逐字稿 🔗
speaker-0 (00:00.226)
Message is both magic of like, like taking it seriously as like magic and sorcery, but also just software and computation is this magical, wondrous thing. When I think back to like my childhood of like getting involved in computers, but a lot of it was just delight and excitement of like it really feels in many ways like this is the embodiment of all these things that I have been thinking about for all these years. I no longer feel like I'm like managing it. I'm almost just like a patron for the creation of software. If our brains kind of I don't know, quell before that.
It probably means that there's going to be limits to how we think about systems with millions or billions or trillions of interacting components. But yeah, there's always going to be this kind of weird messiness in terms of how we think about the world around us.
speaker-1 (00:48.098)
Well, first all, welcome to After Coding. We have the perfect guest today, Sam Arbusman, most recently author of The Magic of Code, but a renaissance man a bunch of different ways. You for 11 years you're a scientist in residence at story deep tech firm Lux, which, you know, is is an interesting job description shop mixture, right? Like as a systems programmer who is a GP at a venture capital firm, I know what that's like. But you know, wanna talk about your your most recent book, first of all, which is incredibly germane to our topic, which is the fact that, you know, coding is
Seemingly not a thing that people do, at least for industrially productive purposes anymore. And I'm just curious, like, you know, first of all, the thesis of your book as I understand it, the kind of metaphor that everything that glues everything together is the idea that code is sort of the ancient idea of words of power or a magic spell made real. It's it's you know, literal linguistic stuff that transmutes the world around us. Is that a kind of a fair view of the of the large scale picture of the book?
speaker-0 (01:44.638)
Say that's definitely one aspect of it, right? Of this idea that like we have had this desire for bulenia through our stories and the kind of things we we tell ourselves of like, yeah, controlling the world through text and now it's a reality. But I would say more broadly, the message is both magic of like, like taking it seriously as like magic and sorcery, but also just software and computation is this magical, wondrous thing. And we should take this seriously. And I I feel like when we kind of talk about like computing and technology.
There's a lot of talking past each other of like saying, like, it's like it's very useful, or we're concerned about it, or we're worried about AI and all the attendant things. but when I think back to like my childhood of like getting involved in computers, I'm sure there was like a certain amount of like cautious worry and things like that. But a lot of it was just like delight and excitement of like fractals and screensavers and I don't know, like my early Commodore of Vic 20 and just like like SimCity, like all this weird stuff. And
And I feel like we've kind of lost that. And so my my my goal with the book was both to kind of talk about like the magic stuff specifically, but also more broadly as just recognizing that computing, it's not just this branch of engineering. It's really this like humanistic liberal art that connects to so many different things. Like, like when you take it seriously, it should connect to language and philosophy and biology and art and how we think and I don't know, the nature of reality and all these different things. And so the book kind of tries to.
Like slowly but surely kind of spiral further and further outward until it says, wait, everything at least has some relevance to how we think about computation.
speaker-1 (03:19.212)
I was often drawn to the hardware, software interface and sort of still am, because it almost feels like this interface between the world of ideas and the world of things, right? It feels like kind of the code is this almost platonic stuff that's is made of almost pure thought and pure ideas. And yet it makes your floppy drive make sounds and you know, makes the fans spin faster and stuff like that, and does these things in the real world in a way that sort of still feels a little magical to me. And I think that idea of kind of code and and the world of computation as a fit subject of wonder for me,
I think there's a generation of folks where like Goldulusher Bach by Douglas Hofstetter was like sort of an introduction to this kind of way of thinking about computing where it's just sort of expansive and totalizing and you know connects all of these big questions in philosophy and physics and things like that down to computation in a way that's relatively approachable in a single book. And it's exciting to see somebody doing this for an audience that doesn't presume a little bit more of the STEMI orientation that that GEB seems to sort of presume. so I think it's a it's a really great contribution.
speaker-0 (04:17.739)
I appreciate that. Well, thank you. Yeah. And I yeah, I I read that book when yeah, when I when I was younger too. And and in many ways, for me, writing the magic of code, it was also like I when I was younger, I had to kind of cobble together lots of different of these ideas of like, I don't know, doing weird things around like artificial life and thinking about like creativity and code and all these other different kinds of things. And from in in some ways, then this book is almost like a
a gift for me to like pay it forward of like this is the kind of thing I wish that like young Sam had. If I couldn't have it, then at least hopefully future young Sams can have that. Now, of course, I mean the world of code is changing drastically, even like since the book came out last year. but I like to think that there are still some sort of perennial truths in it in terms of like this this way of exploring the
Load full transcript
the wondrous and delightful side of of computing.
