Building an Intelligent Business OS, with Runway CEO Siqi Chen

Hello and welcome to The Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week we'll explore their revolutionary ideas, and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my co-host Eric Thornberg. Hello and welcome back to The Cognitive Revolution. Today I'm speaking with Siki Chen, prolific Twitter commenter, active angel investor, and founder and CEO of Runway, a finance platform that's revolutionizing how businesses understand and interact with their financial data through a combination of traditional software and AI. Sikhi's journey from math major to game developer to 4-time CEO, including a stint at Sandbox VR during the onset of COVID has given him a deep appreciation for the pain points of traditional methods of financial planning. With Runway, he's not just building another spreadsheet replacement. He's crafting an intelligent business operating system that can generate complex financial scenarios, explain intricate metrics, and ultimately help all employees understand how their work contributes to the business's bottom line success. In this discussion, we explore how Runway is layering AI onto their product, how they are using AI operationally, How far Siki expects AIs to progress in the near future, and what all that implies for the world at large. Siki describes Runway's ambient intelligence product strategy, which aims to surface AI-powered insights without users needing to explicitly query a chatbot, the prompt engineering strategies that they're using to present financial and other structured data to LLMs in a usable way, and the reasons that Runway has integrated more than 680 different data sources to date. After that, we get Zeki's perspective on the possibility of self-improving AI systems, how his AI worldview informs his approach to company building, investing, and even parenting, his expectation that relative status and things provably made by humans will increase in importance and value, and how he thinks at a meta level about which AI scenarios are worth thinking about and planning for, and which are, even if not entirely implausible, simply too crazy to spend time worrying about. If you're finding value in the show, we'd appreciate it if you'd take a moment to share The Cognitive Revolution with friends or leave us a review on Apple Podcasts or Spotify. Meanwhile, we invite your feedback and your resumes for our experimental private job board via our website, cognitiverevolution.ai. And I always welcome your DMs, whether by email or your favorite social network. Now, I hope you enjoy this conversation on the application of AI to finance and much more with Siki Chen of Runway.

Siki Chen, founder and CEO of Runway, welcome to The Cognitive Revolution.

Thanks for having me. It's great to be here.

I'm excited about this. You are building a company that I think is at a very interesting intersection of problem space that's really important that people have to get right. And you're also building a lot of AI features into that product. I think that's a really interesting place to be right now. So I'm really excited to get into that, get into some of the weeds of what you're building and how you're thinking about that. And then hopefully at the end, zoom out and talk about some of your big picture views. For starters, you want to just tell us a little bit about what Runway is? Like, there's an interesting backstory as to why you decided to start this company, and maybe you can start to tee up some of the AI integration points.

Sure. Yeah. So at a high level, Runway is a finance platform. For the finance team, the main goal is to make the business understandable to not just people inside of finance, but outside of finance too. This seems like a very dry and boring problem, but it's probably the most interesting one I've worked on in my career. The backstory is I've been a 4-time CEO and a 3-time founder now. My background is in math. I majored in math and I have a technical background. I was never comfortable with business and finance. I never thought of myself as any kind of business person. As my career progressed over all of these companies, I felt increasingly insecure about that identity. And this came to a head when I was appointed CEO of a company called Sandbox VR shortly before COVID hit. And when it did hit, we had to do all kinds of scenario planning on these spreadsheets. What these spreadsheets do is, you know, probably heard of the term of financial model before. And not only is that used by investment bankers on Wall Street, but it's also more importantly used by CFOs and founders who run companies. And what these things do is they are a simulation of a business. It basically lets you time travel in the future and understand the implications of the decisions you make today about how much money to raise, how to change pricing, how much to spend on marketing, how many people to hire. And so when COVID came to a head, we had to scenario plan for how long COVID is gonna last. Is it gonna last 3 months, 6 months, 12 months? And we're doing them on these spreadsheets. And long story short, we had to lay myself off and most of the company went from 400-some employees to basically around 10 or so. And the founder hibernated the company over COVID. Shortly after that happened, literally half an hour, I was talking to both our CFO and then the Andreessen team who invested in Sandbox. And I asked them like, surely there's something better we could have used for this problem to like scenario plan and see what's going to happen with COVID Because Figma exists and Notion, and we have all of these great products, surely there's something that is intuitive, understandable for the business itself. And I was surprised to learn that actually wasn't anything better than just simply using spreadsheet, which is amazing, but also like it can do everything. So it's not really good at any particular thing. And that's how Runway got started. The more we work on it, this problem, the more we realize that this is actually a really interesting and deep problem. And the impact could be really large, right? As employees, we work in these companies and we're just used to not understanding finance and not understanding business. It's supposed to be mysterious. But if we can understand it at a deeper level, we can make better decisions and we can align and collaborate better. And there's a lot of interesting ideas and AI plays a big role in making that a lot more accessible.

So listeners will know, and you're probably getting the sense that I always really try to do my homework and get a very hands-on sense of the products and services and technologies and models. research that our guests have worked on. In this case, I am not yet a Runway customer, but certainly the pain point resonates with me a lot. Having been a founder CEO, I can recall scenario planning late night before the board meeting. Like, I was already supposed to have this done, but, you know, there was more. I always had the attitude that I got to actually run the business more so than I need to polish this stuff up for the board. They didn't always necessarily agree with that or love that attitude on my part.

Yeah.

But I can remember quite a few late nights tinkering with spreadsheets, either for historicals or more often for forward-looking scenarios. So I definitely identify with that pain point. And then I did take the demo last week to get a hands-on look at it. Oh, great. Wonderful. A couple things definitely jump out. One is that it does have this somewhat of a pattern of people identifying things that are done in spreadsheets and then saying, can we build a better dedicated product for that specific problem. This definitely feels like that. I think the way that it connects historical data to present and future views is really nice and definitely feels like the future relative to what I used to be doing. Also, the design is really great. The design is really strong and the branding is sort of—

This is starting to sound like a runway ad.

