Imbue: AI Agents That Can Reason with CEO Kanjun Qiu

Full episode

Matt Turck [0:56] Hi, Kanjun. Welcome to The MAD Podcast.

Kanjun Qiu [0:58] Thank you.

Matt Turck [1:34] So to set it up, you are the CEO of Imbue, formerly known as Generally Intelligent, which is an independent AI research company building AI systems that can reason and code. And we'll talk about what that means extensively in a minute. And for the people like me who care about those things, you just announced a large $200 million Series B round at over a billion-dollar valuation. So congratulations on this. And we'll talk about this as well. But maybe to start, as I was researching this and following your work from afar over the years, you are a woman of many talents.

Matt Turck [2:01] You're a serial entrepreneur, but also a VC. You also have a podcast, which, by the way, is fun to be doing a podcast with a podcaster. It's a little bit of pressure. We'll see where that—

Kanjun Qiu [2:02] No pressure.

Matt Turck [2:13] So let's talk about your journey and, most importantly, the founding motivation behind Imbue.

Kanjun Qiu [2:38] Yeah. I guess the quick background is I went to MIT, and actually, out of MIT is where I started really interacting with computers that process data effectively. I paid for MIT by trading high-frequency trading algorithms, so by designing high-frequency trading algorithms. And that was really an effective way for me to learn: how do you actually make machines that are intelligent and can make decisions? And I think that got me into really thinking seriously about what has the computer done for us?

Kanjun Qiu [3:08] What is it? And I dug into the history of computing. I think I kind of have this view that humanity, as a species, we've built tools over time that allow us to become more human. In the 19th century, we built a lot of machines, post-steam engine, that help us do physical labor. And I think as a woman, if I were born in 1900, my life would not be as good as it is today. I'd be spending most of my time doing physical labor.

Kanjun Qiu [3:40] And today, we have computers. But computers have only started to take over a little bit of the mental labor that we find tedious. And I think there's a future where, just like we invented things like the stove or the dishwasher, refrigerator, the train, the automobile, all of these things that helped us do physical labor and go really far, there's a world in which we have computers that can help us accomplish much, much bigger goals. So that they can take away a lot of the tedium, like paying bills or making the fine details of the spreadsheet when I really just want to know one number, or all of these things that we have to do today.

Kanjun Qiu [4:11] We don't even realize that our computers have to be micromanaged. So from MIT, I went to Dropbox. I became the chief of staff. I started a company that ended up having a nice exit. I started a different company called Sourceress, which was a YC company, and we were building an AI recruiter. And that's where we started thinking quite seriously about what it would actually look like for a system that is able to take over much larger goals and what's required for that.

Kanjun Qiu [4:51] And so after Sourceress, we really felt deeply—this was in early 2020, and our housemates were working on GPT-3 actually at that time, and maybe in late 2018. And so we'd been kind of thinking about, hmm, seems like self-supervised learning works well on language. And in early 2020, SimCLR came out. For the first time, self-supervised learning worked well on images. And self-supervised learning is interesting because it's how humans learn. Most of the time we go around, we Google stuff, we learn on our own.

Kanjun Qiu [5:22] We're not giving supervised labels all the time. And so we felt this is actually an important, kind of what feels like a historic thing. This is a technology that might actually be able to, by itself, process the world and develop something that is kind of representative of an understanding of the world. And that's going to unlock so much. So that was the initial impetus for what was at the time Generally Intelligent, now Imbue.

Matt Turck [5:44] Okay, great. So we're gonna go into the details of what it is that you're building and all the things, but maybe before we jump into this, a quick word on the company itself. So you started in 2020, you just raised a couple of rounds. And so what's the company makeup? Is it like everybody's a researcher kind of thing?

Kanjun Qiu [6:11] So the company is mostly engineering. I think when people think about AI, they're like, oh, it's research, but really 95% of the work that we all do is engineering work. And I actually think the engineering is what drives the research in a lot of ways. We build a lot of tools for ourselves, really good infrastructure for running experiments and for being able to replicate our own experiments very carefully. And that infrastructure is what lets us build models.

