[00:00:01] SF: All right, Amanda, welcome to the show.

[00:00:04] AK: Thank you. Happy to be here.

[00:00:05] SF: Yes. Thanks so much for being here. I've been really looking forward to our conversation. I’m recently become a little bit of a Streamlit user. I can't say that I'm an expert by any means, and that's why you're here. But maybe we can start by having you introduce yourself. Who are you? What do you do? How did you get to where you are today?

[00:00:21] AK: Oh, okay. Well, that's a bigger intro than I was planning. But Hi, I'm Amanda Kelly. I'm a co-founder of Streamlit. I'm also Director of Product over at Snowflake, a citizen of Northern California. I'm enjoying the weather and doing things here. I have a dog. I have kids. Those are things I do in my spare time. So, how did I get here and kind of where are we today? Well, I never really made kind of an explicit choice that was like, you know what, I'm going to focus my career on data and on AI and on applications. It just kind of naturally evolved, where I think by going after more and more kind of interesting problems and meeting interesting people, doing new things, and just wanting to help there. 

So, my journey, really, to Streamlit and where we are with Snowflake today, started after grad school when I went to Google X. And Google X at the time was just kind of a lot of random people who were thinking of a lot of ideas of the future and I happen to meet, who would eventually become my co-founders, Adrien and Thiago. Through that, I ended up working on assistant, next generation assistant technology, self-driving cars, cybersecurity, and over the next few years, where I put them and then others, and really kept seeing the same problems kind of over and over again, that we really were not empowering data teams, ML teams, people who are working at the forefront of what you could do with data, to do the work that they needed to do, and to have the impact, kind of in the broader company. So, it was really kind of out of those frustrations that I ended up where I am today.

[00:01:46] SF: Awesome. I see there's probably been thousands of startups that have come out of people working at organizations like Google, or AWS, any sort of large company where they have some – they're counting some sort of pain in what they're doing, especially at the scale that you might be doing at one of those organizations. They kind of like realize, “Oh, like I need to figure out how to solve this.” And that essentially leads them to founding a company and hopefully going on and being successful. 

[00:02:14] AK: Yes. There are many of us. We are legion, working with data, and then the problems that we solve. So yes, it's awesome. It's very exciting, I think, to solve a problem that you have yourself, because that's just on its own, kind of sense of satisfaction. But I think one of the most amazing things kind of in my journey over the past few years is just finding out how many other people there are, that kind of do the same things every day and have the same problems, and getting to be a part of changing some of their lives and having them change mine as well.

[00:02:40] SF: There's this, I think, concept around a date application or this term that we kind of throw out there. So, in your experience, or how you think about it, what is a data application? And then how is that sort of different than other types of applications that, other types engineers might be working on? 

[00:02:59] AK: Yes. I think it's different in two ways. One of which is the fact that it's the purpose, right? That it's coming from data and for data. The second is around the people. The people that are creating it, and the people that are consuming it. So, when I think about a data application, really, its purpose is to do something with data, right? It might be a tool that's working to transform some data to write some data back, or it might be more about kind of consuming an insight or an exploration, and driving deeper understanding, getting to a decision, for the business that can be made right on top of that data.

That I think, it a smaller scope than all of the web applications of the world, but it's really focused on data, and therefore, it has some different properties in terms of how you connect, what you need to do with the data, the types of things you want to build on top of it. But the other really key distinction is about who's using this. The first is that these are applications built by data teams themselves, right? So, it needs to work in the workflow of what a data team is doing, and the types of skill sets that they have. Also, who it's for, right?

A lot of times it's for themselves. We see a huge amount of data teams just building things for themselves to unlock the work that they are doing. But also, who is going to be consuming that? A lot of times, you have data teams, you have IT teams, you have a number of these groups that are working with data, and their purpose really is to empower the rest of the organization. So, the needs of the marketing, the sales, the finance, the product teams that are there consuming this are also really important in terms of what makes a good data application.

[00:04:33] SF: Yes. We think that, this isn't true, necessarily, of every data application at every, I don't know, product application, or whatever you want to call like, maybe a more traditional application that we might be thinking about. But that scale on the end users is different, where the data application is probably going to be a smaller scale than users. But then, the amount of data that you might be actually processing it to power that data application, that's kind of where the scale might come into play.

[00:05:00] AK: Yes, absolutely. I think that it goes. It needs the entire range, right? It’s since an application, it's just, this is something that would be better not run in a terminal. Not screenshotted out and put in a PowerPoint. But it's better to be run kind of with the UI on top of it. It's better for it to be interactive and that could be for a number of reasons.

One could be that it's just going to be better to do this with a drop down of the button, because if we do it by writing the code into the worksheet or the terminal, it's going to increase their error rates. So, it's going to make it a less joyful experience, for everybody who's running these processes. But a lot of times, it's also just about, like, we need to see these things in a visual manner. Printing out the data in a table is not enough, to really kind of convey, “Hey, it's this area of Philadelphia that's really surging in rideshares.” Those are things that are really visual and are much better done in an application. That can often be really serving one or two people. There are applications that we have internally, which are almost entirely made for me, in terms of specific decisions I need to make as a product leader. Then, we also see things that are scaling out right to the entire company, and have some of those basic needs that you have for web application in terms of concurrency, and load and things like that. And absolutely, the amount of data being processed is a huge part of that.

