24: Linus Lee - Engineering for Aliveness

28:57I think one interesting avenue of this is there are a lot of tools, you know, there's a handful of really interesting tools out there that try to help researchers understand excellent literature about a topic in like medicine and machine learning. And some of them, they stretch a gamut. Some are more sort of power usury and some are more tell me the question and I'll deliver the result. And I think depending on, and all of those are, whether intentionally or not, expression of some value that the creator of the tool has about what the right level of detail is for the reader or the user of this tool to have to contend with.

29:40A tool may give you the answer, or it may say, here are 20 papers that agree with you and 10 papers that don't, and now deal with this. And depending on who you are, you may want to force the researcher to deal with this conflict in the literature. And so the fact that the spectrum exists, I think, gives people that build tools room to express what they think is the correct level for that particular task for that particular person. Totally. It also sort of seems almost that the instrumental engaged idea is a gradient in and of itself. And that maybe as tools become more adaptable over time, like a good video game, there's a notion I really love about games from this guy, CT Nguyen, where he talks about like video games designers sculpting agency.

30:21And like, If a game's too easy or too hard, like it's almost adapting with you, which I think is important. Yeah. There's a related kind of rant that I've given to some of my friends recently, which is there's this like, and with the other given like level of technology, there is this frontier of like, you know, if you want to build something that is easy to use, you need to sacrifice some level of complexity. And then if you want to make something really accurate and detailed, you may need to sacrifice some like learnability or usability. so there's this frontier but this frontier changes over time and this frontier itself is a thing that you can move yeah so i'll give you a concrete example i was talking about again kind of coding ai with a friend recently and i was so so one kind of advancement that you can make is to build systems that are better at using current technology to build software and so i went on this thought experiment where i said let's imagine that we're in an alternate timeline where we have built super intelligent, super capable coding systems, code generation systems.

31:27But the best programming language technology that was available was C. And so you have this super coding AI. You can tell it, please build Google Chrome. And it'll write a perfect C program in one shot that is Google Chrome. In this world, would you still want to invent Python? Would Python still be useful? And for some people, no. Like if you're just like a normal person going around your day, like not building software, then like it does not matter. But if you're a person whose job is to like think about what software to build and how to build robust, resilient like software systems in the world, then actually Python is a huge advancement because at the same level of complexity of software, it lets you build it much more easily.

32:08And conversely, at the same level of like ease of use of software, it lets you handle much more complexity. And so it's, I think, a notational innovation or maybe an innovation in how we represent and model software. It's an innovation and abstraction that pushes this frontier trade-off between simplicity and complexity forward so that you can... I mean, in a way, it goes back to the seeing and exploring thing. In Python, you can see much better what the system is doing and how it's composed, and you can express your intent much more fluidly without having to worry about things like pointers i love that example it's a little zoomed out but it makes me just like it feels not that far from the notion of like if if we could communicate telepathically would we have needed to invent language and it's kind of silly the first example that comes to mind is there's a there's a sci-fi trilogy called the three body problem and it's a slight spoiler but like the aliens we find out like can't really lie because they communicate more or less like they have no difference between thinking and talking.

33:11Yeah. And so it's really interesting to think about, like, there's just so many ways you can back into the interesting ways that we shape our tools and our tools shape us. Yeah. And language, language and abstraction or core tools. The next thing I want to talk about is maybe I'll start with a few quotes from you about thinking tools. You say, despite its ubiquity, the most interesting and important part of creative knowledge work, the understanding coming up with ideas and exploring options part still mostly takes place in our minds. With paper and screens serving as scratch pads and memory more than true thinking aids, there are very few direct manipulation interfaces to ideas and thoughts themselves, except in specific constrained domains like programming, finance, and statistics, where mathematical statements can be neatly reified into UI elements.

34:00And you go on to say, in the best thinking tools today, we still can't play with thoughts, only words. Finally, While building tools to solve hard problems for humans, we should strive to also improve people's depth of engagement, going back to earlier conversation, with those complex problems and their solutions as a way to preserve human agency when working with increasingly capable aids for our work. Otherwise, we risk losing touch with and therefore understanding over critical decisions. tools for thought is a can of worms as well talk about agency but nonetheless it's it's an area you've spent a lot of time and it's i think an area that a bunch of curious people are very interested in and i think it ties really nicely back to our conversation about representations can you talk about why tools for or rather representations for thinking maybe it's just notation per the thing we were just talking about.

34:56Why is that such a compelling dream? One thing I'm really personally afraid of as we build more technology is, is I'm really afraid that as we build stuff that absolves us of our need to really understand what's going on, that we are going to be pushed to not care about understanding what's going on. Again, this comes back to this like instrumental versus engaged thing, right? But like, I think for all critical systems in the world, like things related to money, things related to health, things related to how we govern ourselves, we should have like really smart people who are who really deeply understand and try to advance the frontier of our understanding of how these systems work.

35:46And I think a core to my belief, maybe just based on intuition, is that agency or ability to understand and influence what's going on is really important. And that to have that level of agency, we need to also have a full and detailed understanding of how these systems work. If you want to write economic policy, you need to have a really deep understanding of the economy as a machine. If you want to write software, you better have a really deep understanding of how computers work at a mechanical level. And so I don't want to be intermediated or automated out of understanding. Maybe more concretely, this like kind of dream of like a UI where I'll describe kind of visually what I imagined when I talk about this.

36:27Like, I think I have a blog post that's called like runtime for structured thought or something. This dream came out of, you know, at some point I was like building a note taking app because it's like what I do. And I had this idea at some point that I could not let go of, which was like, okay, I dump all these like notes into my notes app. And then, but then like, in the process of me actually writing these thoughts down, all the thinking is still just happening in my head. Like if you're doing long division, there is some part of the long division thinking process that is taking place on the piece of paper.

37:01Yeah. Like there is some mechanical thing that's happening with your paper and pen that is like, not a thing that's happening in your head. I don't know if that makes sense, but it's like - It's almost between your head and the paper. Yeah, it's like between your head and paper. There's like, we've like invented a way to externalize some kind of thinking into this like way of writing things down. I think programming can sometimes feel this way at times. Like you've like found a way to like mechanically externalize something that used to be a part of just stuff that happened in your head on paper.

37:26And I really like the feeling of that. I think that's so cool that you can - Writing can do that at a slightly lower, slower latency or it's like on a slight delay. Like it captures a little bit of that, but it's more like, oh, looking what you've already written. Like writing is different than thinking, but there's like a lag or something. Yeah, it's more in that direction. What I really wanted, so the cool thing about long division is that it's like, it's really obvious. It's more obvious than without it when you've like done something wrong or there's certain kinds of like invalid things that you cannot, yeah, yeah.

37:58You can't like do if you're like writing down long division or if you're like, you know, doing math on paper. Same with certain kinds of algebra, but this is really restricted to math. And I wanted some way to like think out loud it on paper where if I wrote something that was obviously logically incorrect I physically would not be able to like write it down like that would be so cool and so like if I if I devil or angel on your shoulder yeah no except the way I wanted to happen is not that I would write like two sentences down and then some chatbot on the computer will say hey I think these are incompatible clippies like no thanks if what I instead want just like like for me to write these things down And then for me, like some, I don't know how this would happen, but it would be so cool if like, as I write my second sentence down, I like run out of room or like, it's like geometrically not possible for me to continue writing the sentence that is incorrect because it's just not compatible.