Kanjun Qiu [6:41] So, at Imbue, what we do is we train large foundation models, and those models are optimized for reasoning. And the reason for this is because what we're interested in is AI systems that can help us do much bigger things, accomplish much bigger goals. And what we have today is we have these quite interesting, powerful systems that are generative, that I can give some text and it'll give me back a better-written version of that text. But it's kind of like I give it a prompt and it dumps something back to me, and now I have to figure out what to do with it.

Kanjun Qiu [7:13] It's on me. And that doesn't have to be the case forever. And it's kind of like a computer I need to micromanage right now, like an intern. And there's a world where I don't have to micromanage all my computers. So the question is, what is the difference between today's world where I have to micromanage my computer and I'm glued to my screen all day because I have to micromanage it, and the world where actually I can give my computer something, trust that it'll do a good job, have visibility into what's going on, and be able to go off and do other things that are much more interesting and that are much bigger?

Kanjun Qiu [7:53] And for us, our belief is the difference, the distance between where we are today and that future, is kind of distilled into a core set of things that we call reasoning. So reasoning involves skills like knowing when should I ask questions. Let's say you're trying to book a flight. I would not trust my AI agent to book a flight today because how do I know? Are the flights real? Has it analyzed the flights around my preferences? Does it even know what my preferences are?

Kanjun Qiu [8:13] Did it find the best-value flight, or did it find an extra-expensive one? Does it know what seat I want? Maybe I want to be in a window seat at night, but an aisle seat during the day. Did it put in my TSA PreCheck number? There are, like, so many details, even for something so trivial as booking a flight. And so for me to trust a system that's able to do these bigger things, it actually requires quite a bit of reasoning.

Kanjun Qiu [8:43] That system needs to know, I don't have the right information for this. I need to ask this question. But they can't ask all of the questions all the time. That would be annoying. And they also need to know, like, when is the situation risky? Oh, when I'm about to buy something, it's relatively risky. And so that's a very simple example of reasoning. There are more complex examples of reasoning that we encounter all the time.

Kanjun Qiu [9:05] If you're trying to do a science experiment, collecting the data, analyzing it, how do I analyze the data? What are the factors that are important, et cetera, et cetera? So we really optimize our foundation models for reasoning. And then on top of that, we build—

Matt Turck [9:20] While we're on the topic of reasoning, does it matter what kind of reasoning it is? Reasoning being such a fundamentally human thing, are you trying to build systems that reason like humans, or is that not even the question?

Kanjun Qiu [9:42] Yeah, so it does matter what kind of reasoning. We use reasoning as this umbrella term that's really about: does it have good judgment about what to do? The way to think about today's models is they're book-smart, but not street-smart. They're like a person who's read the whole internet, but they've never done anything in the world. As you would expect—

Matt Turck [9:43] I know a few people like that.

Kanjun Qiu [10:03] Yeah, you do. I'm sure you do. And as you'd expect, people like that are wonderful and have really weird judgment when trying to execute on things. And this is true of models today as well. A lot of what we call reasoning is really about: how can it have good judgment that I can trust when it's executing on things? And how do we get it to do that?

Matt Turck [10:37] Okay. And so let's get into a little more of the weeds of how that works. As I was prepping for this, I read about your unique five-pronged full-stack approach, which is a combination of a theoretical approach and then trying to build that very much as a commercial and practical product. So what is a five-pronged approach, and how does that work, and how do you marry theory and practice?

Kanjun Qiu [10:49] Yeah, so I just mentioned that we train foundation models. On top of those models, we actually build our own agents internally that we try to use and try to actually get to work.

Matt Turck [11:01] And this is not to interrupt you, sorry, but just a quick definition on what an agent is for anyone that may have heard the term but not understand the specifics of what that actually means.

Kanjun Qiu [11:13] Yeah, it's honestly not a very well-defined term. We use it to mean trying to get our models to be able to take actions and accomplish goals. So can it help me—something that we tried to get one to do recently was, can it help me read all of the proposals that were submitted to the AI policy request for proposals given by the Department of Commerce, and analyze those 20,000 pages and try to help us figure out what people suggested and give that to us?