Yeah, it's not, but I was impressed. I was genuinely impressed. And I did actually send it over to Waymark CEO, took over for me after I got the AI bug so bad that I couldn't think about anything else. And so we are officially in your pipeline. Anyway, all that said, not a paid endorsement. I'm just enthusiastic. Let's get into some of these AI features. And folks can hear any number of other podcasts and conversations that you've had to get other angles on what you're doing. We'll focus on the AI angle. The feature that I saw that I thought was like, whoa, this really is a shifting moment for how I would think about finance and think about how not just how, but also how much of this sort of forward-looking financial scenario planning I might do was when we just did a text prompt to scenario, and we just asked the AI that's built right into the Runway product to create a scenario where we ramp up our marketing spend enough to 3x our revenue over the next 12 months.

Oh yeah.

And just with that kind of high-level thing, it spits out a scenario for us that we're now looking at. And I'd love to go under the hood there and understand what is happening that is leading to that scenario popping out.

Yeah.

So there's a few ways to look at it. So if you think about what a CFO would typically do, right? The CEO would have a question like, how do we triple revenue next year? And the CEO would talk to the CFO. Or whoever owns the model. And then what a CFO would do is go to this model that has all of these simulations, these relationships around how different things affect revenue and growth 2 years down the line. And they'll literally try some numbers out, right? What happens if we hire more salespeople? What happens if we increase the marketing budget? What happens if we increase conversion? And say, if we do roughly these things, we'll need to raise more money. We'll need to increase prices. We'll need to hire more salespeople and we get there. And here's how it affects all of the other metrics that you Yeah. And it's a very like iterative sort of manual process, right? And if you think about what's going on in a CFO's head, you have some kind of intuition about what are the likely drivers that most ballistically impact those constraints that you're looking for, right? So if you think about revenue, then you think, okay, sales, marketing, pricing, and that kind of thing, conversion rates. And then you go to the spreadsheet and change those numbers, and then you visually look to see if the output matches the constraints that you're asked. And so when you think about it from a technical perspective and how you can use an LLM to make that better and faster, what are LLMs good and not good at? It's good at making a guess. It's good at common sense reasoning about what are things that are probably related to revenue. What is it bad at? It's really bad at doing math, right? It's not super intelligent yet. It can't just do math in your head, which is Interesting observation because humans also generally aren't that good at that. We're good at language, but you would have to be— there's like Rain Man-like savants out there, but most people can't do very complicated math in their head, which is why spreadsheets exist that helps us do these calculations. And so one way to think about how this works is we're effectively giving this LLM a calculator and having the LLM do what it's good at. It can understand a constraint. It can look at the constraint and understand what are the drivers in this model that most likely are going to impact a constraint. It can evaluate whether the output of changing those constraints match the constraints that you want. But the actual calculation is done by the model engine. So that's like one very high-level way of thinking about this. But to make this work, the reason why it's actually substantially more difficult to solve that same workflow in a spreadsheet is the context doesn't necessarily exist in a spreadsheet. You need to have semantic labeling of the drivers and sure, you can look at the rows, but you also need a lot more context. You need to know like, what is a product roadmap? What is a hiring plan? What is a fundraising plan? What is a marketing plan? And plans usually actually don't exist in that spreadsheet. It lives in like places like Notion and Google Docs. And so in order to make this work, we have to invent new abstractions and primitives that can contain business intent context, like a product roadmap, like a fundraising plan and a marketing plan. And with all of that context, plus the logic of the driver, then you have enough context for an LLM to do something reasonable to answer that question.

Yeah. Okay. That's really interesting in multiple ways. The sort of Definition of a syntax for something like this is— that's a challenge that I've been through a couple of times with our company, Waymark. We create videos for small businesses, and in some ways there's like a similar problem where we have the sort of form that we want the AI to write a script for and select the assets for and ultimately create the voiceover for and all these different things. But we need to make sure that it respects the right format in its generation, a lot of time actually, especially in the early days, went into defining what that syntax would be so that it would be as lightweight on the AI as possible so that it could actually do it. Also, of course, it had to be something that we could rehydrate deterministically with code on the backend of the generation.

Yeah.

We started relatively early. I'd be interested to hear where you started and how that syntax has evolved. I imagine in our case, it started out really simple because it was all that the models at the time could handle. And if anything, I think now we're trending toward a more rich representation because the models are getting better at handling those. So I'd love to hear what your journey has looked like to figure out how to represent this in a form that the AIs can work with.

Yeah, it was somewhat of a technical challenge in that, like, The, on the backend, these drivers are represented by UUIDs, right? So even hydrating the UUIDs into semantic descriptions of what these things do was some amount of work to just get stored on the backend. But we go beyond that because the way a model works isn't just about a single row or a single formula, right? You can say runway equals cash flow by burn, and that represents that row. There is an entire dependency tree on like how runway is calculated through burn, and then burn is affected by expenses and revenue, all these different things. And so one of the things we've done that we use is actually we proactively use AI to describe parts of the submodel. So we say, this is runway and this is how it's calculated. This is burn, and in this particular model, it's how it's calculated. And so what we found is that the more context we give the LLM, the better it's going to be able to figuring out what are the right guesses on what to perturb first. And you can generate that context. Also through AI, which is a really interesting thing. And that's what we do. And we find that that is like a really interesting direction in terms of how AI is expressed in our products. And so the way I think about this, and I know I'm like deviating a little bit from your more technical question, but the opportunity we've seen in the product is if you look at how AI has been expressed since GPT-3 has come out, initially people had did a bunch of experiments and were cool demos. You call the API. But nothing really broke out until ChatGPT, right? That was a moment where, okay, this is like a very interesting interface that people can understand and get a lot of value out of immediately. And what's happened since then is everyone has basically added a copilot to their product. There's chat with whatever, and there's like a side pane on the side. And then now we're in the world of like agents, except like the models aren't capable enough to self-reflect. And so agents by and large, like— It's hard to get real productive work done with them. I think the next range of models, when it's able to self-reflect a lot better and plan a lot better, I think we'll see a lot more economic value. But that's how things have evolved. And some observations about that is, one, the thing about a chat interface is that it's really accessible, but the 2 biggest opportunities for improvement is, one, it doesn't have a lot of context in most cases. ChatGPT has some memory now. It doesn't really know a lot of business context about who you are and personalization and like all the meetings that you have in your calendar. And so a bunch of companies are working on that problem. The more context, the more helpful and proactive it can be. And the second thing is its proactiveness, right? What I've found really surprising is despite how many people use ChatGPT, how few of them really understand what you can even ask it to do. People are generally underutilizing what ChatGPT can do. And so the way we think about how those two, those two things connect is where we are with models today. I think the best use of AI is not through necessarily just a chat interface or through agents, but we call it ambient intelligence. It just does the work in the background without you having to ask. And the way it's expressed is through the interface itself. So a real example is. When you look at a driver in Runway, like a row, and you can see the details of it and trace through it, there is just a description there. You didn't have to ask it for a description, right? It just told you and it just looks like text. It doesn't even look like AI. And I think the way Apple has thought about how you express AI is very similar to how we think about it. In the course of your normal workflow, you could surface it as text. You can surface it as suggestions. Without someone having to like have a conversation with a message window. And I think that is like the best way to create value from AI today. And that's a lot of how we think about it.