38:49And I want this kind of fitting the puzzle pieces together feeling for thinking out loud on paper or the way that you write things down is like mechanically helping you think better. And, and that's, that's like an aesthetic thing that I want to proliferate in the world. and so that that that's kind of ultimately what i'm seeing and the and the reason that i got into at some point we might talk about embeddings the reason i got into researching embeddings is is because it felt like it was the closest thing that we have so far to this dream of turning conceptual compatibility and incompatibility and and coherence into like geometric things that can exist in the real world.

39:31When I was in the depth of latent space stuff in 2022, I would tell everyone I could about this idea of like, imagine that in the future, the way that you read a book is not by opening a bunch of pieces of dead trees, but instead you like walk into a room and against all the walls, it's a huge room against the walls are these like sculptures. And each sculpture is like a big idea and different ideas have different shapes that you can tell from a distance. And as you get closer, you see more and more the detail of the shapes and that detail corresponds to the details in the claims. And maybe you in your head have your own sculpture that corresponds to kind of your belief in the world.

40:13And it's like very obvious visually where they're incompatible because they look like things that wouldn't fit together if you like tried to match them together as puzzle pieces or something. This like reification of not just like the words, because the words are just like, they were like what we sound like when they come out of our mouths, but like something structural about thoughts, how can you express it visually? And it felt like embeddings could capture a lot of that where similar ideas are physically closer together and maybe you could turn them into shapes in some way. And there was some research that I want to continue to do over time in that front of like turning the meaning of ideas into shapes and thereby like, yeah, like thinking should be like putting puzzle pieces together.

40:53It's pointing at a notion that I think, I believe, which is that we are fairly underexplored when it comes to spatial interfaces. And I think there might be a lot of heavy lifting that broadly AI might be able to do to help us better investigate them. And there's other things that are going to have to happen. And maybe you need to have a VR experience or whatever, but it almost feels like those two categories of technology are things that might feed off of each other over time. We're like, there might not have been that much work to be, I don't know, your example makes me even think of something that didn't require computing technology for people to sort of come up with like a memory palace, but it sort of felt like we stalled out on like really pushing the boundaries of spatial interfaces for a little while.

41:36And I wonder if that will be something that we return to in the near to medium term, if you can just like generate a world really quickly based on the embedding of a book absolutely i think what happened i think the way that humans interact with information of all kinds has been taken over by writing um i i'm not i'm not very deep in the pure math world but an interesting kind of i don't know discussion tidbit that i heard about was was like there are certain kinds of proofs that you can express as a diagram like if you want to prove that when you have like i'm about to lie i'm literally about to describe something that can't be put into words but in words in a podcast but if you have like a triangle and like a line that's like tangent i don't even know if i can do it but basically there are there are certain kinds of like statements about geometry of like triangles and angles where if you draw a diagram anyone can look at and be like oh it's like so obvious that like this is the case but to like prove it in writing requires a lot of mechanics.

42:43And there's an interesting discussion. Different representations. Different representations. And some are more efficient at certain kinds of claims. But like the culture of, a lot of the culture of pure mathematics, as I understand it, is predicated on written proofs and proofs and prose. And it's a particular kind of written tradition that's biased against this other way of arriving at conclusions. And I think in general, we are so used to the like conflation of rigor of thought and writing lines of text on paper but that doesn't have to be what thinking feels like and what communication feels like and i think actually that the way in which this rubs me wrong of like oh we've like turned everything into reading and writing text is the same thing that i like it's like it's like the same kind of frustration that I feel about everything turning into chat.

43:38Some things are fine as chat. Like some things are the best expressed by like writing things down. But in like writing and in books and in the way that like thinking has come to look like, so much of it is just reading a bunch of text, loading all this complicated state into your brain, and then like doing a bunch of like abstract amorphous stuff and then like writing more ideas out. That, if you have to like turn that into the shape of a software interface. That is what like chatbots are. And instead I want to work with things like charts and tables and plots and diagrams that are just totally other, still very rigorous ways of expressing ideas and working with them.

44:16Or maybe even these like sculptures and things, but that are more direct and feel more like they exist in the universe of like things that the human body is good at working with. Yeah, there's a lot there. I mean, it's, I couldn't help but think while you're talking like it does feel that we're nearing the end of of at least the written tradition being the sort of dominant way of media being consumed and created and we're moving to something that's actually more oral like audio and video but it's experienced through on one hand that's experienced through a two-dimensional screen so there's maybe less of this spatial interaction and then simultaneously as you mentioned we have large language models showing up at maybe the tail end of the text world, society, whatever, grounding us back in text, at least in some part of it.

45:01But yeah, it feels that all of that is at least very flat, which is interesting. Yeah, there's a lot there, as you said. I'll add maybe one more bit onto that, which is, again, coming back to building technology that's a reflection of how you want the world to look like. even in the history of the written tradition technology has further and further constrained what writing looks like obviously in the beginning we had like handwriting for a long time we had handwriting and then we had sort of like what i'll call like more structured handwriting which is like okay you're like writing you're like manually copying down the bible or something But like you have to let you know roughly like you have to put things in lines and there are these like guardrails around the paper and all this stuff But you're still like handwriting you can like if you make a mistake You just like write over it and then now we have and then and then we had typewriters and we had type setting and now we have digital documents And like if you're writing an HTML document, it's really hard to make stuff.

46:04That's just like freeform where if you look back at Really old books. I went to the London National Library Public Library British National Library library it's one one of those things they have an exhibit of old books i totally butchered that but i still love the library where the really cool thing about these books is that they were so expensive to make that they were like heirlooms and like yeah it's almost like a painting yeah exactly but these are also fully handwritten and so there was no such thing as like margins and like lines and line height and spacing on all this stuff you just like wrote stuff where there was space if you look back at books where this was like the way of producing books and like there was no preconception of any of this typographic stuff you just kind of wrote things where there's space and over time we've and and also if you wanted to draw a diagram in the middle of your writing you just like draw a diagram if you want like if you want to emphasize a word just like write it bigger you know and then it's gotten more and more difficult to express to like take advantage of these other richer axes of expressing yourself and so instead you have to like you know italicize in one specific way or bold in one specific way yes and we were kind of like the the richness of writing has calcified into into the set of markup that we have today but again like doesn't have to be that way and it would be cool to like someone should write a actually this is even continuing today markdown is so ascendant i remember when i was in high school markdown was not so everywhere and like most people are on google docs and word and google docs and word you can make it's so easy to make any text any color any shape any font you want and you can like put it anywhere.

47:44Markdown, you can't color your text. And so people, there's like, I think there's like a real cultural shift, again, where like the aesthetics of this like really messy Word doc or Google doc is kind of getting sucked out. And in some settings, that's really nice. But in general, I think that's kind of a loss. Yeah, it's not, I didn't realize that was the way you're going to go. It makes, I'm very pro Markdown, but that's a great, and there are a lot of benefits, but yeah, it's really interesting. We're circling representations. A few more kind of ideas on representations at a super high level and maybe leads to where we're going to go later in the conversation, hanging over so much of how we think about stuff today as LLMs.