Kanjun Qiu [11:56] Policymakers as a tool. Or another example of a goal is, can it go through our codebase? A lot of what we do is work on agents that code. Can it go through our codebase and find type errors, generate tests where there's no test coverage, help us make infrastructure more robust, do code reviews, et cetera, et cetera? And so we really prototype lots of systems on top of our models that we get to try to do things that we try to do in daily life.

Kanjun Qiu [12:34] Most of our agents are agents that code because most of our work is code. So there's this idea called serious use, where you want to basically, to push the boundary, try to very seriously actually use our systems. It's more serious than dogfooding, in the sense that we're really trying to push them as far as we can go. And I think one thing that everyone sees very quickly is the agents don't work. They're not robust. They're not robust because language models are stochastic.

Kanjun Qiu [12:57] And they output random stuff. And I cannot trust—80% of the time they might fix my type errors, and 20% of the time they do something really weird. And this is actually quite problematic. And so a third prong—so I mentioned models and agents—a third prong of our work is actually on interface. So internally, we have a really interesting agent interaction and debugging interface that lets us build toward a system that we as a user can really trust to do bigger things.

Kanjun Qiu [13:32] So that means seeing how the agent is executing everything and being able to interfere and being able to fork the execution and all of these things. And then on the theory side, I think what's interesting about theory is the way I think about where we are in history is it's very similar to the moment between the analog computer and the digital computer. So in the 1930s, Vannevar Bush made this thing, an analog computer, which is literally a system of levers and pulleys that did an integral.

Kanjun Qiu [14:11] And he set up the system so that it specifically could only calculate this one integral. But that was really powerful because it could calculate this integral over and over again. And that was so cool. And to solve some different problems, you had to reset it up. And what was interesting about the analog computer is you get the situation where actually, if there are errors, they compound, because if it's analog, errors compound. You have to reset it up every time.

Kanjun Qiu [14:45] There's no concept of software. And there was this transition from analog to digital that actually depended on very important theoretical breakthroughs. So Shannon's idea of relay circuits mapping to binary, Turing's idea of the Turing machine and what is computable versus not, and a couple of other very important theoretical breakthroughs. And what that gave us is the ability to build abstractions that led to the digital computer being reliable and programmable. So I think today, we're in that analog-computer phase of AI where these systems output things, errors compound, and they're not very reliable.

Kanjun Qiu [15:22] We have all these issues. It's pretty difficult to get them to do something totally different than the path they're going down. And I think at least some of our theoretical work is about figuring out what are the right abstractions to make agents robust. Once the human is out of the loop, we actually need non-leaky abstractions. Right now, abstractions are very leaky. You can't build on top of them. And so theory is actually a very important part of getting to working agents.

Kanjun Qiu [15:46] And I think ultimately what we hope to end up with at Imbue—the name is Imbue because it's about imbuing computers with intelligence and rekindling the dream of the personal computer, like the truly personal computer that can program itself and help us do the things that we want—I think there's something that's like the set of tools and abstractions that are usable for other people to build agents on top of that, that we hope to provide to others, kind of like an operating system for agents.

Matt Turck [16:24] Great, great. And to play it back to make sure I got it right: so part of the idea is that the agents would be able, for example, to detect that what the foundation model produced was not correct, was a hallucination, and then be able to act accordingly? And if so, any sort of inkling about how that might work? Is that an AI that checks another AI, or are there rules involved or parameters or something that the system administrator can control?

Kanjun Qiu [16:56] Yeah, there are tons of techniques to make the outputs more correct. And it's actually not just about hallucination. I think we use hallucination to say, like, oh, it's just making something up. But often the model will output something just wrong and be like, can you solve this programming problem? And it makes a function that does not solve the problem. And so we do a lot of stuff around critique and a lot of other techniques that aren't in the literature that maybe we've kind of come up with.

Kanjun Qiu [17:25] And this is also part of the work of theory. And those techniques help us get to outputs that are much more reliable, much more reliably solve the problems that we give the model. And when I talk about non-leaky abstractions, that's what I mean. Like, an abstraction is leaky if it's not quite right and you can't really trust it.

Matt Turck [17:50] So, okay, great. And then to take up some of the stuff that you alluded to in passing: you mentioned code as a way to train models to improve reasoning. And again, as I was prepping for this, you referenced models that are pre-trained on code as being demonstrably better at reasoning than those that are not. I was curious if you could talk a little bit about why that is.