I really like the idea of using AI in the background to synthesize metadata. That to me is a definitely a strong emerging pattern. And we do that a little bit at Waymark 2 in our case. I'd be interested to hear how much like you had anticipated building models into the product when you started with the product. In our case, case, we really didn't. I was always interested in AI, but it seemed like at the time that we were building the product that anything that could actually help our users create content was like a long way off. And then it showed up a lot faster than we anticipated. So we created this template system and initially users would fill it out. And then we learned later that there's an opportunity for the AI to fill it out, but they weren't yet multimodal. So it was like, we need to represent this in text. Well, how do we represent it in text?

Yeah.

And there became this, this whole kind of elaborate pipeline for figuring out what exactly is on screen. We know what our wonky JSON backend representation is, but what does this actually look like to a person? That was a whole process. When did it become clear to you that the models were getting good enough that it was actually going to be able to help with a finance use case?

We started Runway maybe just right before GPT-3 came out, I think maybe a few months before. And Naval Ravikant was our first check, and he actually named the company Runway. There's a funny story about this because when I was talking to him about this idea for Runway, I asked him what he thought about the name CFO.ai, which is a domain we still own and we don't use. And his feedback was, no, don't call it .ai. This is back in 2020, right? And he said the reason why you shouldn't call it .ai is because it's gonna be dated really quickly. Everything's gonna be AI. And so the company with .ai as a name, it's not gonna be a sustainable name. And I thought that was, in retrospect, that is such a prescient observation because the way we think about AI today is you don't talk about your product as powered by JavaScript, right? Or AWS. It's just how it works. And that's the right way I think about AI. Yeah. It's not like a separate thing. It's just part of the capabilities of the product. And one of the many avenues that you can use to create value. That was part of our thinking in the beginning. The mission of Runway has never changed. The mission has always been to make business accessible and understandable to everyone. And when you think about AI in relation to mission, it's not just a thing we added in later when it became hot. What large language models are really good at today are precisely those 2 things, making things more understandable by taking something that is maybe complicated and explaining it in plain English, right? And also making things accessible by shortcutting these complicated workflows so you can ask it to do something in plain English. That's something that it is really good at. In fact, I thought it was so good, like I currently have the number one GPT in the ChatGPT store under the education category called Universal Primer. And the entire thesis is that is something that LLMs are good at. It can take something very complicated and explain it to you like you're 5 and also fill in the blanks of any prerequisite concepts you're missing. And so that's how we thought about LLMs from the very beginning of Runway is how can we make this very complicated business understandable to everyone?

Cool. I missed that in my scouting report on you. So that's cool that you have the, the number one GPT. Are you cashing any checks from that yet? Still waiting on our mailbox money from our GPTs, right?

I am. I can't disclose how much or the method under NDA, but I can at least confirm that yes, I am. Cool.

I haven't even heard that they're actually paying out. So that's also a little—

Yeah, they said specifically don't talk about the method or the amount, but they didn't say nothing about that. Can't disclose. We're not getting— we're getting paid. We are getting paid.

Cool.

Interesting. So a couple of great little tidbits there. Going back to how it works when you are generating a scenario, I am interested in what the language model is being asked to generate. Are you using a tool use or function calling paradigm? And then you said you're giving the model a calculator. So In this case, like Runway is the calculator, right? But yeah, interested to hear a little bit more about what that interface looks like between the model and the more deterministic, explicitly coded product. Is that like an iterative goal finding sort of thing? If I say I want to get to 3x revenue, is it cycling until it kind of hones in on that end state?

Yeah, exactly. The great thing about LLMs is you can have some intuition about how it should work based on what humans do. So a CFO understands what are the— has some intuition about what is going to drive revenue. And with a model, we can do better than that. We can provide the entire dependency tree of a model to the LM and say, here is roughly how the model works and here's the different drivers. Take some guesses on what you think are going to perturb revenue or whatever constraint you want the most. And once you've selected that. Take some guesses as to when. So if the constraint is like 3x revenue next year, then reasonably you're probably not gonna try to do it 5 years from now. You're probably gonna do it like 3 months from now, 6 months from now. You're gonna try a few different things, right? And when you try a few different things, we run it through this calculator that is a model. And then we look at, okay, what was the constraint that you were asked to optimize against? It was revenue, right? So we said, okay, which one is that? It's gonna be the one that says revenue. And then we compare the outputs before and after. And say, is this 3x? Right? That's a calculation we can trivially feed it. And so the way this workflow works is exactly what a CFO does, right? We are not asking it to do the math. We hand it basically this black box calculator and we ask it to basically figure out what are the most levers, likely levers to pull and try pulling some levers. And we're gonna tell you what happened. And then you're gonna look at that. It was a higher, lower, how much was it off by? And then you're gonna try some other things. And you iteratively get close. It doesn't work all the time, but that is definitely something that as the model capabilities get better, it will get better at. And the way we orchestrate it, as that gets better, it will also get better.

That sounds very much like the Assistants API. I don't know if you want to comment on what underlying technology you are using, but if I was going to point to a single technology that maps onto that most closely, it would probably Be that?