48:19You maybe got at this a little bit with the embeddings note. You say language models decouple the way information is stored from the way information must be presented and consumed by humans. What does that imply about how we interact with information? I'll turn the question on you. What does that imply? I think a lot of ink has been spilled around the fact that language models are good at translating one form of information to another. I think that's correct. I think that's also maybe not the most exciting way that I look at LLMs. It is a thing that LMs are great at. Yeah. I think many of the people listening will know that the transformer architecture was originally invented for translation, and that's the ur form of data transformation, at least in the like language modeling realm.

49:09So this like data transformation way of looking at models is fine. The more exciting thing to me is like, again, going back to this embedding stuff, the more exciting thing to me is language models in the process of being trained, they have to somehow internally derive some way of like doing a thing that looks like thinking in some like mechanical way. like i always love physical metaphors and if you like if you imagine a language model the most visceral metaphor that i have for like why a language model my understanding language models are cool is that imagine you have like i i love it if i was like infinitely wealthy and i just had a bunch of side projects i could spend money on one of them that i would do is to build a physical version of gpt2 where imagine you have like a building that is like maybe like i don't know how many layers distributed to like 12 layers they imagine a 12-story building and it's like you know one of those like marble rolling down the tracks things like imagine it's just like one huge root gold machine and you put like a token you put a token at the top that's like a ball with like the token number like 32 or something and it's like rolls down and somehow there are all these like rails carved into the building such that maybe it takes like three days for the ball to roll down but by the time the ball rolls down it like rolls down very neatly into like the next token like there is a the the reason i think language models are cool is because there is a mechanical process that they refer that they like are it's like fundamentally geometry happening inside the model that somehow embodies all of this stuff and obviously it would be really hard to turn it into a physical building but the intuition of like okay there's something like physical and geometric feeling that is much more physical and geometric than like what writing feels like that is happening oh this is so cool yeah there's one kind of data transformation that language models do is like inputs to outputs, which is fine.

51:00And that's sort of like maybe less surprising because that's what the models are trying to do. But a more surprising kind of data transformation they do is they like language models use their first few layers to translate, again, kind of simplified, but they use their first few layers to translate the input tokens into some representation space that is more geometric that the model can contend with. and then spends most of its compute working with ideas in this domain is like shapes. And only the last few layers are used to then like translate that its own kind of internal thought back into the human writing.

51:37Yeah, yeah, yeah. And so this idea that there's this like other way of thinking about stuff that is way more mechanical than writing that exists that we can kind of like do science on and figure out is so cool. Yeah, they're sort of like distilling it into essence space and doing a bunch of work with it in essence space and then like giving us some, there's a, you've talked about so many different, you've built and written about and talked about a lot of different examples of ways we might create new representations and new tools from the prison project to broadly like what would a synthesizer for thought be, I think a video you shared with me early on was this great talk on liquid art.

52:18Like all capturing a lot of these ideas. There's a tool called Loom that you refer to Lume's brilliance is that it lets the creation happen in a different perspective, painting in time first, then filling in the details in space. My phrase for this perspective shift is a perspective transformation because it reminds me of coordinate transformations. Synthesizers, you say, because synthesizers are electronic. Unlike traditional instruments, we can attach arbitrary human interfaces to it. This dramatically expands the design space of how humans can interact with music. All circling this idea. There's one final one you're talking about.

52:52Your friend who is working on Flora, the founder, they told me about a mission statement I always found really inspiring. He wants to allow artists to speak beauty into existence. I love this phrase because all of the focus here is about the precision of the words and the ease with which the words can conjure ideas into being. There has got to be a way to get there by finding better ways to speak of beauty rather than by mechanizing the means of beauty production oh i love that i know i said it but it's i like and like man i mean we could talk about that for a long time but like it's it's really interesting like you're you're you're dancing around actually both the mechanical side of and these like way more mystical like ambiguous almost ineffable side of it my my question might actually be a little lower level which is just like you're describing some kind of cool toys or some cool theories or ideas.

53:49Synthesizers took a while to be taken seriously by musicians. Yeah. Where do you think we are with these types of ideas today? Like what are the seeds that are actually promising beyond like fun blog posts? Yeah. First of all, I'll comment on the, the like better ways to speak of beauty rather than mechanizing the means of beauty production thing. I think it's kind of mystical, but there's like a concrete way to think about this, which is like Python is a better way to speak of software ability. Mechanizing the means of software production would be writing a language model that's really good at writing C.

54:25Wow. And I think we're all better for Python existing in the world. Yeah, elegance doesn't mean ambiguity. Yeah. It's actually more clear, right? It's more useful abstractions of a certain kind. If you're writing a database, Python is too messy. But if you're writing Instagram, Python may be right. And when I think about what advancement has a civilization, maybe this is too grandiose, but when I think about advancement as a society or a civilization, what would really, if an alien species landed on Earth tomorrow, what would really impress me is it would be really cool if it had a language model.

55:04That would be sick. But what would be more cool is if like they had really elegant physical theories or they found some more elegant way to think about music or they have a thing that's like the periodic table, but like even more elegant. Like these things would be really impressive because they are kind of more impressive further compressions of knowledge. Have you read Story of Your Life or Seen Arrival? Yes. A little bit of that in there. A little bit of that. And these things coexist, right? Like mechanizing the means of working with this knowledge is useful, is useful for like scaling these things and automating these things.

55:41But like that by itself feels kind of empty aesthetically. Yeah. And then just seeds of any places where you're starting to see interesting things that could soon be real in this category, broad category. Yeah. I mean, this is what I've spent the last, I guess, two years thinking about. I spent the prison project, which is about figuring out ways to read interesting ideas or concepts or features out of latent spaces of models. That was really fun. There was actually also the first time I ever wrote deep learning code was like this research project, which is kind of wild. Also kind of hilarious project.

56:23Like a lot of its outputs are hilarious. Oh, yeah, yeah. It's so funny. I remember like, like you can do things like you can put in my like bio from my blog and then turn off the coding feature and like turn on the culinary arts feature. And then it's like a bio that's like exactly the same style. I made a hundred dishes. Yeah, exactly. Exactly. So it's super, super fun and interesting. But like, I think the reason that it's worth digging into like why it feels interesting is it's a totally different way to interact with ideas. Now, the challenge that I've spent the last two years with is, okay, you have this kind of cool toy.

56:57It can let you play with ideas in a really fun way. How can I make this more of a thing in the world? And that's partly about packaging the idea. That's also partly about building a thing that's actually useful in some way that people want to use or people want to play with or people need to use because it's valuable for their business. And that has turned out to be the difficult part. I think there are some avenues that are interesting. There's a really cool company called Goodfire that is applying this in a more research domain where they're applying interpretability techniques to very non-language modalities, like models that understand molecular biology and cell biology and other kinds of things.

57:39But that's more in the research realm. I'm still very interested in language. I think the most viable path to production impact that I've arrived at is using this kind of thing to help people make sense of really large data sets where nuance is important. an example of this is like understanding people a big part of venture the business that thrive is engaged in is about understanding people and how they work and where they are and what drives them and data sets about people are really interesting because there's a lot of structure but all of the important signal resides resides outside of that structure like like sure i can get a list of everyone that has ever worked at OpenAI, but like, that's not what makes them interesting.