Kanjun Qiu [18:15] Yeah, it's quite interesting. Some companies have tried training language models with no code because they're like, oh, the product we're building, we don't need code. It's a therapist or it's a question-answering thing, and we don't need code, so we should take code out of our training data. And what they find is those models are really bad at logical reasoning, and that's really interesting. So why does this happen? A way to think about it is basically the models, what they're doing is they're kind of reflecting the training data.

Kanjun Qiu [18:46] It turns out that code is the most explicit type of reasoning data on the internet. So, in code, I'm reasoning step by step, like this variable is this, if this happens, then do this, if that happens, then do that, I'm referring to this other thing. There are a lot of reasoning primitives in code that are implicit. And so, in some ways, I think of code as kind of like a curriculum for the model to learn reasoning. It turns out that there's not a lot of explicit reasoning data on the internet.

Kanjun Qiu [19:14] People, like, when I'm writing, I'm not writing out explicit reasoning. And so most of the internet is kind of like an associative thing, and code is kind of the most reliable source of reasoning. And so, yeah, part of why we work on agents that code is we kind of feed the code back into the model training, and we train the models on quite a lot of code. And that does help with reasoning. We also do a lot of other things on the data side to help with reasoning.

Matt Turck [19:21] Okay. Can you go into some of that?

Kanjun Qiu [19:38] Yeah. We generate and kind of come up with much more explicit reasoning traces. So that's one thing. A second thing is we do a lot of question answering. Question-answering sequences are kind of like reasoning sequences. Things like that.

Matt Turck [20:02] Okay, very good. I read, I think you guys wrote a post a little while back, a few weeks ago, I guess, in the era of generative AI, about a system called CARBS, Cost-Aware Hyperparameter Optimizer. What does that mean, and where does that fit in?

Kanjun Qiu [20:27] Yeah. So one thing we do a lot of is build our own tools because tools, it's the whole idea of we want tools that help automate the work that we don't want to do. And that's why we build agents as well. So CARBS is an automated hyperparameter optimizer. What that means is that you can give it a model, and the special thing about it is it'll find the Pareto front between cost and performance. So basically, it finds, at every cost of training the model, whether it's small, big, how much compute it takes, how much data you're giving it, at every compute cost, the best-performing model.

Kanjun Qiu [21:10] So it finds the hyperparameters for the best-performing model automatically. And what it will do is this kind of local search, so that once it's found good performance for a smaller model—it usually starts smaller—then it's actually able to find good-performing hyperparameters for large models without trying the bad-performing hyperparameters. So that saves us a lot of time. And what it gives us actually is kind of like scaling laws for almost all hyperparameters for these models. And so that means part of what it means is that we can train at smaller scale and kind of have an estimate of what parameters, what hyperparameters matter at larger scales.

Kanjun Qiu [21:44] There's a second piece of it that's compelling, where CARBS, it's an optimizer, so it optimizes for some metric. And so we've defined evaluation metrics internally that are kind of smooth evaluation metrics. And what that gives us is it gives us essentially like a lens. So I guess we define an evaluation metric that's a smooth metric. And the reason that's important is because I think a lot of people, they think, oh, as we scale up these models, there are these emergent capabilities.

Kanjun Qiu [22:19] And there's a really good paper, I think, "Are Emergent Capabilities of Large Language Models a Mirage?" And what that paper shows is essentially, if your metric is smooth, you actually see slow performance improvement over time. And if your metric is relatively discrete or not smooth, that's where you see the emergence. And it's actually more about the evaluation metric than about the emergence of the capabilities. And so we actually do a lot of experimentation on smaller models and have a really fine instrument for being able to tell, did an experiment improve things?

Kanjun Qiu [22:34] And then that lets us scale those improvements to larger models.

Matt Turck [22:55] And again, to take some of the prongs, so the models themselves, are those models that you build from scratch internally? Or would you take some stuff in open source and most of your work is at the agent level on top? What is the core of the effort?