Yeah, we don't use that. It is a very unique use case in that we have a custom model that's different from company to company. And so we built a lot of very custom orchestration systems to make that work. But yeah, that is a pretty good analogy.

Hey, we'll continue our interview in a moment after a word from our sponsors. In terms of the improvement, what do you think is Maybe what has and what do you expect to drive the most improvement in the overall value to users? Like, how much is coming from scaffolding orchestration and how much is coming from just raw scaling and more fundamentally capable general models?

I think generally for most problems, including ours, the answer is both. So what I mean by that is obviously, As the models get better, intuition, reasoning, planning, self-reflection, all of that has incredibly high leverage. And it also drives the architecture of the orchestration. Orchestration will need to change as it's able to self-reflect better. But for our uses in particular, orchestration is particularly interesting and useful. And I'll say it goes beyond orchestration. It's like, what is the substrate that you're orchestrating on top of? And what I mean by that is it would be very difficult to create that workflow on top of the substrate of simply a spreadsheet. And that's because the abstractions that exist in a spreadsheet are cells and rows, and they are not well suited to contain business intent contexts like a product roadmap or fundraising plan or marketing plan. And so job one was actually creating the right substrate so that you can connect that context With the business model and the plans together and pass it to an LLM. Once you have the right substrate that can contain that context, then orchestration is important in order to feed the relevant context to the LLM. And at that point, larger context windows help, better understanding helps, but a substrate is not something that an LLM is going to handle for you outside of the box. And this is where I'm a big believer in this idea that you can't just jump straight to an LLM, at least for the next year or 2. What you can get an unfair advantage of is building a tool for thought that has the right abstractions to contain more and more context that people already use. And that's a much better starting point to feed to an LLM because you can feed in more context.

I'm looking in my notes to see if I have the number of integrations that the product supports, but I believe it's in the hundreds, right?

It's about 680 today.

680. So I assume that is also something that kind of started for non-AI related reasons, but turned out to be a perfect Swiss Army knife to pull in context. And I'm assuming that this is where the marketing plan comes from or the product roadmap, right? People aren't entering that stuff directly into Runway in most cases, right?

Actually, a good number of customers are. So we integrate with just about everything because finance is not about really just finance. I think the right way to think about it, the way we think about this, is that it is a simulation of the entire company. And so on some level, everything that happens in the company, every person, every department falls down to the bottom line eventually somewhere. And so it's all kind of part of the finance model. It's just that today the tools that we have make it so that the finance model is really isolated. And as a result, if you're a salesperson, the sales team has a model, the product team has a model, the growth team has a model, the marketing team has a model. All these are in different spreadsheets everywhere, and they really ought to connect to each other, right? So that you have one view of how the marketing plan affects the sales plan, affects your overall financial health of the company. And that doesn't happen today. And so we integrate with a large number of data sources because Everything that happens in a company is eventually part of that financial model. And you want to be able to track progress against all of those things. I think the most interesting thing that we've seen is that we have companies integrating Jira into Runway, which for a finance platform, that's unheard of, right? Like finance people barely use existing finance platforms because they're all really difficult to use and quite slow and inflexible. But the reason why is because we have an abstraction called plans that actually can encapsulate business intent context. So you integrate with Jira so you can attribute a ticket or a project to financial outcomes, how it affects conversion rates, how it affects growth, and then see how it affects margin. And so all of that context is useful for an LLM and it'll only increase in utility over time because it's hard to make decisions off of just a financial point of view. Right? It's a lot easier to make decisions if you have a holistic understanding of what is a product team's intent and roadmap, what is a sales team intent and roadmap and growth and marketing, all these things. And then as the LLMs get more capable, it can create richer plans that reflect that business intent as opposed to just changing the numbers.

So the vision that I'm starting to see through the fog for myself would be like, if I've integrated my task management platform into runway, I might be able to go in and ask something like, what's the financial impact if we do the ticket that's on the onboarding funnel and it improves signup rate by X, Y, or Z%? What's that look like? And how does that compare to if we handle longstanding pain point, which might improve retention by A, B, or C increments? And then the idea would be like, I can see the range of these 2 things and start to prioritize my— actual work with this kind of visibility into long-term financial results?

I think a nuance there is that's something that you can do today in Runway and you don't even need AI to do it. So if you can say this Jira ticket is gonna increase conversion by 5%, right? You can just do that and look at instantly the impact on Runway revenue margin burn by just clicking on the thing. There is no AI required there. I think what's more interesting though is, for example, when you have a bunch of possible potential tickets you can work on and you wanna know on which one should you work on first, and there's 100 of them, then we don't actually do this today, to be clear. But in theory, an LLM can look at that, all of those things, and look at what is the most impactful based on ROI ticket you could work on. What we do do today is something interesting though. Which is, for example, if you have something like budget versus actual. So as your time progresses, where you thought you were gonna be in July a year ago is different from where you are actually are, right? And what, there is a variance there. Let's say your revenue is off by 20%. What you can do if you have all of your data integrated in through your model is drill in and figure out, okay, why was it off by 20%? Was it the pricing that we got wrong? Was it like the number of leads? Was it a conversion rate? And we have all of this information because we integrate with analytics systems and your data warehouse, all these different things. And through that, we can just tell you, right? And traditionally the way you would have to do is you would have to trace through your model by hand in a spreadsheet, which is really annoying and slow, but we can just automate all of that, which is pretty handy.

Yeah, that's cool. I can definitely see going back to the theme of automatically AI doing stuff in the background for you, creating context, or in this case it'd be maybe more creating connections or even helping manage. So many companies struggle with, are we doing things that actually advance our stated goals at any given point in time? I can imagine like a really interesting background process that could look at all the tickets and say, first of all, we hypothesize this might connect to, and then also, I don't know, I'm assuming there's OKR ability among those 680 integrations as well. Do these things seem to be connected to the top-level goals that we have set? Which I think soberingly enough for many Leaders in many companies would be like, not, they don't always, sometimes a surprisingly low percentage of them do. Yeah. So it almost sounds like this is, you're sort of angling not for a finance platform, but really for an overall like business management platform. Like how big does this vision get in your mind?