58:29What makes them interesting is like, like, what game are they playing? Like, why are they interested in this company? What are they looking for? What are they, what, what really excites them? And those things are all between the lines. Yes. Perhaps using some of this kind of looking inside the model stuff, we can learn to work with these data sets, not just at the level of like, where do people work and like, what name are you? What name do you have? But more at the level of like, what seems to be like between these types of things that we're looking for, what seems to be the one that's like driving this person?

59:00And rather than just like throwing it to this black box, that's going to tell you a bunch of names, you can actually roll it out in a sheet of paper and like different motivations and different intensities of motivations can correspond to, for example, different places on this chart. And that there are so many reasons why I think that is much more exciting to me. The obvious This one is just like, it's more visual and more detailed, but also working with people and making judgments about people is a really high stakes thing. And it's really, it feels really important to me from like a principle perspective that the people that are making those judgments are exposed to the full complexity of people that this data models.

59:37People are such a good example because they're the rare category where there's incredible complexity. many i shouldn't say everyone but many people actually have a strong degree of intuitive high resolution thinking about it but they can't always put it into work like there are there are people who are incredibly high emotional intelligence for example yeah and they might even have and you see this in all categories but maybe especially people they have this sort of high dimensional non-verbal vibe or sense or notion um that somebody's special i mean in the words all the words that you're using yeah like yes what do they even what are they holding yeah or i don't know investor firms they'll say that's a person special or good or even smart like what what it was inside yeah or like they might use the word smart in like a really specific way or like spiky use people say crafty or like cracked like what are these words like don't we don't we deserve better vocabulary for talking about these really important things yes but it's also cool because there actually is a huge latent space there that is in certain people's heads.

1:00:44Yeah. And it's like, it hasn't been extracted. Yeah, there is complexity there that I think my, I guess the value that I want to push out into the world is that there is complexity there that we currently have very ill-suited tools for. Yes. That like, I want anyone working with these, you know, data models of people and CRMs and tools like that. I want to, I think the world would be better if we like forced them to contend with the complexity of what what's inside people. But I also want to give them good tools to be able to do that. Totally. One theme that runs across your work that you actually started to maybe allude to when you were talking about warming up to instrumental interfaces is this sort of balance between very philosophically inclined, highly principled thinking and being pragmatic and actually making stuff that's useful.

1:01:33I think you've done that for your own tools. You build for yourself. I'm sure you're doing that professionally, you might even be doing it ideologically. And obviously, that very much applies to understanding when an instrumental tool might be useful versus a more engaged one. I think this is especially relevant when it comes to building with LLMs, which is what you're focused on, both in your personal and professional work. You alluded to it earlier, like LLMs are sort of this magic box, sort of, at least that's how a lot of people feel that's what it feels like to use to a lot of people. And it seems to me and we talked a little bit about this that you are focused on how we can make these LM powered tools more robust and engineered.

1:02:15Sometimes that means making them more instrumental. Sometimes that's with engaged versions of it. But broadly moving away from this sort of mystical ineffable hallucinatory thing to something that's more reliable and predictable. You say natural language interfaces feel like they should be super easy to use, but in practice, they can feel confusing and frustrating because they leave no room for affordances that tell you how exactly to command or control the tool. How do we fix this? Obviously, that's, as you said earlier, specifically referring to chat. I guess at a high level, before we go more into the weeds, why are LMs so different than most of the rest of software tooling and infrastructure?

1:02:54At least when it comes to tactically solving problems. There's a cultural aspect and then there's a technical aspect. the technical aspect of me comes first, which is that this thing, in an academic sense, the thing that's exciting about language models is their generality. But generality is actually a really undesirable property for an interface to have. You want an interface to be intuitive and intuitive means, in a lot of cases, very obvious. Generality gives you power, but there's a reason that Final Cut Pro looks very different than like an IDE. And in some ways, the thing that makes language models as a technology academically appealing and what is pushing the frontier of research is antithetical to the thing that makes really great interfaces for humans to use.

1:03:45And then there's, I think, a cultural aspect, which is this like academic desire for generality and simplicity of interface has bled into people who are building products where there is a similar kind of desire for generality without really considering what end users feel like when they're using it, which is like, okay, you have a box. You were telling me this thing is useful, but I don't know what to ask. You can do anything with it. You can do anything with it. And real physical objects have less of this problem because there are physical constraints. If you're making a hammer, a hammer needs to have a thing that is like the affecting end that is like the thing that hits and then like the handle.

1:04:28If you had a box of tools that was just like a bunch of squares, it would be really, maybe like a terrible box of tools. But I think both the kind of academic culture around machine learning and the way that the technology originated has influenced this. I think it's worth talking about both the instrumental side and the engaged side with regard to LLMs or at least kind of ends of that spectrum. You say the interesting details are in the necessary trade-offs between how well you understand the user's intent and how cheaply, quickly, and reliably you can deliver the result. I'm curious how you think about what ideal instrumental-shaped LLM tools look like today beyond just like an agent.

1:05:11This applies a little bit to what we were talking about before, but like more practically, if the idealized magic button perfect agent is the perfect instrumental LLM tool, Like what is a good enough instrumental-ish LLM tool look like now? This is kind of an oblique answer to that question, but I think gets at what I want to say, which is because the technology is so general, I think it's tempting to make your, as someone building something with language models, because the technology is so general, I think it's tempting to want your solution and your problem framing to also be really general.

1:05:50What I mean by that is if you are using a language model to build something to help people write programs, it's really tempting because the technology is so general to want to preserve that generality, which is maybe equivalent to power in some senses, and say, I want to build a thing that's generally useful for anything in programming. It's hard to resist the temptation, by the way, of the fact that LMs really are really good at being general. Like it has the like, yeah, the Cs or the potential to be all the other like elements of goodness are in there for generality. But it turns out that's just like tactically, it's just really hard tactically to make a product that is so general and good at all those general things.

1:06:35because personally, I think when the times when I've built products that feel the best are when I've been really specific and precise about what the task that the model is doing is. It's kind of funny because this is the, in the more kind of pre-language model, pre-generality pill, classic ML world, the task shape is what you start with. You have a bunch of inputs, input domain, and you have a bunch of outputs. And you try to like, what the model is doing is it's like learning a function that maps the inputs to outputs. And because you're building this thing and you can't train over everything in the entire world, you have to be really precise about what the inputs are, what the exception cases are on those exception cases, what the right behavior is.

1:07:17And you have to really master the task domain. Like you have to have a lot of really fine opinions and there's complexity in the inputs and outputs that you have to have opinions about and deep knowledge of. and I think that principle is still true today. I think if you want to still build good products on top of language models, it really pays off to be an expert in what the right inputs are, what you want the users to type in. You should be opinionated about what the users should type in, what they should want to do with this thing and you should also be opinionated about what the right outputs for those inputs are.

1:07:56You should not just leave it up to the model defaults or whatever OpenAI or Anthropics, Google's post trainers want the model to say. That's the thing that feels like it's missing the most in a lot of AI products today. Yeah. And again, there's this temptation of like, oh, let's just like preserve, try to preserve the generality of the models, which is possible. But it just means there's no true generality. It just means that you are letting the post trainers of whatever model you're using be your product designers, which I think is not the best. One thing that sort of, I'm not sure where it sits on the sort of gradient from instrumental to engaged, but is like this broad frame of AI as a collaborator, which we're increasingly contending with.