Kanjun Qiu [23:21] Yeah, we do both. We pre-train our own large models, so very large, 100-billion-parameter-plus models from scratch with our own data. We also use open-source models and fine-tune them to try to see how far they can go. And we also build agents on top of both fine-tuned models and our own models. I think an agent uses multiple models, or can use multiple models. And so it's not just that there's one massive model.

Matt Turck [23:49] Great. So, how is that all going to manifest? I guess, what is the medium-term strategy for the company? Do you consider yourself to be mostly focused on research, or are you going to build specific products, ChatGPT-style kind of thing? Are you going to be focused on the enterprise? And if so, what is the timeline?

Kanjun Qiu [24:07] Yeah. So I think the way I think about it, the core hypothesis is there's something that's blocking agents from really working that well. I think when we try to build agents ourselves and talk to everyone else building agents, it's just very hard to get them to be robust and reliable and trustworthy. And this is not a new problem. Like, back in the early computer days, it was really hard to get programs to work and be robust and reliable and trustworthy.

Kanjun Qiu [24:31] And so, in a lot of ways, this is about building something that's a little bit like an operating system, maybe a set of programming languages, that allows us to actually be able to build agents that are robust and reliable and trustworthy, and that I, as a user, can actually use. And so some of the way we think about our work is assembling that operating system so that we can build agents, so that other people can build agents—first developers, then maybe eventually regular people.

Kanjun Qiu [24:51] And because these models can write code, you don't necessarily have to be a developer forever to be able to build agents.

Matt Turck [25:14] Okay. But the timeline for all of this depends on how, I guess, tractable that problem is around having truly reliable, safe agent systems that you're building. So the exploratory research part has to be solved before you think of enterprise use cases.

Kanjun Qiu [25:37] Yeah, I would say the timeline—it's not like 10 years. I think that's very unlikely, but it's not like in three months. So somewhere between three months and 10 years. The way I think about it is there is a right time to bring a technology to market. The iPhone was in development for 10 years, AirPods for five years before they released it. And it's because the touchscreen wasn't good enough yet. And I think there's a very similar situation going on here where the models aren't quite good enough.

Kanjun Qiu [25:52] We don't have the tooling in place yet. There are a bunch of pieces to assemble, and those pieces are appearing very quickly, but who knows how long that will take.

Matt Turck [26:12] Yeah, a little bit to those pieces appearing very quickly: how does one manage a company like this, especially in a time like we are right now, where new things seem to be appearing all the time? And do you have a series of discrete teams working on different parts, and then you're trying to put everything together? So again, I find it fascinating, this intersection between fundamental research, turning this into products, and how you land it all.

Kanjun Qiu [26:43] Yeah, it's pretty straightforward, honestly. We have projects, not teams. I don't like teams as a concept because they're like self—never mind. We have projects, not teams. One project is pre-training, and fine-tuning is a little bit attached to that. One project is agents, so getting agents that we can use. And then we have a project around infrastructure, and we have a project around data collection, and they all feed into each other.

Kanjun Qiu [27:10] So the purpose of data collection is to get good data for the pre-training, and that data can come from our agents, actually, and that data can come from everywhere else. The purpose of pre-training is to make the models better for our agents so that the agents actually work. And we have a lot of evaluations. So creating evaluations is a big part of what we do in order to figure out, like, are we actually improving things or are we not improving things?

Kanjun Qiu [27:41] The purpose of agents is to get agents that we are able to use every single day. So internally, every single day, that's across analyzing policy, recruiting, writing code, and all sorts of other things. And that's how we organize it. And so the serious use of these agents drives improvements in everything else. And then the purpose of infrastructure is making it so that we can very easily, reliably run experiments, reliably run agents, developer productivity is really high, et cetera.

Kanjun Qiu [27:53] So those are the pieces.

Matt Turck [28:14] Great. And what is your fundraising strategy, and how does the recent round fall into all of this? Among other things, I'm curious about your lead at Astera Institute. Do you want to tell us more about who they are, why you chose them, and how it all came about?