So LinkedIn, I describe it as disrupting this $70 trillion business industry. I think about it one another way. I do think about it and like Stripe's mission and vision is to increase the GDP of the internet. I think in a real way, our mission is to increase the GDP of the world because there is so much impact in having everyone in the company understand how the business works, what the plans are, what is impacted of their own decisions so that we want to live in a world where Imagine a designer, a PM engineer could say, if we work on X instead of Y, here's how it affects margin and growth 2 years down the line. You can't imagine having that conversation today. And that's really interesting. Like, why is that? Everyone works in a business for the most part. And yet we accept the, this reality that like, it's okay for very few people to understand that you can just be heads down and focus on your own thing. And we think that the evolution of information technologies, especially like since computers, has been making like very complicated things more and more accessible and understandable. That's what a spreadsheet did. And also that's what Figma did for design through components, right? And we think in a real, very real way, what we're doing is inventing new information technology to make One of the most complicated businesses understandable, and we think that has a really big impact. People, and you're right, like the outcome of this, we don't actually think of it as a finance platform. We market it today because that's the most understandable way to think about it. But the way we do think about this is a business operating system, and an operating system maybe sounds more dry. But like the way I think about it is like, what is the video game view of a business? My background is on video games, right? I was at Zynga and Sandbox because like, if you ever play Civilization or RollerCoaster Tycoon or like Pizza Tycoon or whatever, all these games, they're fun games where you run a business. So what makes that fun? What makes it fun is it's understandable. It's visual. It's interactive. You get responses immediately. You feel like you have a, you're building up a mental model of how things work and you can have some agency and control over it. Businesses don't work that way today. And it totally can if we have better tools for thought that map to how people think about their businesses. And we would argue a spreadsheet is not the best form of that.

Hey, we'll continue our interview in a moment after a word from our sponsors. Yeah, it's a really interesting comment on video games and how the sort of business intelligence in video games is always 10 out of 10 in contrast to real life where it's so often far from it.

Yeah.

So I love the hugeness of the vision. If we reel back to today, I was just at the AI Engineer World's Fair last week, and the big topic there is like, Demos are really easy. Making stuff that's actually reliable in production is hard. That's the chasm that so many businesses are trying to cross. Where are you guys in that journey today?

Yeah.

When I generate a scenario, does that come along with a, you better fact-check this sort of warning, or do you feel like you have it under control enough that you can be confident that it's pulling the right levers or things are executing properly? You said earlier it doesn't always work. I don't know if that means there are errors or just sometimes it can't converge on your goal, which would be 2 different ways in which it might not always work. But to what degree have you tamed the unwieldiness of language models in this product?

I would say we have not tamed it very well at all. This is early, and I think the model's capabilities aren't quite there yet. It hallucinates; it will misunderstand; it will get things wrong. I generally would not trust it out of the box at all, to be very clear on that particular use case. And so that is like our version of the shiny AI demo, and like it will get good as model capabilities increase and orchestration gets better. But I think like what people. And what we're focused on now is like, how do you get value out of it today? And the way we think about this is there's so much that you can do that leverages what LLMs are good at immediately. And so we have a few ideas. So we have this thing called explain mode. And what that does is when you hold down the control button and you mouse over any part of the product. It'll just tell you what you're looking at and explain it. And so that idea comes from this idea that you, you shouldn't have to like type and have a chat with the system to get value out of it, because most of the time you don't even know what question to ask. And so there is these ambient, we refer internally as ambient intelligence. There's these ambient things that just exist without requiring any user interaction. That takes advantage of known context, where you can make things more understandable, explainable. Another example is we have pull requests between models. Like when you look at that, it should just explain, and it does. What's different? What changed? That's something it's good at today, right? And where we are seeing value and it's like fairly reliable is just making things understandable and accessible. To people who aren't familiar with the product, aren't familiar with the business. It's very quite, it's quite good at that today.

Yeah, that's interesting. I read a lot of ML literature and these days my reading workflow is to take the paper, put it into both ChatGPT and Claude and ask for a summary in each one for starters. And then just go back and forth with kind of tinkering back and forth between the 2 models and just asking lots of follow-up questions. And I would say it is getting extremely good. And what's notable about this ML literature is for one thing, it's moving super fast. So like a lot has happened since the knowledge cutoff date at any given point in time. It's like by definition, not in the training data, but even just the overall gestalt or whatever of the field is shifting on like a month-to-month basis.

Yeah.

And yet I would say I see precious few outright hallucinations or errors when I'm working with something that is, this is the document that we need to focus on. So for finance, there certainly are these intricacies where I could see it being a challenge with formulas, drivers, exactly how things are calculated, but it checks out to me that it would be very good at explaining these sorts of things in the world today.

Yeah.

Given that at least among those in the know, those concepts are pretty well established.

I'm curious to see how my GPT will perform with that workflow. If you upload a PDF to Universal Primer and see what it does, I'm curious to get your reaction to that. Because the prompts that I have by default for Explain are very complicated technical prompts. Like, explain to me how transformers work in the context of LLMs, or explain how Navier-Stokes is derived and things like that. So yeah, I think that would be a pretty good use case. I'd love for you to try. Okay, cool.

I'll check it out. When you're in discussion with Sam Altman, as one is from time to time, and they ask you, what do you want GPT-5 to do that the current models can't do? I'm being somewhat silly with my premise there. I don't know how often you're in discussion with Sam Altman, but they talk about—

Not much at all.

Abstracting away from that, I was just joking, but the They've said in public forum a few times that one of the ways that they evaluate companies that they're interested in is the companies that are like, we need GPT-5 because we need it to do X that your current model can't do. So with my silly premise aside, what are those things for you? It sounds like it's this scenario planning, but can you be more precise or detailed in terms of, like, Where it's failing now that you think GPT-5, quote unquote, might do for you?