speaker-1 (05:10.83)
And I think in some it's like searching for those truths is kind of the purpose of this podcast, right? The the project here at some level is people from my generation and and my cohort who spent a lot of their adult lives trying to get great at building software kind of the old-fashioned way, you know, one character at a time in Vim or Emacs or what have you. A lot of them went through sort of some kublarass grief at some point over the last year about, you know, having this thing with this this skill set that they've cultivated seemingly get commoditized. And I'm curious where you're at on that spectrum, right? I mean, I know you're not a professional programmer, but you know, you
use code in your work and and obviously use it for your books and and appreciate it as a as an aesthetic medium. Like are you at denial, anger, bargaining, depression, acceptance? is it sort of a good thing actually? We know where where are you at?
speaker-0 (05:55.736)
So I I would say I'm probably closer to acceptance. for me, the way I kind of think about this, I mean, certainly when these tools first came came on came online, yeah, I initially with all, this is great. Like I and then I thought, maybe like they don't quite do what you expect. And then there was kind of this like, I know, trough of despair. And then, then it got really, really good and you're kind of trying to figure out all this kind of stuff. I for me though, one of the things that I've been
always thinking about is this idea of like how do we kind of democratize software creation. Like this is one of the things that I think a lot about. And so I so for me, I mean, I as you mentioned like I I'm never I've never been a professional programmer. I've been sort of an amateur my entire life and like even like when I was like working in the sciences, it was and programming for some of my research, it was the way a scientist programs, which is not so pretty. and so and so I I kind of recognize that yeah, the the kind of coding I'm doing is very, very different.
That being said, going back to kind of this democratization idea, so my my first experience of programming was with this piece of software called HyperCard, which kind of it was I think it was bundled beginning in like the n late 1980s and with the Macintosh and went through I think like the late 90s. That was my on-ramp. And so I mean, I guess for for those of the audience who's not familiar HyperCard, although you you're nodding, so you're definitely familiar with it. It's basically this like authoring environment where you are building stacks of virtual cards.
Where each card can have text and buttons and things like that, and you kind of move between the cards and the stack. it's kind of like a primitive, like primitive on your computer website that you're building. And so, and the nice thing about HyperCard is it embodies this this phrase attributed to the computer scientist Seymour Papert, who was involved in building logo, this phrase low floors and high ceilings, this idea that there was like low floors. It was a very easy on-ramp to get involved. So you could just start by drawing or typing in text.
But then it had very high ceilings. And so you could actually, you could drag a button and make that button do something like make some beeping or or take you from one page to another or one card to another. But then under the hood, there was an entire language. There was an entire language of hyper talk. and so it had this idea of like low floors, high ceilings. And that was my on-ramp to program it. And so I found this kind of thing wonderful where I didn't really know how to program before playing with this.
speaker-0 (08:18.786)
believe. I I'm trying to remember the the actual cadence, the the actual timing of this. but it really got me comfortable with this kind of thing. And I've always wanted more of that kind of experience of like this democratizing programming or or some people talk like end user programming where it's like the the end users are actually able to modify and modify the software that they're working with. And we had never really had that for a very long time. The closest thing that was in widespread use, even though many people don't think of it this way,
was like Excel and spreadsheets, where because like they were highly programmable. they're kind of messy, but they're highly programmable. And they were programmable even by many people who never even thought that they were actually doing anything like programming.
speaker-1 (08:58.998)
people who are who can build incredible things in Excel who don't think of themselves as programmers were happy to explain why it's Turing complete and you know
speaker-0 (09:05.838)
no, it's it's unbelievable. And and it's also one of these things where you know that it's very, very powerful and open ended because the kinds of things people are building in it are things I imagine no one in their right mind in Microsoft could have ever imagined was be like would would be actually be able to be created. And and so for me, like that open endedness, whether Turing completeness or not, like is a hallmark of like a true, a truly expressive language. And for a long time I had been like looking at people who had been doing certain things in kind of
Program synthesis are kind of like ways of making it easier to generate code using using computer computer programs themselves. It was still like there wasn't that much going on. And then of course, now with the advent of like generating like generative AI and code generation, it really feels in many ways like this is the embodiment of all these things that I have been thinking about for all these years. Now, I I feel a little complicated about it because on the one hand, it it is, it has become so incredibly easy.