Kanjun Qiu [28:45] Yeah, so we were kind of preempted for this round by our Series A investor, Jed McCaleb, who runs the Astera Institute. And the reason why we chose Astera Institute, which is a science nonprofit, is because these timelines—in order for us to really build something that is essentially an operating system for agents—it's kind of like building the first personal computer. And I think it can be a very historic kind of thing if done well. I think the current way that startups are encouraged to be built—

Kanjun Qiu [29:17] I've been a founder many times, and the way we encourage startups to be built is: find something, build something people want, be very locally focused on a problem. And that's good. And it causes a kind of myopic thinking, I think, in this kind of opportunity. And so we wanted to choose a funder that could have slightly longer timelines and that was not immediately focused on commercialization. And then NVIDIA is also part of the round because we have a very large GPU cluster of 10,000 H100s, and it's very helpful to have NVIDIA as part of that.

Matt Turck [29:38] 10,000. Okay. And so the $200 million is going to go to some of this. So it's sort of GPU compute cost kind of thing.

Kanjun Qiu [29:39] That's right.

Matt Turck [30:11] How do you think about—I love the ambition and how you're trying to solve absolutely fascinating fundamental problems—how do you think about the current landscape in terms of competing with just about everyone, the OpenAIs and the Google DeepMinds and research and all the things? Are there different approaches? Are they 100% competitors, or do they do things in a way that can be compatible? How do you think about it?

Kanjun Qiu [30:38] Yeah, the way I think about it is, if we're in 1980, it really doesn't matter what other people are doing. If I'm Apple in 1980, it's irrelevant what people are doing. What matters is what we're trying to do and whether we're right. And so, I mean, it matters a little bit if OpenAI is trying to do exactly the same thing as what we're trying to do, and they replicate exactly what we're doing and there's no moat. In that situation, okay, maybe that's an issue and we have to figure out how to execute.

Kanjun Qiu [31:11] But I think the way I think about it, this is like huge blue ocean. This is like a foundational technology that's once a century, equivalent to analog digital computers. And the opportunity is essentially like free intellectual energy. I think whenever we have free energy appear, this is a very powerful thing. With the steam engine, we got free energy. With petroleum, we got free energy and the automobile and everything. And with the personal computer, we got a little bit of free intellectual energy.

Kanjun Qiu [31:32] And now what we're unlocking is like almost infinite free intellectual energy. And that's incredibly powerful. And so what we're trying to build toward is a computer that is able to unlock, for every person, for every company, that free intellectual energy. And I think we have a pretty specific thing that we're trying to do, and I don't know what other people are doing, but it seems—I think a lot of our focus is on serious use because it teaches us what actually needs to be done in order to get these systems to work really well for people.

Matt Turck [32:12] So, all right, we talked about GPUs as used for the $200 million, and obviously a big part of the current competition is the war for talent. How do you recruit in this environment? What kind of people are you looking for? I've heard interesting things like, for generative AI, you almost want people that actually haven't spent too long in other parts of AI, and you want younger folks that are going to be fresher.

Matt Turck [32:31] Is that true? Is that not true? Like, who are you looking for?

Kanjun Qiu [32:57] I've always struggled to describe the type of culture we have. I think it's expressed relatively well if you go to our careers page, but we hire people who think from first principles and kind of don't accept the current state of things as given, and who have very high agency. So everyone on our team, I like to think of people as creative agents.

Kanjun Qiu [33:28] I think a lot of companies, they're very proud of thinking of their people as assets. And I think, like, if you think about what an asset is, this is such a low bar. Like, an asset is a thing you own that provides value to you that you can discard at any time. Like, people are not that. And so, thinking of people as creative agents, our projects are very dependent on who's on the team. The reason we have CARBS is because Abe was a plasma physicist, and he applied some ideas from plasma physics to make this local search algorithm work well.

Kanjun Qiu [34:06] That was crazy. And Abe kind of made an Abe-shaped blob at Imbue. And almost every one of our projects is because somebody was kind of uniquely suited to doing something like that. So Bartosz on our team, there was a broken NVIDIA 3090 GPU, and he was able to debug it to figure out that there was one capacitor that had blown out, and he replaced the capacitor and it worked. And I was just like, what? How did you do that?

Kanjun Qiu [34:38] That's crazy. So part of why we're able to go all the way down to hardware is because of that. And so, the team is kind of like—a way I think about it is what we're building is the team, and each is the expression of each person. We're certainly not a factory-style, top-down, like, everyone's going to do a particular thing toward a very specific roadmap. I think we learn collectively. We really rely on collective intelligence.