Yeah. What is known about orchestration is that one of the biggest lifts that you can get through better orchestration is reflection, right? Look at it. How do you feel about it? Is it good? Is it accurate? How would you change it? And then come up with a better prompt. That I think is the most important thing to get better in a new model. And I think it's probably likely to be the highest leverage thing. And actually it crosses into my personal definition of what AGI is. As the models get more capable, it should be able to look at existing outputs and figure out why it was broken and how to make it better. And below a certain threshold, you just get nowhere, right? Like it just devolves, it hallucinates, it loses track. I personally think that is the biggest thing I want to see. And that reflection capability, I think once it crosses a certain threshold, is basically the barrier towards AGI. If, if it crosses a critical threshold where it can continuously look at its own orchestration, look at those prompts and just get better, then you approach it self-bootstrapping, right? The separation between an LLM and AI system today is basically, it's already, I think, smarter than many humans. You ask it questions and like probably half the human population couldn't come up with as good of an answer for any given problem that you ask ChatGPT. I think it's like more capable than people give it credit for. But I think what we really want to look at is on the extreme end of capability for a human being. And I think one of the more extreme ones are the human beings capable of improving the LLMs themselves. And once an LLM or properly orchestrated some kind of agentic loop can get to that point, then we're there. And I think that's basically the downstream of better reflection capabilities, which in turn is driven by like planning capabilities, reasoning capabilities, and things like that.

Yeah, it doesn't seem like we're too far off. I definitely agree that all things considered, one of my kind of standard talking points is that it was actually, in my view, as of summer of 2022 with the PaLM paper, that the frontier model passed the average person on the average task. I evaluate that with PaLM on BigBench. And now I would say that the models are closing in on expert performance.

Yeah.

On routine tasks. But the word routine is doing a lot of work there because obviously the models really excel where they have lots of examples to learn from, and comparatively they struggle when they have to figure out entirely new things. So that's exactly the shortcoming that you were speaking about.

Yeah.

Before we return to that and project into the future a little bit farther, the UI that you mentioned around just mousing over and hitting control within the product, I think, is a really interesting one and one that maybe a lot more products should adopt. That also seems like the kind of thing that might not be secret sauce, that you might be able to tell us a little bit more about how that works. I'm curious, is that like a— did you have to go through and take a ton of screenshots of your product and just manually label them and throw those into a big prompt? Or any lessons learned there? Because I think a lot of people would actually do well to implement something like that if they could figure out how.

Yeah, it's mostly a design problem, right? And I think like when you look at the way Apple implemented intelligence in the iPhone, they never said AI, artificial intelligence, just like Apple's intelligence in general. It is very similar to how we think, been thinking about internally for a while now. The, it's not a very complicated multimodal science fiction thing. We just literally trace through the dome of what the object tree of Whatever your mouse is looking at. And we say, what is this thing? Is it a driver? Is it a report? If you're looking at a modal around a pull request, we just, we know you're looking at it. It's already pre-generated. And so we define the key entities in the product, whether that's a page, a report, a pull request, a driver. And we say, okay, let's generate some explanations and just show it to people. So it's fairly uncomplicated. But You know, people have, I think, systematically undervalued design, right? What Apple did was interesting, but there was not a whole lot of new foundational model capabilities there. It was just a very well-considered design, and that's what made the iPhone work. And in general, that's one of the bigger opportunities in applied AI today. Everyone's like looking for the more capable model. The better fine-tuning orchestration system or whatever it is. And reality is you'll have a few AI labs and they'll have the most capable models. And there needs to be a lot more work in figuring out how you can use these new AI capabilities and express them to customers in useful ways. And we barely scratched the surface of that. And right now the bulk of that is Chat with a bot.

It's a profound point that if the surface area isn't that big, a lot of times things don't have to be nearly as complicated. That's my takeaway from your comment there that like, just work harder on defining your problem and mapping out the internals of it. And in many cases, like you can be pretty comprehensive about that. And then you've made life a lot simpler on whatever, whether it's a user or a language model, just do that hard work. It's finite and it can really pay off without having to invoke any science fiction. I think that's a good upshot from that. How about just to sneak in a couple more practical things before we do the big picture questions? I've seen you comment in a couple other contexts, like very aggressive early adopters of ChatGPT and other AI tools, For operations within your own business, what are like your favorite products, your favorite use cases? Where are you getting the most value from applying language model or language model workflows in your business today?

Yeah, we, the good news is there's a lot more ready-to-use tools out there than there was a year ago or even 2 years ago. But we've always thought about AI as an incredible tool to increase increase the leverage of our company. And so we have this internal goal of never having more than 100 employees at Runway. And it's possible today because you can scale with increasing capabilities of AI, right? As you mentioned, like we're maybe not that far off. If that's the case, you have a very small company where everyone is like just so empowered, has so much leverage and so much freedom and ownership over what they do that they can create so much more impact than at a large company because the biggest factor in making like work less enjoyable is just like more people, I think. More people, more problems. So we've always thought about and been excited about it. 2 particular use cases, like we've talked extensively, I think externally about the AI SDR that we built when our website went viral and we had so many leads coming in, we had to qualify them. So it's a very simple automation. Have a new sign up. We enrich it through Clearbit, and then we have a prompt to qualify this lead whether they're using a free email address or a work one. What kind of work do they do? How big is the company? How is it growing? What's their role? Are they a key decision maker? All these different things. And the wonderful thing is we can describe our qualification criteria the same way I would describe it to SDRs. Roughly, we want like companies that are growing and pass a million ARR. Want them to be a key executive decision maker or in finance. They should be largely based in the US and ideally more than 50 people. And it'll do that. That's a pretty simple thing for an LLM to understand. And it does a great job of filtering. And then once we think they're qualified, then we, and this is where it took a bit more time, we developed this prompt that where we gave them context about what we do and our value proposition. And what we know about their company through their website and all the other contacts that we have, and say, please draft an email responding to this customer and make it very clever. Don't make it cringy. Don't make it mid. Make it like, just connect what we do and what they do in a really clever, novel way. And it took a lot of tweaking to make it like sound reasonable. And the out— The outcome is that when our sales team wakes up in the morning, we automatically post that draft in the Gmail drafts folder. And so they wake up and in their drafts folder is a bunch of pre-written emails, all pre-qualified, all pre-customized, and they just press send. So that's one of them that's interesting. More recently, there, I do like cultural interviews with candidates and I was taking a lot of notes as I'm like looking for certain traits in candidates. And I was thinking about, okay, like I don't really want to take these notes anymore because I want to pay attention to the candidate and have more of a connection instead of typing. And I know Grain records the video. And so what I did is I just made a GPT that's internal and I downloaded a transcript, paste it in, and just gives me the notes. In exactly the format I want, pulling out exactly the things I'm looking for. It doesn't do a very good job of making the hiring decision, evaluating it, but it's really good at taking the content and extracting out like the key things that I was looking for and the angles I was looking for around motivation and values and things like that. That's been like incredibly convenient for us.