1:08:39You have a bit where you're sort of suggesting that working with AIs or collaborators or a set of agents or whatever ends up being might be more like managing people or designing an organization. You say there may be craft to building software at a distance with things like LM agents chugging along beneath your hands. But that feels to me like a distinct kind of joy from working directly in the medium of code, then you go on to say, and just as we shouldn't expect the same people to enjoy both coding and managing all the time, we shouldn't expect people to enjoy doing both coding by themselves and coding with a team of agents the same.

1:09:14And I think that's okay. I think the people like you and me who like the craft of working directly with the system will find ways to accelerate ourselves in our craft. I don't know when you wrote that. I don't think it was super recent. No. Yeah, I'm curious how, like, have you found yourself doing more of both of these sides? Or are you more still kind of in the, like, accelerate your craft? I am definitely. It's funny because I think this was a while ago, as you said, and I think it's actually kind of come true. One of my friends, Dan Tripper, has a company called Every. And somewhat incredulously, all of the engineers at that company only use coding agents.

1:09:53I think it's incredibly rare, if at all, that they open up an ID and write code, which I am somewhat skeptical of, but I believe Dan. But me personally, I really enjoy understanding and having opinions about how our architect systems. And I like looking at how things are organized and how things work under the hood and being in the weeds. And so I've, I think, gotten better. And I've still accelerated myself. Like I'm way faster at building things than I used to be, but just in a different way. And I think it's a different kind of joy as I wrote that. And I think both things will hopefully continue to be accelerated.

1:10:31At the end of the day, I do think it's important that like software is one of the most. I mean, not even one of, I think software is probably the most complex kind of machine that humans can ever make just because you can pack so much complexity into such a small amount of space. And so I think it's important that as we build more software and more critical infrastructure software that humans are forced to contend with the full complexity of the systems we're building. And I think there will be, it'll be important for that some people, I think that's good that some people will continue to get joy out of like working with that complexity, including myself.

1:11:10A few bits that I think tie very closely to that and the idea of accelerating craft, all you when possible directly manipulating the underlying information or objects of concern the domain objects minimizes cognitive load and learning curve and then separate quote what direct manipulation is to the graphical user interface we have yet to uncover for this new way to work with information then you go on to talk about good complexity you say a second type of surrogate object is focused not on showing individual attributes but on revealing intermediate it states that otherwise wouldn't have been amenable to direct manipulation because they weren't concrete.

1:11:47Indeed, I think it's fair to say that direct manipulation is itself merely a means to achieve this more fundamental goal. Let the user easily iterate and explore possibilities, which leads to better decisions. All of this to me, I think is like around, we talked about a little bit with the people stuff and the embedding stuff. I think you call a data set a bag of concepts which is which is great um this proximity and latent space another framing or metaphor you use is like a brain what if we could do a brain scan of an lm yeah i'm curious what your early conclusions are on like maybe going back to the earlier one of those quotes is about like we have the the gui most people today even engineers are either like using it to generate some amount of code, tab autocomplete, whatever, full agents, like get answers, like the broad shape of the use case for these things doesn't seem to capture a lot of this, like exploring latent space idea.

1:12:48Yes. I'm trying to figure out what the right question is. I suppose it's part why is that, part do you have any hunches of where we're going? Are you doing stuff? We've talked around some of this, so I don't want to be too repetitive, but yeah, I'm curious where you sense we are on this. I feel this really personally. I've spent so much time thinking about how to work with ideas in this more like geometric format. And I'm so excited by it. And yet, like, even I'm like building chatbots, you know, I, I, so like, at times it feels like kind of very deep cutting personal hypocrisy. At times it feels like I just maybe have it, like, maybe it's like a skill issue and I just like haven't been good enough at like injecting some of these ideas and research land into real products.

1:13:33I just think it'll come. Phil Wadler, who's a researcher who contributed to Haskell, has this great quote about doing computer science research. It's actually specifically about building a great programming language, which is, I'm going to butcher the details, but it's something to the effect of, if you want to build a great programming language, first do some research, invent a language, and then wait 40 years for the ideas to mature. I think surely a part of what I need is more patience. And I think there are ways for ideas to collide into the right ingredients for it to become successful over time.

1:14:14I also think it is still what I'm working on, which is to find like for a while I was just doing like in my head exploration. generation and then for some time you know like three years ago two years ago i was working in a very horizontal company notion and trying to find ways to make useful things with my ideas then now i'm in a very vertical very concrete set of use cases and still trying to find ways this can plug in be useful and i think i've gotten closer like the the understanding people thing i think is concrete progress in where i think this can be applied but uh i'm even more excited to actually build things in this direction and then see what breaks and ultimately that's how ideas get out into the world that's a great transition um you work at thrive um if that weren't obvious you previously spent some time at notion which you just alluded to is sort of being a little bit more wide and abstracted the thrive works more deeply kind of embedded part of that i think implicitly is just like thrive has less than 100 people most of whom you're with in a room or at least in a building you I think you frame this somewhere as like a much smaller TAM for the things you're building yeah like 80 people yeah I've spoken to every single one of my potential users it's amazing super high level like what what drew you here and what has that shift in in this sort of scope of of your work what has that been like yeah so maybe important to know that at the top given the engineering that we do at Thrive is not super visible is we have a team the team is five or six of us, including a designer and many people who are great product engineers and data folks.

1:15:56I'm a part of that team for most of my work. And the team existed a while before I came, but I came in and I think kickstarted a lot of the more LLM-related parts of what we're building now. The really interesting thing that I've learned about building internal tools is that when you're building anything i think you have a kind of a fixed budget for surprise or uncertainty and when you're building at least my kind of brief experience in the world when you're building a product to go have its own life in the world and be used by hundreds of millions of people you spend a lot of that uncertainty on just like the problem like the problem different people are going to bring different problems in different use cases and contexts of use to what you build and you have to prepare yourself all of those things so it's like the problem is a theory yeah or is this like broader zoomed out abstract thing that has to hold a whole bunch of potential people's needs and because the uncertainties in the market and the users you want to i think be pretty stable and predictable and kind of mostly what you build with where in internal products at least in my experience at thrive there's a lot more like the problems are so clear they're also kind of i think just barely outside the reaches of boring technologies that there is a lot more room that i found to throw weirder ideas or more frontier ideas against those problems and see what sticks and see what works and so that's been that's been really fun you can go as specific or not as you would like and you briefly alluded to it but i think it would be helpful to hear a little bit more about like what you actually work on yes and then And as a second piece, I think you also kind of implied this, but like you aren't exactly the typical profile of like a somebody, a VC firm hires to like work on internal tooling.

1:17:46I think like that match on both sides a little unique. And I sort of like a meta thing hanging over all of this, it would be what does it mean for a investment firm to take building software very seriously? Yeah. Yeah, so we, it's kind of funny. I find it always really hard to, I had this problem at Notion and I have this problem again at Thrive. It's really hard to explain what the thing actually is. Like if you work at a calendar app company, you're like, I'm building a calendar. Notion is kind of a little bit of everything. So it's like, you can say, some people call it a note-taking app, but it's really like, like the thing that people say internally is like, oh, it's like Lego bricks for software, but then like everybody's like, you know, pretentious description.