Matt Turck [35:04] Well, thank you for all of this. One thing I'd love to do before we close is to talk about some of the other projects and stuff that you have been working on. As I alluded to at the beginning, you are a person of many talents. I read somewhere that you're working on a neighborhood project. What is that?

Kanjun Qiu [35:29] Yeah, so a couple other projects. One where I work on kind of trying to cultivate a neighborhood in San Francisco, really basically like a mile-by-mile radius where all of our friends live. And the reason is, there's actually a study by the military about what leads to really close friendships, really close relationships. And it turns out there are three things. There's shared experiences, a space where you can feel like you can let your guard down, and spontaneous interactions.

Kanjun Qiu [36:03] And so, spontaneous interaction is actually not something we get very often. It's why a lot of our closest friendships are formed in college or potentially in the workplace, if the workplace feels safe, which satisfies the second criterion. In 2015, I started a house called The Archive with a bunch of friends. And the reason we started the house was because of this study of, like, oh, okay, that actually allows us to form really close friendships. And out of that house, our housemates, Tom Brown and Ben Mann, are the first two authors on GPT-3.

Matt Turck [36:10] Mm-hmm.

Kanjun Qiu [36:35] And so out of that house came a lot of really interesting stuff. And I think the neighborhood is kind of like a scaled-up version of that house, of bringing all of our friends together into a place that's—we call it scenius. It's like genius, where if genius comes from your genes, scenius comes from your scene. And so we learn a lot from the people around us. The idea flux of the neighborhood is excellent.

Matt Turck [36:47] That's so fascinating. So how does that work practically? So you just encourage people to move within the same neighborhood, and then you create community events of some sort? I'm just guessing, is that how it works?

Kanjun Qiu [37:06] Yeah, most of the credit goes to Jason Benn, my former housemate, who's the one who's driving everything full-time. And essentially, he goes door to door to get to know every landlord, and the landlords tell him when units are free. And then when there are good units free, he helps find people to move in who are on the list. And yeah, so we're trying to be able to hang out with all our friends, kind of like an adult college campus with lively intellectual life and events and everything.

Matt Turck [37:32] Fascinating. Another area where you seem to spend time and do really interesting work is meta-science. Do you want to talk about this and maybe define what that is?

Kanjun Qiu [38:03] Yeah, sure. So, a lot of what I think about is how do you design social processes that allow people to unlock their potential collectively and individually. So companies are comprised of social processes. Actually, at Imbue, we have all sorts of weird social processes that are awesome. For example, when we do quarterly planning, we actually have everyone write simultaneously in a Dropbox Paper document, like, write out ideas for things to do, problems we're facing, and then respond to each other simultaneously. So it's like a giant synchronous-asynchronous conversation where everyone talks at once.

Kanjun Qiu [38:33] And it's so awesome because we're not waiting for people to stop talking. The loudest voice does not dominate. And we do this simultaneous writing for all sorts of things. We do it for standup meetings. We do it for planning. We do it for, like I mentioned, quarterly planning. We do it for product thinking, all sorts of things. That's an example of a social process. Another social process we have is Feelings Friday, where we share our feelings and how we feel, our stories about our feelings.

Kanjun Qiu [39:13] So the work on metascience with my collaborator, Michael Nielsen, who drove most of it, is about essentially how do we design the social processes of science such that they actually lead to fundamental discoveries, better, more interesting discoveries faster that are more fundamental. And a thought experiment that might be interesting, that kind of motivates this problem is, let's say we encountered aliens and the aliens do science. Would we expect those aliens to have PhD programs or tenure or universities?

Matt Turck [39:13] No.

Kanjun Qiu [39:35] It's actually interesting to think about what we might expect those aliens to have. We probably do expect them to have mathematics. We probably do expect them to know what atoms are. We probably do expect them to have something that's kind of like a microscope. Those things are much more fundamental than the social processes like tenure and PhD and peer review and things like that. And so the question is, it's kind of like these social processes of science have happened accidentally.