Do you use no-code tools to build those things? I feel like a lot of software companies are reluctant to use platforms like that, but maybe your example will inspire people to build more on existing platforms and do less kind of code it all from scratch.

Yeah, we do both. So for the 2 things I mentioned, one's just ChatGPT, the other is Zapier plus Zapier automation, but we also build other tools like that's written in code depending on capabilities, right? I wanted to build a thing that will leave feedback on documents in whatever persona that people wanted to. It's, I want Jeff Bezos or Steve Jobs to critique this thing, my Notion document. And that required code. I wrote most of the code through with ChatGPT and understood how the Notion API worked. And I said, I want you to monitor like all this entire workspace and look for mentions You, this bot, and when it says like, the way it works is you call @jam_session with whoever, with Steve Jobs, right? You just leave that as a comment on any document. And then the bot will come back in the voice of Steve Jobs and excerpt out the critical parts of whatever document and leave feedback critiquing it however you like. So it really depends on the use case. We're pretty pragmatic if it's possible to do and quick to do with no-code. code tool, we'll do that. If we need to write code, we'll do that too.

Do you find that those personas are best delivered by any model in particular? I've recently been experimenting with the RLHF models versus the purely pre-trained models, and the pre-trained ones we have access to are not as powerful, but they do seem to capture the stylistic elements, at least of a lot of different personalities, a lot more closely. What have you found there in terms of how do you actually get the real value of the Bezos or the Jobs perspective?

I find the prompt has so much impact on the quality of it. The prompt that we, I think we share publicly is really long and you have to work really hard. You have to put every trick in the book. For it to get like some interesting insight out of it. We say, do not say anything generic. Make it like the most interesting novel insight, even to an expert. And talk in their persona, but not in their media persona, but look at like their internal persona on how they actually speak. And so the more specific detail we add and the more stringent we are on the quality and insight requirements. Interesting.

I'm really very interested to see what Llama 3 400B does in that regard, especially if they do release the purely pre-trained version. They may or may not in the end, but if they do, things will start to get a little weird, I think, in the impersonation game.

Yeah.

I think Jimmy Apple said that's not gonna happen. Right. And I think strategically I would probably bet on it not happening.

How about getting any models to write as you? You're obviously a prolific tweeter and I've heard that's good for lead gen for your business. Folks internally want more tweets from your account. Have you gotten anything to write as you in a way that's compelling to you?

No, I don't think we're quite there yet. And I think this does speak to Maybe a cutting-edge limitation on models. I think maybe it speaks to probably more a lack of context than anything else. The problem with having an LLM generate interesting content or just in general a good piece of writing is that good writing isn't about the writing, it's about the ideas. And the most interesting ideas tend to be novel a lot of the times or phrased in some kind of novel, interesting contextual way. And that gets at the tail end of the distribution in terms of training data. I think more data, more reason is going to help, but I think more context and some way of continuous learning and interacting with the world so they can form these experiences is all going to help. And I find working with a ghostwriter, so like I actually have a ghostwriter on LinkedIn because it's a very different style. And it took a long time. And the main challenge is it's not the writing, it's imparting the lived experiences, the stories, everything I've seen, and then expressing that in a style that's like very difficult to get to. And when working with a ghostwriter, you're basically offloading a lot of context. And you probably will need to do the same thing with LLMs. But the great thing about LLMs is that You can build a lot of context just by connecting it with a bunch of existing data, which will be much harder for a human being to do. It'll take longer. And so as we have more context and more and better capabilities, I think that'll get better. But today it's not there. I would not like my content if I want it to work. And you've seen this, right? Like you can tell when someone posts something with ChatGPT, it's very recognizable.

Yeah.

It's very mid, and that's probably not going to work.

I'd say that's mostly been my experience too. The one exception that I've had with that is when I write the short intro essay that I put at the top of each episode of the podcast, I now take 30 of those that I originally wrote by hand before Claude 3 existed, and then provide also the transcript of the current episode. And ask it to write a new intro in the style of the 30 that I've provided, using, of course, the new content. I would say for 3 out of 4 episodes, I could probably just use that. 1 out of 4, I have more of a strong opinion or a take or perspective. If we're talking about regulation, we're talking about something where I want to weigh in in the intro, then it doesn't get that right. Although it can still write it pretty well if I instruct it upfront of here's what take I want to infuse. But if it's a more like standard episode where we're talking through research result or we're talking through a product, it actually does a pretty good job. And it is uncanny in picking up on my written mannerisms, which is a bit of a funny experience to be like reading what really does feel like my own voice coming back at me out of an LLM.

Yeah, that actually totally tracks. It's probably if you can Give it enough context, and what you're we're saying is you give it a bunch of examples, and it has a context over your current episode. Can it like write something reasonable most of the time? That makes sense, and that's probably like close enough to good for something like a LinkedIn. I think Twitter is really different because what context do you give it for any given shit post that is it going to be a banger, right? That's. Timely, it's new, it's kind of clever, it's surprising. And so I would say that works pretty differently, but generally, yeah, if you have context and you prompt it, it can do a pretty good job.

So last couple minutes, zooming out, big picture, you're on record with an expectation that AI capabilities increases are not at their end and expecting, it seems like pretty powerful AIs to come online and then not too distant future. I don't want to get bogged down in specific timelines or whatever, because obviously that's always got a confidence interval on it. But I am interested in how you think about making your investments in light of possible AGI soon, how you think about raising your kids in light of AGI soon, and how you think about your company in light of AGI soon. All these sort of most fundamental activities in life seem to be Under the specter of possibly radical change that I think about this basically nonstop, and I still am more at a loss than not as to how I should adjust my approach to life. So what, if any, changes have you made to your approaches given your outlook?