But in many ways, it almost feels, and I think someone wrote about this when they were using
Fable when it first came online. They're like, I no longer feel like I'm like managing it. I'm almost just like a patron for the creation of software where it's just like, here's the idea, and then go off and do it. And then it kind of can do things you never imagined. Now, of course, there's still a lot, like if you want it to work really well, there's still a lot of things you need to do. but I think another way that I think about this in terms of, I don't know, maybe not quite stages of grief, but maybe more taking my
philosophical step backwards is like programming, it is deeply changed and and changed in so many ways in the past like few months, let alone like the past few years. it's always been this moving target. Like like when I look back to like the early days of programming. So like the very early days when you're like plugging in different cables and doing things with I don't with any acquiring like I can't do that. And then there's like
speaker-1 (10:58.196)
was a electrical engineering activity.
speaker-0 (11:01.678)
Deeply physical. And then you had punch cards and other kinds of things. and then I look at like the early days of like the things that like Steve Wozniak was doing and unbelievable. Would I ever want to do it? Would I could I ever do it? No. And like I like this is like one and and even when I think back to like my own programming experiences, like there's many programming languages that I learned when I was younger that I would never want to go back to. and and I don't necessarily want to say, like we have to subject every single person who ever programs to those earlier experiences. And so the way I kind of think about
Programming is yes, it has changed. I'd like to think the fundamental idea of like trying to take some sort of idea in your mind and like embody it within, within software, in some sort of rigorous and precise way. That kind of thing is ultimately still software development. What it what it entails has changed. And there's been and we've changed kind of how we abstract away details, things like that. I'm also reminded in this is there's this story from the Talmud where
and so the Talmud and this like Jewish repository of knowledge has like a ton of different like stories and and and Jewish law. There's all these different things. But in one and at one point, there's this kind of crazy story where God and Moses are having a conversation and God is doing something with some text. And Moses is like, What are you doing? That he's like, Well, like, what are you what what's going on here? And he's Well, I'm gonna I'm gonna do this because like it like I think like a thousand years, two thousand years in your future.
Which is our past, there's going to be some rabbinic scholar who is going to interpret things based on like adding like little like crowns to the different letters. And so Moses says, like, I want to see this guy. And so there's this weird act of like time travel where God kind of takes Moses, transports him like 2,000 years into his future, whatever it is, and then puts him in the back of the house, like the house of study, where this this rabbi is giving like giving a giving a class. Moses realized he doesn't understand anything this guy saying.
And and it's this weird moment where he's like, I'm the one who has received the law. And like, and I don't understand what this, what this rabbi, like what this scholar is saying. And then until the scholar says, and how do we understand this? It is because of the law received by Moses from Sinai. And at that point, Moses is calmed because he realizes, okay, I might not understand what is being done right now, but there is this clear connection. There's this clear tradition and like and and path of like,
speaker-0 (13:30.236)
this thing led to this thing. This thing. I might not understand how we got to where we are now or how where we'll be in the future, but I know I'm on this continuum. I'm part of this larger tradition. And I wonder if, to a certain degree, programming is kind of the same thing where I don't necessarily understand how things were done back in the day. I might not understand how coding is done far off in the future, but I kind of know I'm part of this larger tradition of programming. And
And that makes me feel kind of okay. I mean, in terms of like how I think about what I learned and what is lost, yeah, there's some of that too. I mean, like I there's certain amount of delight sometimes and sometimes exhaustion and annoyance with like debugging, but like sometimes there's fun in like debugging a little bit. Like a a a really good debugging story is a beautiful thing to behold. Indeed. And it's like not doing that, not not necessarily feeling.
The need like it's fun to also just like tinker with the code itself, which I still sometimes do because I like doing that, even if it's no longer necessary. And I think this is also people are beginning to tease apart like different kinds of programmers. Like there are some who are like focused on the craft and the process, and other ones who are focused on just the outcome. And it often depends on the kinds of things you're building. Like if you are just trying to build software that is going to be in commercial production, being used.
By a huge number of people, or you're a startup and you have to get something to market by a certain amount of time, then you're maybe going to be a little bit more focused on the outcome. But if you're, I don't know, thinking about procedurally generated art, or you're just building something for yourself, sometimes you might care about the outcome because you just want to make the thing and then get on with your life. Other times it's kind of fun to just dive into the details. And so I think we're still grappling with all this, but in terms of how I feel, I'm I'm delighted, I'm excited.
certainly sometimes worried and I sit with some loss every now and then, but but I think that's
kind of that's just the world's how it's always been.
speaker-1 (15:36.182)
It's just change. Yeah. I I I it didn't come out with as well here that it seems like in practice we're gonna be it's obvious we're still gonna be building a lot of software, like some of that you know, some of the process of that has changed a little bit, but some things are are gonna remain constant and trying to find that those fixed points, I think, is is some of the project that we're under here. The you're also a man of letters, and you know, you've spent a lot of your awaken life crafting sentences and trying to get ideas across in a clear way and
unwittingly training our generation of of text generators as well, with or without your consent. In many cases, there are differences. Maybe it's partly the sort of like artistic temperament and the technical temperament, but I've noticed there's often like kind of an air gap between people who, you know, self-identify as authors and people who self-identify as programmers here. how do you feel about the fact that, you know, a lot of the emails you're gonna read and a lot of the blog posts you're gonna read and a lot of, you know,
Scripts like the script we sent over for for our conversation today and things like that are going to have kind of the fingerprints of these of these weird text generators on them.