1:18:31here there's a different kind of the same problem which is the thing that we're building is kind of a suite of different tools that together help a bunch of different teams inside thrive work and so it's not really a calendar it's not really a note-taking app it's not really a chatbot it's kind of an amalgamation of all of these things to the extent that they're useful for different teams the mandate of a team is to build a kind of iron man suit for everyone inside thrive and also to make Thrive's business itself much more software-defined. And there are a few different layers to this. One layer is just to kind of organize the information and data we have about the world and about people and companies and investors around Thrive to be more accessible.

1:19:13The next layer is a kind of core product that includes a thing that looks like a calendar, some kind of a newsfeed, a search product, a research agent tool, all of these things combined to a single coherent interface that helps people find information quickly, prepare for their day more efficiently, know what's going on, be more prepared in their job and spend less time working on drudgery. And then there's a third pillar, which is more embedded, which is we have this foundation of data and automations, but there are really specific workflows where if we just like pick get that specific workflow and fully or mostly automate it, it's going to save someone tons and tons of time.

1:19:53And maybe only two people of our 80 people are going to use it, but it's going to save them so much time that it's worth it. And so we build those kinds of one-off automations as well. The way that different people use it differently, which is what's cool about building for 80 people, is that you can kind of build for every single individual person. But the way I use it, I use it a lot to just like look up and understand different companies that we're partnered with and involved with and invested in different people around our universe. also just generally to like our research agent's actually quite good and so i use it sometimes as a replacement for like o3 um because we have some proprietary data and also to just organize my day and like prep for meetings and things like that so that's what we build yeah very very classic internal products but a lot of the way they would like broken down the problem is as like a data layer and intelligence layer and then like a ui on top and then automations on top which i think is a really clean separation of concerns.

1:20:48And then your second question. Yeah, I mean, I think inside my second question would be like, most investment firms, certainly VC firms these days, have internal tools. That's true. You take building software, you personally, I know, take building software really seriously. And so my assumption here is that that is, to some degree, to attract someone like you, that is internalized here. And so, yeah, the broad question would be like, what does it mean for particularly an investment firm to take. Facebook builds really nice internal tools too, but what does it mean for an investment firm to take building software seriously?

1:21:24I think there are things that make it personally really interesting to me as a place for me to continue to validate some of my greater ideas about information tools. And then there's also, I think, ways in which this is just a really cool opportunity for Thrive and for any investment firm. For me personally, So I've already spoken a little bit about how internal tools give you, in my experience, a little bit more flexibility and the methods that you apply to solve hard problems because the problems are so fixed. And because internal tool teams tend to be smaller and a little bit more exploratory and agile.

1:22:00I also think the kinds of intellectual work that a lot of people do inside at least Thrive, because this is really the only firm that I know with this level of familiarity, is deeply qualitative. It's not just moving numbers around and optimizing numbers and sheets. It's like you have to understand people. You have to understand products and problems. You have to understand research. You have to understand news and events. And these are all really qualitative things. And at the same time, there are clear signals for when decisions are correct. There are also in place of like raw signals of successful decisions.

1:22:41There's also lots of like really high quality proxies. And I thought personally, it was a really good place to continue to build like pretty high stakes tools to help people work on pretty deep intellectual problems and knowledge bound problems and but then be able to flexibly explore inside that i think more generally i've been noodling on this idea of of for like fundamentally venture is a services business it's a financial services business and i think right now is a really interesting time for services businesses because when you're trying to get really good at a service there are two things that are really scarce especially in an ai world one is really really detailed understanding of what it means to be really good at that task.

1:23:27Like, what does it really mean to get really good at investing or get really good at researching a market or get really good at legal contract review or whatever? There's a lot of detail in there. A lot of detail is not like written about on the web. You just have to kind of like do the job and learn. And then there's also, so that's one thing that's scarce. The other thing that's scarce is like a sandbox environment or even like a real environment for you to then make decisions and learn, maybe not through mistakes, but learn through feedback. This is not just true of financial services. It's true of like, if you're like trying to run like an AI managed software building agency, the two things that are scarce are like, how do you do the job of like getting a client and writing the software and delivering it really well.

1:24:10And then also just this, this playground or environment for, for writing software and then getting feedback from, from clients or from, from compilers or what have you, the environment. And it feels like a really interesting technical problem given where I is at now. You're also, sorry to interrupt you, but you're always, both of the examples you gave, in part because they're generalized, you're in this constant state of, you have this frame of like scale X or explore. And you're in this constant sort of that type of situation because the next incremental thing might always be new. If you are a consulting or an investor or something like that.

1:24:48Whereas inherently big software companies have new things, but they also are like really they're just really good at like mechanical reproduction in some sense which is interesting yeah yeah i think at the end of the day it's an environment that's like very rich with really intellectually demanding tasks with a lot of detail where if you built some software that either was really good at the task or helped a human get really good and efficient at the task both those things were really interesting innovations for me yeah so um for it to work on it. I also think more concretely, in general, Thrive is good.

1:25:22It's filled with really thoughtful people. And from my experience, at least in the last year so far, we've been able to build something that's really functional and valuable, but also in a really high craft way. Like one of my favorite things about our app is that when you, it's one of these apps designs where like there's a ribbon of like four tabs at the bottom. And when you tap the tab buttons, there's a little like haptic vibration. And it's just so nice. And I love that I can work on things like that. Little animations. We've done a lot of work to make our chat interface itself really, really robust errors.

1:25:52And I like that there's space afforded to, even if it's only for like 100 people or 80 people, to make something that is just objectively a great piece of software. And that applies to the way that we build its language models as well. Yeah. There's a sort of broader thing that your earlier comment about the Iron Man suit made me think of, which is one of the more unique things about an investment firm is that investing in many ways is sort of like the most generalized, job. Like you sort of definitionally are a generalist. And as a result, you're sort of building tools for this like group of high agency, fast learning, fast moving generalists, which is kind of just a broadly interesting design problem.

1:26:40I don't know if I fully agree with that. I think maybe that, especially at a generalist fund, I think maybe the topics that we want to research and understand are general. Like one day we'll be looking to understand batteries and the next day we're looking at fashion. But the skills involved are not super general. I think the skills involved, I mean, I'm not an investor, so maybe, you know, but maybe I don't have as much detail as someone who would be. but it feels to me like the core skills are around relationship building, things that kind of resemble sales, rapidly understanding and researching new topics, kind of decision-making under duress, maybe.

1:27:23You're describing something that sounds pretty generalist to me, but... I don't know. I think, like, if you really think about, like, what it takes to be good at, like, programming... Mm. Ah, this is a good counter. ...the skills. It'll also be, like, understanding really complex systems and being able to like describe that complexity in words really well. And obviously there's a concrete, I don't know, like the concrete skills in programming are like, you got to like use an IDE. In investing, you got to like use Excel. I don't know. These are like, there are these super concrete skills, but I feel like they're a little bit too in the weeds.

