipped. There's some noise along the channel. I can come up with some way of coding it that's resilient to that noise that's very good. And then he quantitatively describes what very good is. What's funny about how he proves the existence of good error correction codes is rather than saying like, here's how to construct it, or even like a sensible non-constructive proof. The nature of his nonconstructive proof is to say if we chose a random encoding, it would be almost at the limit, which is weird because then it took decades for people to actually find any that were anywhere close to the limit. And what his proof was saying is choose a random one and it's like the best kind of encoding you'll ever find. But what's”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, okay, so that from a Kalmogorov complexity standpoint, almost everything will be interesting. What's fascinating is to find the stuff that's describable with low information, but still does interesting things.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“But the reason the gradient descent works well with neural networks and not just choose however you want to parameterize this space and then like apply gradient descent to it is that that layered structure lets you decompose the derivative in a way that makes it computationally feasible.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, so on the one hand, it is, but also it's like not terribly surprising that you have these interesting points that exist when you make your space so high dimensional. Like GPT-3, what did it have? 175 billion parameters. So it doesn't feel as mesmerizing to think about, oh, there's some surface of intelligent behavior in this crazy high-dimensional space. It's like there's so many parameters that, of course, but what's more interesting is how is it that you're able to efficiently get there, which is maybe what you're describing, that something as dumb as gradient descent does it.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Exactly. It's automated abstracting, which, I mean, that just feels very powerful. And the idea that it can be so simply mathematically represented, I mean, a ton of modern in-mmel research seems a little bit like you do a bunch of ad hoc things, then you decide which one worked, and then you retrospectively come up with the mathematical reason that it always had to work. But who cares how you came to it? When you have that elegant piece of math, it's hard not to just smile seeing it work in action.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Math on each one. So that's a pretty beautiful idea that you can have a generalizable object that runs through the layers of abstraction, which in some sense constitute intelligence is having those many different layers of an understanding to something.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Well, I think what I would go to is the layered structure and how you can have what feel like qualitatively distinct things happening going from one layer to another, but that are following the same mathematical rule. Because you look at it as a piece of math, it's like you got a nonlinearity and then you've got a matrix multiplication. That's what's happening on all the layers. But especially if you look at some of the visualizations that Chris Ola has done with respect to convolutional nets that have been trained on ImageNet, trying to say what does this neuron do? What does this family of neurons do? What you can see is that the ones closer to the input side are picking up on very low level ideas like the texture, right? And then as you get farther back, you have higher level ideas like what is the, where are the eyes in this picture? And then how do the eyes form like an animal? Is this animal a cat or a dog or a deer? You have this series of qualitatively different things happening, even though it's the same piece of”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Sorry, I didn't interrupt. You should get Juan Bennett on the thing and then talk to him about permanence. I think you would have a good conversation. So he's the one that founded this thing called IPFS that I'm talking about. And if you have him talk about basically what you're describing, like, oh, it's sad that this isn't forever, then you'll get some articulate pontification around it that's been pretty well thought through.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Think it will be. I mean, you know, another variant that this might take is that you fast forward 50 years and Google or Alphabet isn't the company that it once was and it's kind of struggling to make ends meet. And, you know, it's been supplanted by the whoever wins on the AR game or whatever it might be. And then they're like, you know, all of these videos that we're hosting are pretty costly. So we're going to start deleting the ones that aren't watched that much and tell people to try to back them up on their own or whatever it is. Or even if it does exist in some form forever, it's like if people are not habituated to watching YouTube in 50 years, they're watching something else, which seems pretty likely. Like, it would be shocking if YouTube remained as popular as it is now indefinitely into the future. So it won't be forever.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“This is why you need IPFS or something where it's like if there's a content link are familiar with this system at all like right now if you have a URL it points to a server there's like a system where the address points to content and then it's like distributed so you can't actually delete what's at an address because it's content addressed and as long as there's someone on the network who hosts it it's always accessible at Address that it once was.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Which is again circling back. I do think the pandemic will serve to force a lot of people's hands. You're going to be making online content anyway. It's happening, right? Just hit that publish button and see how it goes.”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“I mean The honest problem is like a lot of the educational content is posted by people who were just starting to research it two weeks ago. Are on a certain schedule and who maybe should think, like, who am I to explain and choose your favorite topic, quantum mechanics or something? And the people who have the self-awareness to not post are probably the people also best positioned to give a good, honest explanation of it”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source
“Is scary. I honestly wish that more of the people who had Modesty to say, Who am I to post this were the ones actually posting”
2020-08-23 · Lex Fridman Podcast · #118 – Grant Sanderson: Math, Manim, Neural Networks & Teaching with 3Blue1Brown · IDENTIFIED FROM THE TRANSCRIPT · source