speaker-0 (16:42.082)
So that also I think is it's it's complicated. Okay. I'm not even sure I have I and I would say, I mean, well, first of all, more broadly when it comes to AI, like I don't think I have like a single perspective. I kind of feel like every single perspective, probably multiple times a day, which I think is I think everyone should feel that. Like I concerned, excitement, worry, like just like being blindsided by things. when it comes to like the tech stuff, I mean, I think and certainly, yeah, when you get.
some email that's kind of long and you're like, okay, clearly it's been AI generated. You're like, okay, this person maybe hasn't spent the effort in writing this and now is just offloading all that effort to me. And like I now I have to read this long thing. And and that feels a little bit selfish or disrespectful to the reader. but I would say separate from those kinds of things, like in terms of like writing, for me, like writing is like the act of writing is
The act of thinking. and I think there's many writers who they feel that kind of way. Like, and like, and if there like without that, like certain amount of like struggle, not like a not struggle and difficulty, like it's not one of these things where I'm I'm gonna say, like, like, I don't know, like certain physicians who like they had to like stay up for 36 hours straight and do crazy things during their training. And so therefore they impose it upon the n the next generation. Like I don't wanna do that. Like, and I and I'm also a believer in like the fact that like ultimately.
And computers are tools for thought. Like we can use
them in this very powerful way to enhance our thinking, and enhance connectivity. and Steve Jobs has this whole idea of like the bicycle for the mind. And like and is just very briefly, like the way that that kind of works is he read this essay in Scientific American, I think back in the 1970s, where it was looking at like energy efficiency of locomotion and it was looking at different species. Humans were kind of mediocre, but then they also measured like a human on a bicycle, and the human on the bicycle kind of blew everyone away. And so his idea is like, like computers are sort of bicycles, but for the mind.
Of like enhancing it. And so I think we should definitely think about using computational tools, whether AI or not, as bicycles for the mind. The problem becomes though is like when you're not even steering it, then you kind of lose that aspect of like actually like working through what you actually are thinking through the writing. And so, yes, using tools to kind of like help like find interesting articles or connect different ideas or remind you of things that you might have like.
speaker-0 (19:10.39)
written about in your notes a while ago or whatever it is. I think those kinds of things are really exciting. And we need more of those kinds of tools and we need to experiment more of them. But I think we still kind of by analogy with the bicycle of the mind, like still need to spend some time like making sure that we're like, I don't know, steering the bicycle. and and when yeah, and when we just get all these texts that feel like, no, this is not a bicycle for the mind. I'm not even sure this is like a driverless car for the mind. Like a
Who knows what it is? It's just like something on Rails or whatever. I didn't know I'm struggling to find the right analogy here. that then doesn't feel as as meaningful or as proper for for writing. but yeah, but I would definitely say like ultimately, like if a tool is kind of like helping you to figure out what you yourself are thinking about something, then probably that's like a right like that, that's a useful indicator, a useful signal for maybe this is something good. If it's basically,
telling you what to think, then maybe that's not as useful.
speaker-1 (20:13.582)
And in some ways, this is like a there and I could ask this question about programming too, maybe pumping back out to coding for a second. One of the things that I didn't necessarily expect beforehand that seems to be manifest in in lots of people's experience with coding agents is that your expertise in the before times, right? Like your your kind of level of virtuosity as a computer programmer in the old-fashioned way seems to almost be amplified by these things, right? It seems like the sort of people at the high end of of those skill sets, if they
Once they get serious about it, once they take the sort of coding agent workflow seriously, get more out of it than people who are kind of using it to avoid learning how to code
at the very bottom. But let's say like your your entry level intern or something. Your intern seems to be just sort of using it like grind out their geotickets and then go back to watching TikTok or something.
speaker-0 (21:00.62)
There was some writer who was talking about yeah, like with like on the one hand, right, like you should view these tools like, it should be helping everyone. On the other hand, though, if this this person was saying, like, if someone who was not trained in programming was looking at the conversations that he was having with Claude Code, he'd be like, this is like, this is like a foreign language, like I don't understand anything here. And so right, there is this kind of amplifier. and I wonder how we prevent AI from like like from creating that like that bifurcation where it's like like
We we want AI to make people who are good at something even better. But we don't necessarily want people who are bad at something to just like make them even worse. and and I feel like we can see paths where it's kind of creating that thing where like, you're not really good here. Now you're not gonna be you're gonna be really not good at like a crazy scale. and and that that doesn't seem like a useful way to think.
speaker-1 (21:53.794)
There, we we just talked with Kent Beck on the show yesterday and you know, sort of T V D Agile XP and so on. And he's a fascinating guy to talk to. He's been he's always been sort of relatively creative and broadminded about programming workflows. And so interesting time to to have him on the show. And a thing that he's sort of trying to make happen is casually referring to the coding agents as genies. Okay. which is a little bit of an interesting connection to kind of your your code is magic idea, right? Which is, you know, there are these wielders of magic.