1:27:52And I think the way that we describe skills, they all kind of can sound general. But I do think there are aspects of it that are like, I don't think investing is fundamentally more generalizable than many other skills. But I do think it's very, like, qualitative and, like, rich in detail. It makes it an interesting kind of task. On that last note, how do you think about bringing an engineering orientation to any kind of work or this kind of work specifically? one way that I think I've grown as an engineer in the last year which is how I'll answer this question more directly later is I think when I was really young I viewed engineering as like okay my job is to write some code that does the solving of the problem and so the end artifact the deliverable if you will is like the source code I grew a little bit a little more mature a little wiser.

1:28:47And then my perspective was actually, my job is not just to write the source code, but to deliver a running system. The system is a thing that's derived from the source code, but it has to be reliable. There are certain kinds of operational metrics that you have to hit. It has to be understandable and debuggable. If something goes wrong, you need to be able to root cause it and fix it. If you want to make a change to it, you need to be able to make that change. You have to be able to make that change and be confident that it's going to be good, and so on and so forth. So there's like second layer, which is the operational aspect to the software.

1:29:19And then more recently, I've been thinking about, like the first two are really, like they make sense in the setting of like an individual coder. Like as a solo developer, you have to write the source code and then you have to like maintain and run the source code, the software. And then there's third layer, which is especially important if you're running, if you're working in a team, where the third layer they have to deliver is to build like a people system or an organization that is capable of continuing to make changes to that software and evolve it as the world evolves. And in some ways, the thing that you have to deliver is not just the software, but like the processes and the culture and the team itself, such that if you were removed from that equation, this whole system would be self-sustaining and resilient and robust and would continue to deliver this like second tier thing of like really resilient, operationally excellent piece of software and also the first thing that is really good source code.

1:30:18Yes. And so building these systems, the source code, the running system and then the team, all three systems in a way that is easy to change and understandable and debuggable is I think what good engineering is, at least the way I understand it at this point in my life. And that's kind of the concrete aspect. I think more specifically and maybe more interestingly in my specific job right now and building a tool for Thrive. In the near term, I think a lot of what we're building is this like Iron Man suit idea of there's a ton of things that people do inside the firm that we want to make easier and give them more leverage.

1:31:00In the long term, I'm also really excited about are there parts of Thrive's business that we can fully turn into like a scaling equation where we like spend an incremental dollar of compute and get some incremental predictable like amount to return out. Like it would be really exciting if we had some machine, some black box where you could put in like X thousand extra dollars and like for every X thousand extra dollars you put into compute, you get like one new extra interesting founder that we discover in the world. I don't know how you would do that yet, But like that is a very, that's like a fully software defined version of the business versus like a thing where you have people.

1:31:46And in that world, you would still need the people and the judgment of the people. Yes. And so it's not about automating, but it's about like in order to build that thing, you're going to have to have really deep understanding of people and of the job and of the decisions. And so I think it gives, especially as we want to like stay as a small team, as we grow the business, we have to like lift everyone up to work at a slightly higher level of abstraction. And as we lift those people up, there's going to be perhaps this like underlying engine. That's like an embodiment of what makes people really good at their current job.

1:32:20Also, just another cool example of the ways that like better modeling latent space might produce really interesting outcomes. In order to build systems that are really good at these very nuanced jobs, you're going to need to give the people that are operating these tools the ability to express their nuance, which goes back to what makes a good tool. Yes. Has working in an investment firm broadly or Thrive specifically made you more commercial in any ways or even just like learnings about that side of the world? I don't think it's a consequence of working at an investment firm specifically. I'm sure these are all related, but I don't think I'm more commercial now than I used to be.

1:33:00I think part of it is the growing pains of having an idea that I was really enticed by, but then not being able to find great use cases for it. Yeah. Yeah. And a part of it also is like, I think Thrive in particular being very generalist and doing everything from like fashion to, I don't know, yeah, compute. So many of the things that humans are engaged in have like nothing to do with software or technology. I mean, everything is technology in some ways, but nothing to do with like software and AI or even like automation machines. Like software is such a small part of the human experience. And there are a lot of other kinds of businesses out there.

1:33:45And I think that has been, that's like a continual learning that I find really fun. I want to spend our last few minutes on an idea that I think has actually been inside a lot of what we talked about and is expressed in that final answer you gave a bit, certainly expressed in the tools you build for yourselves, yourself, excuse me, including stuff. I mean, stuff you said, all great tools must be built in a serious context of use. Good tools are transparent. They let ideas through. but zooming way out, talking about kind of technology and humanity and how they interact with each other. And in this section, I'm gonna have a bunch of quotes for you.

1:34:22So you'll have to forgive me. The first is from an essay he wrote called create things that come alive. You say building technology is fundamentally an affair by humans for other humans and objects of technology ought to be ensconced in a romance and history and all manners of color and details and textures of life. It ought to come alive in our environment. Technology is not what's shiny and boxy and delivered in metallic wraps. And then you quote Ursula Le Guin, technology is the active human interface with the material world. From a separate essay you wrote on Radio City, you say, I yearn to see this more irreverent and humanist relationship to technology imagined more often today, when both technology and the industry backing its progress feel increasingly detached from culture and media and the humanities.

1:35:12I dream of a humanist revival with computing systems at the center, grasped firmly in the hands of wisdom. And then one final pair of quotes where you're kind of pointing at this notion that technology is often thought of as like other separate alien thing that's separate from our humanity. You say technology exists woven into the physics and politics and romance of the world and to disentangle it is to suck the life out of it, to sterilize it to the point of exterminating its reason for existence, to condemn it to another piece of junk. If you consider yourself a technologist, here is your imperative.

1:35:49Build things that are unabashedly, beautifully tangled into all else in life, people and relationships, politics, emotion and pain, understanding or the lack thereof, being alone, being together, their homesickness, adventure, victory, loss. Build things that come alive and drag everything they touch into the realm of the living. And once in a while, if you are so lucky, may you create not just technology, but art, not only giving us life, but elevating us beyond. I kind of just honestly wanted to read all those quotes are amazing. My question is, how do you personally imbue the technology you build with this aliveness, with this humanity?

1:36:32wow what a question um so earlier i mentioned that one way that i've grown is that i've gone from working on the stuff because i was just really excited by tools and productivity tools and the aesthetic of it to thinking of it in the context of trying to give people more power and agency over their life and over the world i think a similar parallel maybe like growth arc that I've had is if I actually think about why I like working on this stuff a part of it definitely is that I like the puzzle of like writing programs and it feels like I'm good at it and I want to continue doing it and be better at it because mastery itself is fun but so much of it also is like everything about the context in which I do this work like I find the like environment of people building companies and then thrive really fun.

1:37:29I love the people that I get to work with and get opinions and feedback from and tell them, it's like fun to tell them about ideas that I'm thinking about. And like all of this exists within the context of like, I'm a human being, there are other human beings. I'm like building stuff for them or building with them or having them disagree in interesting ways. And the context of doing this work is what's fun. and I'm sure I'll over time I'll like learn even more and lean even more in this direction and going back to that where we started this conversation I this I think is really adjacent to this other kind of aesthetic thing that I feel of the point of technology is to like elevate us to give us more prosperity, to let individuals do more than just survive.