Yeah, for investing, I talk in my book in a sense that I think the biggest opportunities in AI is proactivity and context. So I'm looking for companies that take those 2 pillars and drive it forward in new product expressions. So things like an AI chief of staff, things like AI SDR platform. So I invested in Automatica, which does that. I invested in Maestro, which is AI chief of staff. I think about, I think those are still like very interesting opportunities. So mostly it's on the application layer. Although I was a small investor in ElevenLabs and some other companies. But I think there's going to be relatively few new foundational model companies and most of it's applied. And so that's one angle I think about. How does AGI affect that? I have the same answer for investments as I do for the company. So how does AGI affect the company? And I think on some level, all bets are off. Once you get to AGI, it may be a very short shot to ASI. And I actually don't think about it that much. The reason why I don't think about it that much is in that world, my company, in terms of just absolute impact, the impact on me is probably way smaller than the impact on Microsoft or Amazon or literally every large company. Because what we're talking about is, okay, like it can do all forms of productive work. What is then the work that people do? What is the nature of work? What happens to my specific company is almost like irrelevant when you have like very much larger societal level questions that's going to raise. The one thing I do think about is in an AGI world where you have like agents running the world, can there be a runaway agent and what does the wrong agent to. And I think there's an opportunity to have like effectively a CEO agent who understands like the entire business context, because that's a really interesting, unique source of data. There's very few companies that have a holistic view of everything that happens in the company. I think that's one interesting direction, but by and large, I think the societal effects are just gonna be like unpredictable. So I don't think about it all that much. In terms of the family bit, I think maybe this does connect to the company bit. Which is I advise my kids to be makers. And this is what I would do before AI anyway, like learn to code, learn how to make things with their hands, learn how to create games. And it's wonderful. Like over Christmas, we were traveling with a friend's family and I have a 9-year-old and with, there's another 9-year-old, they were They wanted to create a game that's a cross between Transformers and Gundam, and they asked ChatGPT how to do it. ChatGPT generated art, they generated the rules. They asked them, how would you make it? Get Unity. They downloaded Unity. 9-year-olds, right? And they have this thing that like will guide them step by step on how to do things and it'll respond to their questions at their level. It's wonderful. So they can learn how to make things, but like The question then is like, what is the value of making things if AI can just make it from scratch, from start to finish instantly, right? In the future. And my point of view on this is basically things that are provably made by humans will still be valuable because I think what Charlie Munger observed about the world is something I think about a lot, which is people think that greed drives human progress. But it's not greed. It's actually envy. Like comparative, like social competition, sexual competition in societies is something that has existed since people or animals have existed. And so we are going to be endlessly creative at finding new status gains to compete against, right? Even in a world, in a post singularity world. It's still gonna be about like, how can I demonstrate I am the most interesting, fit, creative, smart person out there? And we're gonna figure out new rules. The rules might be the things that's gonna be most valuable is like whoever can paint the best painting by hand, right? Like that's what art does. The software's gonna be valuable is the software that is not created by AI, but the one that was provably coded by hand without AI assistance. We see that today in the sense that if you watch chess, you don't watch AlphaGo playing against AlphaGo. No one can play better chess than that. You still watch Magnus Carlsen because that's interesting as a human. And I think that dynamic will play out on a broad scale. And so there's always gonna be something interesting for humans to do unless AI kills us first.

On that note, that's a very common answer. Actually, it's basically the same answer that I have. Which is my version of it, is just like playing toward the range of outcomes that I feel like I have some ability to influence. If there's AGI, ASI soon, yikes. I don't really know what I can do. I don't think about it that much either. Do you think, though, that means that governments should start to be thinking about this and start to be working on international treaties or whatever? Or would you cast your lot with just kind of technology taking us where it might, how much do you think we should try to be intentional in shaping or controlling versus just letting it happen as it happens?

I'm generally skeptical of government intervention. I don't know that we know what the right move is. The thing that's already happening is regulating compute. I don't know how I feel about that. How do we know this is good? How do we know that's the right thing to focus on? How do we know that with the current amount of compute and right architecture, we can't just get to ASI anyway. So I think there's a lot we don't know. I generally think the— I'm bullish on mechanistic interpretability. And I think really the things that are going to solve it is just better understanding through technology of how this works, more so than government regulation. Because there is no doubt that to me that regulation will happen. I just don't know that the right thing is to get ahead of it when we don't know a whole lot about how the world's actually going to change.

Yeah, I'm pretty uncertain about all that too. I'm also very, generally speaking, sceptical of government regulation. In this case, I feel like I might rather go down trying than not trying. But it does seem also at the same time like we very well could have a lot of unintended consequences or even negative consequences. And I definitely share your enthusiasm for it. Yeah.

I mean, like one worry I have is if we believe this is gonna be as powerful as we think it is, regulation then puts you at a disadvantage against like less favorable actors. Right. And that's true on a nation state level as well as a company level. So if OpenAI decides to be super safe Then some other research group is going to not worry about that and just go faster. And they may be less benevolent. Same with nation states, right? If America regulates compute and AI research, another nation state will not. And what happens at the critical crossing of capabilities when it gets able to self-bootstrap? So it's a very difficult problem and the implication of regulation is like, may not play out the way people want it to.

Yeah. Again, my instinct there is if we're going to go down, I'd rather go down trying to cooperate with China than racing China to create godlike AI. But it doesn't mean you're wrong. It very well could go down that way. I know we're just about out of time. Any other either big picture questions or any other items you wanted to touch on that we didn't, or anything you want to plug before we break?

No, I think we covered a lot. But anyway, yeah, if you're a mid-sized company, pay for Runway.

I was really impressed with the demo. We're in the pipeline, so I am plus one at least in the lead column. We'll see what our key stakeholders have to say about that, but very cool product, really interesting application of AI in a high importance area, and I think definitely sign of things to come. So. Siki Chen, founder and CEO of Runway, thank you for being part of the Cognitive Revolution.

Thank you, Nathan. It's been great.

It is both energizing and enlightening to hear why people listen and learn what they value about the show.

So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.