They're very literal minded and they're not quite on your side. Or at least it's ambiguous whether, you know, whether they're on your side yet or not. and there, and we could kind of like go on with this too a little bit, taking that idea, which is you know, like a golem from sort of folklore that, you know, when we sort of wrote a word and put it in its mouth and it sort of sprang to life, but it doesn't really feel like anything to be it, and it's kind of unclear and alien exactly what its motivations are and stuff. are there any, does the CODIS magic analogy taken seriously unsettle you ever?
Are you ever sort of do because I mean there's a dark side to sort of some actual forces, right?
speaker-0 (22:59.534)
yeah, no, I I think right, like if you take it seriously, right? There not only are there implications for like yeah, how we think about like recombination of text, but then there's also implications of like there's the idea of like sacrifice, like something of value must be given to allow some magic to occur. And one analogy is like all these unbelievable AI systems operate, but like we've had to, I don't know, sacrifice huge amounts of creative texts and creative works at at these altars and like and pour all these things in. Is that is there a downside there? And I think
related to like the genies or I or the golem or all these different kinds of things is this idea of like you are creating something, you don't have complete control. I mean, like you see this in like the sorcerer's apprentice, both kind of I like I mean it can even be like the Disney version with like all the the the brooms and the water
speaker-1 (23:45.934)
Understand in motion and yeah, exactly.
speaker-0 (23:48.066)
Yeah. And and the truth is, I mean, like I mean, people talk about this like as I I think about like the alignment problem. But like people have been talking about these kinds of ideas for a long time. And so like if you look back in so like Norbert Wiener, who is kind of like the developer of cybernetics, he has this wonderful little volume, which I believe is sort of a collected collected volume of like maybe like modified versions of some t of some speeches he gave. It's called God and Golem Incorporated. And in it, I mean, he kind of anticipates all the things we're talking
And he talked he's talking about and he gives the example of like the monkey's paw story, where it's like, yeah, the the shriveled monkey's paw, you make a wish, and then you get the wish, but like in the most horribly twisted version that you could have possibly imagined. in the story itself, like I think someone asked for like a hundred pounds, because it's I think it's a British story. and the next thing you know, like someone like someone knocks on the door and says, I'm so sorry, like your your son has been like horribly die like died horribly in some the industrial accident. we're not gonna admit
Liability, but like here's a hundred pounds for your trouble kind of thing. And of course, like this is like not what he wanted. and and I think right, there's always this sense of like how do we kind of understand these systems that are very powerful and are, yeah, maybe a little bit out of our control. And and this is I mean, another and another aspect of this is also just like how do we think about systems we don't fully understand? the author voted.
speaker-0 (25:18.446)
Well, thank you. Yeah. Yeah. So overcomplicated. Yeah, right. It's so overcomplicated is essentially about right systems that we built or you just work with on a daily basis, like even as an expert, that are not only more and more complex, but increasingly incomprehensible. Like and like what are the forces that lead us to this incomprehensibly? But how do we how do we deal with that? And and so that book, I wrote it in it was published in twenty sixteen. So kind of before the current like AI craziness moment. And so and in some ways maybe it was a little too early. I'd
I I I anticipated and I put a whole bunch of AI stuff in there. And I think now the the thesis of the book is almost just like completely obvious of like, of course we
have all these systems we don't fully understand. The question becomes though, I I do we just throw our hands up in despair, or are there ways of still trying to actually meet these systems halfway or better understand them? And I and for me, one of the one of the analogies, I mean there's both like the magic analogy, but there's also the analogies from biology.
Because biologists for hundreds of years and they have been confronted with systems they don't fully understand. But it's not to say, we don't fully understand them, therefore we are stuck in this mode of complete and total ignorance. No, we don't fully understand the human body, but we know a lot about it. And we've been making inroads and understanding different subsystems and little bits and pieces here and there. And I think that's increasingly the kind of mindset we're gonna need when we try to interrogate these large complex technologies.