1:38:21In some ways, not in some ways, I think most of the time, for most people, technology itself is instrumental. There is obviously fun in building technology. It's fun to write programs and solve these puzzles. But in a societal scale, technology itself is an instrumental tool. and i think it's important not to forget that and the the if it's instrumental there has to be something that is the result that we want like what is the prompt and i think the most inspiring prompt whose response is technology writ large is everything that makes it really great and fun and lovely to be a human being how can we have more of it and how can everyone have more of it and how can we have it and all of the variations of it that exist in the world, that is the prompt.

1:39:11And then the answer is everything the technology has unraveled into. And if we stray away from that, if we develop some other proxy, some other benchmark that we're like optimizing technology towards, that is, that's not good. And so I think people building with technology you need to remember that these are all for for other human beings and and building with other human beings another few quotes on a related theme this is a recent tweet you say more people should create things to proliferate and aesthetic into our future not just to solve problems this is the quality that every artist and engineer i respect shares most universally without this You are doomed to churning out slop.

1:40:01What values do you create to spread? What image do you dream about? What is the feeling of a tomorrow you want to give form to? Have a position, stand for something, don't just create value. Then you're a separate idea about dreams. You say too many tools for thinking, not enough tools for dreaming. But really, aren't some of our best ideas found in dreams? And then finally, you're riffing on this idea of neural media. And you say, could a generative image model, never having seen Voyager 1's pale blue dot, have been used to create such beauty? Can a neural generative model imagine beyond its own world model?

1:40:41How can we build models that can help imagine new worlds, not just permutations of the one we know? Imagery is this pattern or this theme across all of those ideas. um aesthetics but maybe even more so dreams dreams are so image based at least in my experience how does imagery underpin your creativity your work and i realize you're sort of talking about imagery in the literal sense and maybe in this sort of broader sense too yeah i do think i do think in a lot of these cases imagery and dream are interchangeable or at least deeply entangled and and dreams are maybe the more fundamental thing although imagery is the way in which it's it comes across.

1:41:27There's a thing that I feel like a lot of what I read around me is engaged in these days, which I promise this is a roundabout way to answer the question. There's a thing that a lot of what I read around me is engaged in, I feel like these days, which is to pretend that technology has its own arc and its own destiny, and it's going to like go somewhere by itself. And it's like there's some like manifest destiny version of technology that is where like there is some final form that we're trying to achieve. And going back to like humans build technology for other humans thing. Sure. I understand the reason that this is sometimes true.

1:42:06Like there are kind of really powerful structural forces that let certain technology really proliferate really easily and sort of have more evolutionary pressure and tailwind against it. And so in some sense it's true, but I also think this is often used as a way for people to either not be intentional about the direction in which they are building the technology or to sort of forget that they have the agency and the responsibility to try to push against it or try to shape the direction in small ways at all. Yeah, it goes back to the first thing you said at the top of the conversation. Yeah. Again, there is a default path for technology to roll down the hill and And it's still really hard to predict.

1:42:50Like this like kind of fetishization of prediction is maybe most obvious in this like question that you get in San Francisco all the time. Like, what is your timeline? Or is it if there's some like correct answer to this question? And my kind of pee-pee response to this always is like, what do you mean the timeline? Like you're the ones that are like building it. Like it's sort of like if you ask me like, oh, like let's make a prediction. Like when you're going to wake up tomorrow? Like, okay, I understand that there is the spirit behind the question, which is like, okay, there's going to be some probabilistic distribution over like when I'm likely to wake up.

1:43:23Like I'm probably going to more likely wake up around like 8 or 9 than like 5 a.m. Sure. There are some likelihood distribution that we're talking about. But also this is a thing where like if I really wanted to, I could wake up whatever, you know. We have will. There is some will. And not only do we have some will in the case of maybe, you know, waking up, when do I wake up tomorrow is an inane question. But like in the in the case of building technology, I think people should be building technology intentionally to express who they want to be and who they want, what they want the world to look like.

1:43:58and again there's there's a default course but i think the fun in technology is to like use it to try to deviate the world from the default course in whatever way you think is interesting and useful and good for you and good for the people around you and so it has a course but but like the course is ultimately the sum vector of like every every little person that's like pushing on technology in different ways. One of the other ways in which I think I've grown over time is it used to be that the only way that I would influence where technology was going around me was by like building stuff. And then I think over time, I've ended up in this really privileged position where a bunch of other people care what I think about stuff.

1:44:45And so now I have this like other lever to push on technology, which is I can build technology on my own, or I can like try to convince lots of other people that there are other directions for technology to go. in and like get their help in pushing things where I want to go, which is kind of, I guess what I'm doing now. But there are all these different ways that you can, you can tug or nudge technology to go in the direction that you want to see the world look more like. And if you're building technology out of this power and you should take advantage of it, it's incumbent upon you to take advantage of it.

1:45:15I think aesthetics and dreams and all these things too are highly rational, ambitious, smart whatever people can maybe underrate their um their weight or their persuasion or their impact on on what we do too which is cool i think also this is maybe i'm just a romantic but it's like kind of the point of it all like when you're like enjoying a really delicious meal there's like the there's like the functional point which is to like sustain yourself and then there's like the taste and i think it's totally valid to be like i'm eating this thing because it like tastes amazing and and uh sure you can like make whatever meal but but like i think it's much better to like think about trying to make something that tastes really good and the the like taste of the meal in my head is like the the value and the principles behind what you make i have just one final question but i do have some more quotes before we get there uh this is you in these visions i fell in love with the idea that there was some quality other than truthiness that ought to guide our search for knowing more about the universe and about living.

1:46:26This ineffable quality I've come to call by many names. Among them are words like novelty, surprise, and wonder. Also you. Productivity is about the industrialization of creation. Wonder, in contrast, defies systemization because it gets its power from uncertainty and surprise by nature. You can't optimize wonder because to optimize requires knowing the output and the process. Wonder is the discovery of new outputs and new ways of getting there. And one final quote from the beginning of a book I love by Lawrence Wexler is the biography of Robert Irwin. I'm in the middle of this one right now. Amazing.

1:47:07He opens the book with an anecdote about Irwin. During the early 60s, when Robert Irwin was on the road a lot, visiting art schools and chatting with students, he was preferred an honorary doctorate by the San Francisco Art Institute. The school's graduation ceremony that year took place in an outdoor courtyard on a sunny, breezy afternoon, sparkling clear. Irwin approached the podium and began, I wasn't going to accept this degree, except it occurred to me that unless I did, I wasn't going to be able to say that. He paused, waiting as the mild laughter eddied. All I want to say, he continued, is that the wonder is still there, whereupon he simply walked away.

1:47:46My final question is, what does it feel like to be lost to wonder? I think at the root of everything that I find really fun and invigorating is some useful, substantive sense of novelty, which is, I think, if I had to really clinically analyze it, what wonder feels like to me. This is a part of solving hard programming problems. This is a part of trying to come up with new abstractions. A part of spending time with new people is like behind all of those things, there's something new that I feel like I could understand if I just put more effort into it and spent more time with it. And there's something like fundamentally very satisfying about coming upon a thing that feels new and mystical and then kind of figuring out how to model for how to understand it and how to model it.

1:48:40And I feel like times in my life where I feel like I'm having, making the most of my life are when I'm in a repeated cycle or a flow of coming up on something new, understanding it, coming up on something new, understanding it. And to do that continually over and over is to be lost in wonder. That's all I got. Linus, thank you. Thank you very much.