Kanjun Qiu [40:02] What if we treated it as a design problem? What would that look like? What kinds of things could we design? And what are the barriers, and how do we overcome those barriers? And so one of the main ideas we have is this idea of a meta-science entrepreneur. So an entrepreneur that's able to experiment with new social processes that eventually get incorporated into the scientific system. That's some of the work on meta-science. I love it.

Matt Turck [40:07] And how does that manifest? Is that through a series of conversations or papers or?

Kanjun Qiu [40:21] It's mostly an essay. There's an essay you can go to on my website called "A Vision of Metascience as an Engine of Improvement for the Social Processes of Science." And that's kind of part of a separate community, the metascience community.

Matt Turck [40:27] I mean, you have a whole section which I love called Kanjunctures.

Kanjun Qiu [40:34] Kanjunctures, yes.

Matt Turck [40:39] You want to talk about those? Maybe people will go check out those.

Kanjun Qiu [41:05] Yeah, it's sadly underpopulated. I think something my friends experience a lot of is, like, I'll make up theories for why things are the way they are all the time. So, like, I have kind of this, like, trauma is overfitting. Like, we think of people having trauma as, like, whatever's happening is not well-suited to my current situation. And so trauma is overfitting. And actually, a good way to overcome trauma, if we look at a lot of therapy techniques, is by giving more data to me and helping me access the overfit parts and then give them more data.

Kanjun Qiu [41:36] That's an example of a conjecture. It's like a conjecture. I have not run experiments that are rigorous. But I think it's interesting theoretical frameworks for different things. And then we also have a fund called Outside Capital. And that actually drives a lot of the community convening in San Francisco. It's kind of part of the neighborhood, part of Imbue, where we host these Thursday night events called Thursday Nights in AI and bring in speakers and things like that.

Kanjun Qiu [41:48] Great, great.

Matt Turck [41:52] And in addition to that, what does the fund focus on?

Kanjun Qiu [42:00] Mostly focuses on future of work, kind of deep-tech future of work, which is mostly AI these days.

Matt Turck [42:14] And you are a seed-focused firm or later stage? Like, tell us more, and maybe how anyone listening to this can reach out if they think that might be a good fit.

Kanjun Qiu [42:35] Yeah, pre-seed and seed, everything from, like, idea in Notion doc. We have invested in a founder with an idea in a Notion doc all the way to, you know, I have a product, some customers, raising my seed round. So everything there. And you can DM me on Twitter or apply on the website. Honestly, if you apply on the website, it'll probably be much more reliable than DMing me on Twitter.

Matt Turck [42:52] Okay, very, very, very good. And maybe as a last question, a personal question on my end, like, how do you manage it all? You seem to be doing all those fascinating things. Like, what's a day in the life? How do you just fit it all?

Kanjun Qiu [43:18] Yeah. Well, really, the real answer is everyone else does all this other stuff. I just help them. And the only thing I do is Imbue. So all of my day is spent on Imbue, thinking about Imbue. It turns out that those thoughts are useful for other things like metascience or, like, the fund or, like, the neighborhood. But yeah, I don't do any of these other things.

Matt Turck [43:35] Okay. Terrific. Very, very, very exciting. Very impressive. Kanjun.me, K-A-N-J-U-N dot me. We talked about Twitter and how a good way to reach out to you is to DM. Where else can people find you? How do people find the podcast?

Kanjun Qiu [43:47] GenerallyIntelligent.com and click on podcast, or search for Generally Intelligent. That's the podcast's name.

Matt Turck [44:11] Okay, wonderful. Terrific. Well, that was a really fun, multifaceted conversation. Really appreciate your time and very excited to see the next steps at Imbue and what comes out as you build this OS for agents. Absolutely fascinating, and really appreciate it.

Kanjun Qiu [44:14] Thank you so much, Matt. Have a good one.

Matt Turck [44:15] Thank you.

Kanjun Qiu [44:44] Bye. Thanks for joining us for The MAD Podcast. We're back here every Wednesday with new conversations with leaders in the machine learning, AI, and data space. And if you like this show, you can also find a video recording of not only this episode, but many, many more over on the Data Driven NYC YouTube channel. Thanks again, and catch you next week.