That we ourselves have built, or when we unleash little coding agents or whatever it is, we want to make sure kind of we understand th things that are going on. And I and to be honest, like that, like when if you refuse to kind of have that sort of iterative incremental approach to knowledge, you end up kind of zooming to these two extremes of either awe in the face of the unknown, of like, like I don't know, the mind of the AI is unbelievable or the mind of Google is wonderful, or whatever it is, like this kind of unearned, almost like like religious reverence, or
It's fear in the face of the unknown of like, my God, these things are going to destroy us all. And the kind of this sort of incremental humility of like, okay, we don't fully understand these things, but we can still make inroads. For me, like the the two extremes, they cut off questioning. They just kind of say, like, this is sort of an axiom, like, we're never going understand these things. And you can either respond with despair or awe. And neither one is correct. Like, you need to actually say, maybe we not understand that, we might not understand them.
speaker-0 (27:44.888)
But it's still imperative that we actually try our best. And sometimes we might even use need to use other technologies to understand these systems. And I think that's the kind of thing that is going to be a much more productive response to these systems. And I think importing some of the approaches from biology is actually one way of doing it.
speaker-1 (28:04.146)
And I mean, I I I'm this would have been too late for the book to have captured it because it wasn't really a real field yet. But the kind of mechanistic interpretability vein of of AI research right now in many ways resembles biology. Like in many ways you kind of form these hypotheses about like, well, how could this thing represent information and how could this thing reason? And how could this thing, you know, why is it that we see these effects when we train this way and things like that? And then you go out and test them. And it's the it's
a strange thing for computer people to encounter initially because we're just
It's not like a UFO crash landed with this stuff on it. You know what I mean? We these are engineered artifacts. We made them. But the truth is also we made them, but we also, you know, created them in such an empirical way, basically, created them in a way where it was like people like me sitting behind a keyboard being like, what if we try this? And a large enough population of people doing that that we found a few things that worked. and now we kind of have these things that are almost need to be studied almost like natural artifacts.
And I agree with that that's an interesting sort of observation that there's I hadn't really thought to connect the that the fear response and the awe response both kind of shut off questioning in any productive way, or shut off like incomplete routes to understanding that are still, you know, informative. I think people want to be able to For some people, one of the strange things that I really didn't anticipate about this AI moment is that it became sort of culture war-coded in some way a little bit, like that the sort of not liking AI is sort of.
You know, my is the the NPR tote bag of kind of a certain social set that like proves you are a member in good standing of a certain perspective. And you have to simultaneously view them as a little bit malevolent, but also actually overhyped nonsense that can't actually do anything, that nobody understands anyway. and it it's kind of and and watching, you know, a a group of folks that sort of I'm pretty immersed in, right? I live in San Francisco, right? There's a lot of blue state elite folks running around who who kind of hold these views.
and watching them kind of seemingly not update in the presence of a ton of of changes in capability, but also our understanding of these models and our ability to work on them in a somewhat directed way, has been really odd to watch. and I do think like I I used to be one of these mysterians about neural networks in some ways too, where I was kind of like, well, we don't know how they work, but they do work, so let's go. But there are real inroads. Like I think it's it's it is foolish and misrepresented to sort of say they're completely mysterious to us now.
speaker-1 (30:23.434)
we can make kind of, you know, there are hypotheses like the holographic hypothesis about how information is represented and why high dimensionality does what it does and all these other things that are sort of starting to to really hang together into a coherent body of thought. maybe even engineering, if we get better at it. So yeah, I'm optimistic about the the kind of old fashioned reductionist science. Will it eventually make sense of this in ways that we find satisfying? But in the meantime, you know, going back to linguistics for a second, right?
Yeah, I'm not quite sure if you have a I I don't actually don't know which way you're gonna take this, right? But if you
went to EMN LP in 2012, right? So computer scientists who want to do natural language processing, which is a really old vein of, you know, I we think of it as AI now, but it sort of was its own topic for a long time because AI was kind of poisoned. Middle of the 20th century, you like computers, you like languages. You notice there's this text stuff that digitally serializes almost all of language. Like it's there's some nuances, right? Like
The thing Hollywood actors do, where they're like, you want to buy this from him, you want to buy this from him, you wanna buy this from him. Like there are some nuances that the text loses, but for the most part, it's an almost complete serialization of human language. and it's digital, and computers can work with it. They can store it, they can compute on it, they can count things, you can learn things, you can do all kinds of stuff with it. And yet, for a very, very long time, the kind of I view NLP before the sort of deep learning revolution as being a little bit like communism or something, is like a bunch of interesting ideas.
Attracts a lot of bright minds, a lot of talent, not a lot of working systems, right? Not a lot of sort of, you know, end-to-end working things coming out of it. For the first few years of the deep learning revolution in NLP, sort of old guard at NLPists would be like, well, sure, it's great that you're doing well in this benchmark or whatever, but like this isn't science, right? This isn't answering any of our questions about how language works. It doesn't tell whether it's top down, whether it's bottom up, whether it's and and in some ways I think these models are new that way in that they're a kind of knowledge that
That gives us control and gives us power without really giving us understanding in the sort of science sense. Right? We've so, you know, for a few years here, we've had these models that can basically do anything you want to do with text. You want to translate this book into Chinese? Great, here it is. And it's like not too bad. Might be about as good as a human translator does. how does that work? We don't know. Is there some interlingua? Is there some deep grammar? Is Noam Chomsky right? Is Noam Chomsky wrong? Nobody can tell you. We just have this artifact that can do it.
speaker-1 (32:48.072)
And we seem to be happy with that. I mean, it seems like that's that's good enough for everybody. Like we're happy to kind of maybe go back after the fact and understand it a little bit. Do you ever think, do you ever think there's sort of an actual tension with enlightenment ideals here? Do you know what I mean? Do you ever think that there's a a tension with sort of the the scientific worldview and the idea that we're gonna form theories and those theories are kind of that engineering is downstream of that instead of upstream of that? So yeah, broad broad question.
speaker-0 (33:13.774)
No, this is great. I
I mean, so in terms of the language stuff, I mean, I actually think I mean I I agree with you, like that, yeah, it has deeply changed. Like it used to be, yeah, like these like kind of like very clear, like we're gonna develop a grammar and we're gonna understand these kinds of things. and and now it's just a bunch like a bunch of very sophisticated statistics. And and I actually think in some ways, I'd like to think that the like from the mass of like
The neural network that is doing this machine translation, some sort of statistical descriptions of, I don't know, nouns and verbs and adverbs actually will pop out. mainly because I mean, I'm not trained in linguistics. I I took one linguistics course in college. I found it fascinating. It was super interesting. But oftentimes there's kind of like rules and then there's like a whole stack of exceptions. And and I feel like that as this the exceptions sort of swamp some of the kind of stuff, some of the rules, you need some of these more statistically bound like bounded descriptions. and I think
Machine translation as we have it now works well for that kind of thing. More broadly, as we think about these different kinds of things, whether it's machine translation, or it could even be, I don't know, predicting the weather. we have to think about what are models and systems and simulations for. because so, like machine translation, it works. And and I but I think, and that certainly is an engineering problem.
But in terms of like understanding language, I think that is, I wouldn't necessarily say at cross purposes, but it's a separate question. And then maybe it requires sort of the interrogation using biological tools to understand how these models actually get translation. Like is it the kind of thing where, yeah, maybe under the hood, there are these fuzzy equivalents of various grammatical structures in there? That would be really interesting. Like, is it the kind of thing where no, we we think of language as having a grammar?
speaker-0 (35:14.776)
But in fact, those are just like fun simplifications that just make sense of this crazy disordered thing that is language. That would also be really interesting. And I feel like, and so when we do like so like weather prediction. So like traditionally, weather prediction was these incredibly sophisticated physics based models where you kind of take the entire atmosphere and you break it up into like little cubes or squares, and then you kind of do all these different like atmospheric things of like how air moves or whatever. really, really sophisticated.
And it might be hard like in the end of like saying, All right, how is this thing doing what it's doing? 'Cause there's so many different things interacting. But there is some sort of physics underneath it. And then now increasingly people are showing that you can actually do weather prediction very well as a black box using right using AI. And
speaker-1 (36:01.006)
Case is better, right? For some phenomena seemingly better.
speaker-0 (36:03.958)
And it I mean, and a lot of them are trained on kind of these more traditional things. And so you obviously you can't just throw everything away. But but I think that it forget like for me, like it forgets like you forget then they're like like what is the point of a model? Sometimes a model's purpose is prediction. And that's really powerful. But models are also used for many other things, including understanding. And in that case, then you need to sometimes have a very, very simplified model that
might not actually provide any sort of predictability, maybe a hint of predictability, but can give you some sort of mental hook or window into what's going on. And so like and I mentioned like one of my first, like my early experiences in computing was SimCity. And like I would say like that for me, like the the high point was SimCity 2000. I love it. Like do cities operate the way SimCity or SimCity Thousand does? No. But does it give me an insight into I don't know how systems bite back and unanticipated consequences in the nature of complex systems?
Or can it lead me to further learning about cities and urban design and things like that? Yeah, a lot of urban planners got their start playing SimCity. and so for me, like that model, like not only highly simplistic or simplified, but sometimes like probably very wrong, or it has a whole bunch of like like weird assumptions in it. Like that's not predictive, but it's still really useful. And so so when I think about these kinds of things, I don't necessarily view them as at cross purposes. They're more just like they're serving purpose.
different things. Now in terms of like the Enlightenment and things like that. And then there's kind of like bigger questions of, are we get to the are we going to get to the point where scientific discoveries are going to be made that we don't fully understand because the AI is making these kinds of things. And people have written about this kind of stuff too. Maybe we're increasingly getting to that point. It seems like we're sort of getting there with mathematics, maybe. but in terms of the sciences, I mean, so actually there is so the computer scientist Danny Hillis, he wrote this probably like over like 25 years ago.
He wrote this idea that we're we've kind of moved from the enlightenment, when we kind of apply our rationality to the world around us, to the entanglement where everything's so hopelessly c interconnected we can't fully understand it. In many ways, I mean we've been living in that for a long time, like whether it's like socio-technical systems and things all interconnected. But that does all this weather, yeah. Yeah. And that being said, like going back to like what I was saying earlier, like doesn't necessarily mean that because we can't fully understand everything, that therefore we should have
speaker-1 (38:18.296)
Economy, stock market, yeah.
speaker-0 (38:29.9)
a stance of like we should just give up. And and in in some ways, like because we've had this feeling that we should try to understand everything, I think it's kind of prevented us from recognizing that it's okay to not fully understand every single thing. And so, and if you look pre-Enthumment, like you look at like the Middle Ages, there were, I mean, there were people who like scholars who recognized like there were actually limits to humans to human knowledge. Like you look at, I don't know, the Maimonides, like the philosopher, physician, rabbi,
in his like master work, The Guide to the Perplexed, he actually gives a list of like, here's things the human mind will never understand. And he kind of like says, but like the mind of God will understand or whatever it was. And it's a list, I don't like the like the the number of stars in the heavens and whether or not it's even or odd, or like some other kinds of things. And that being said, we actually know the number of stars that are visible to the naked eye now. Yeah, and I and I believe it's and I believe I believe it's even. I believe it's even. I think. but it's like one of these things, but like he was okay.
speaker-1 (39:20.502)
Are they even or upper? Which one is
speaker-0 (39:28.118)
Saying, like, no, there's limits to human rationality. And I think that's one of the like, yes, we should keep on trying to build tools to help us understand things more and more. And and we should try to apply our rationality. But like the fact that if I like try to make a sentence that has several, like five levels of recursion, and you look at that sentence with like multiple subjunctives or whatever it is, it like barely even looks like English anymore. And like I can't, and like that's just
A few, like a handful of levels of recursion. And like if our brains kind of on a quell before that, it probably means that there's going to be limits to how we think about systems with millions or billions or trillions of interacting components. And I think, and we should certainly try to simplify and understand. And I and I love the approaches of interpretability. But but yeah, there's always going to be this kind of weird messiness in terms of how we think about the world around us.
speaker-1 (40:22.88)
One of the things, and we're getting close to the end here, by the way. And one of the things that I really enjoyed from your from your writing, I actually forget which book this was, but you coined the term mesofact at one point as a thinking in terms of the half life of knowledge, a mesofact is sort of something that is sticking around in the pool of common knowledge that is no longer, strictly speaking, true. In twenty twenty six, are there any mesofacts that that you sort of notice floating around?
speaker-0 (40:48.792)
Ooh, that's fun. Yeah. I mean, I mean, the problem with that though is both most of these are like they're like they change on the order of a human lifetime, like on the order of years or decades or whatever. The way in which we learn that knowledge is often the same as we learn things that don't change. And so, and oftentimes we don't you're not confronted with that change until I don't know, like the next generation. So like, your kid comes home and says, Guess what? Like, I don't know, dinosaurs look completely different or whatever it is. I, in terms of like what we're I mean.
I'm struggling to think of like any specific examples, but I but I do think, I mean, again, things are changing so rapidly, and we are as humans, like not always updating our priors as quickly as we should, or or updating our knowledge as quickly as we should. I think there's gonna be a ton of them. but I would say the the optimistic aspect of that is the fact that like, I mean, even though humans I we we're irrational, we have limits to our ability to understand things, we're also incredibly adaptable.
Like we and it's like one of these things where like we like I remember when when the iPhone first came out, like my my grandfather, so hey, he lived to the age of 99. He read science fiction since like the like the modern dawn of the genre. Like he read Dune when it was serialized in a magazine. So he had seen everything. And then I went with him and my father to the Apple store when they when the iPhone came out. And he looked at it and he's like, This is it. Like this is the object I've been reading about for all these years. And we like so rapidly moved from that of like object from the future to
Like the camera resolution sucks, or why can't it do X, Y, and Z? Like we're really adaptable. So I think we're just, yeah, there's all these changes that are happening, but the fact that we can adapt pretty quickly gives me a little bit of
speaker-1 (42:29.27)
Yeah. If if our minuscule but impassioned audience wants more of Sam Arbisman other than your books, where can they get more of it?
speaker-0 (42:38.254)
So if you go to my website, Arbisman.net, so just my last name.net, there's a link to I I have a newsletter where I I write about some of these kinds of things and I have also a list of yeah, other articles and stuff that I that I've written as well. So I would say those are the those are the best places.
speaker-1 (42:52.46)
All right, fantastic. Sam, thank you so much for joining us. This is a hugely edifying. I I really enjoyed talking with you.
speaker-0 (42:58.166)
Likewise, this was so much fun.