I think that that's part of the broader, when you think about serving a data application need well, given that as the data teams building this, we need to – those of us that are helping this, need to not just be building out in like, “Here's a nice chart for that”, but also solving those needs. So, that the people who are building them don't have to consider what does it mean if a thousand people come and try to alter this chart at the same time.

[00:06:43] SF: You talked a little bit about the idea of empowering data teams, and also some of the limitations that you may have, if you're just generating a static chart. But what are some of the major challenges, essentially, that data teams run into when they're either – you essentially, trying to access or make decisions about what's going on within their organization?

[00:07:03] AK: Yes. So, the basic problem is data itself. Your data is in silos, it's in different formats, it's really costly to do this type of thing. The engineering team didn't implement the telemetry properly. So, we don't actually have what we need, and now there's a long process to do that. But even beyond that, I really think that data teams in general are undervalued in organizations. People don't understand the work and how hard it is. When you're like, “All I need is a basic chart that tells me who did X.” The actual work that goes into that, and how difficult that can be.

But also, I think they're undervalued in, in what they could be doing. I think that in most organizations, we're only seeing the very tip of what your data team could be doing for you, in terms of the depth of insights, the way that we could really be empowering, marketing and sales and product decisions. If these groups worked much more closely and really understood, what is the choice that marketing has to make today? Ultimately, that has to be done, or that can't be done at the highest level. It just doesn't scale, because the choices that are needing to be made that we want to act with data or down to like an individual salesperson. What product set should I recommend to this specific customer based on what we know about them and what we know right about other customers?

So, we have to be empowering that, and I think some of the best insights that are coming out, and we'll probably talk about this later. But especially with Generative AI, they're coming from the layer of people working directly with data. It's not the people sitting in the C suite. So, we need the applications, and we need the ability to empower those people.

[00:08:37] SF: Yes, that makes a ton of sense. In my day job, I work in marketing today, and there's a lot of decisions that someone who's maybe running campaigns or running events, essentially, at that level and needs to make about where they fund, essentially, resources or additional campaigns. If they don't have access to the data to make informed decisions, then they're basically using hunches or prior confirmation bias or what they heard from a sales rep. And then there, there's bias in terms of whatever the last conversation they had from that lead source influences, essentially, their perception of how valuable it is. That leads to a lot of, potentially, incorrect decisions and a waste of resources.

[00:09:16] AK: Completely. It's not like you as a marketing person don't know, right? And you're just like, “I'm just going to go off my gut today.” You know the data that you want. You want to do it. It's also not like the data team doesn't want to give you that. The problem is that it's so hard and costly, generally, to go from that very specific, “Hey, I need to know this today.” To figuring out how to scale that. That, I think, is the ultimate thing of how can we make this so easy, so fast, so cheap, for these groups to collaborate together on data, that every single one of those questions that you have about what you should do in that campaign, what you should do with that customer, can be data driven.

I remember, when we first launched Streamlit as an open source project. I mean, you’d never really know. Sometimes we launch it, like how it's going to get used. And we were so excited when we got contacted by somebody out of the community who happened to run ML and data science for a rather large convenience store chain. He was telling us all about how they started using, our tool, Streamlit, to build out these tools, so that marketing could go in and they could say, “Hey, I'm trying to push coffee.” Or whatever it was they were trying to do, and be able to generate email lists, based on models that they were running that said, “Hey, if somebody bought this, they're likely to buy this.” These specific ones would like this type of coupon. These, would probably like these types of coupons.

They just were never able to do that before, because again, it was just so costly to have to build that up, or to work directly with every marketing person, that they couldn't do it. But the ideas were already there. I think the ideas are there in most companies of how we want to run better and how we run faster. It's just a question if we can empower people to do that.

[00:10:54] SF: You mentioned, that when you started Streamlit, being surprised, maybe by some of the uses of it. But what was the original problem that you were trying to address by creating that product in the company?

[00:11:06] AK: Yes. So, like you were saying earlier, it really came from a need that we saw working at Google X, working in self-driving cars, working in other kinds of places where we just kind of felt constantly hamstrung by the tools that we had. A number of times, we ended up in a conversation. I remember very distinct conversation having with some of our computer vision team, when I was working on self-driving cars, where it was like, “Well, we're getting all of – we have this new sensor type. We’re getting all of this data coming in and we needed all labeled, and we need a way to do this.” Our option was basically to go with an internal tools team or go with a very expensive vendor to build a tool. It was going to take months, and we really needed that data tomorrow to increase in the goals. We needed to get it to people who are offshore who can label it.

So, what ended up happening was that head of research ended up in going and learning Flask, over a series of weekends and writing this tool. The tool has so many bugs, but it was so much better than what we had otherwise. It's kind of scary, if you look inside any company, that number of kind of bespoke and bug ridden tools that underpin a lot of our critical processes. But we just saw that over and over again, right? Both this need and this desire of like, “Hey, I'm a smart person who knows how to code. I should be able to do these things for myself. I want to take my destiny back in my own hands, and not have to wait on external teams, or having to learn a completely new skill set, in terms of front-end programming.”

Then, also, this kind of related need, which is I want to work with my stakeholders better. I ran a – there's a couple points of my time in my career where I've run different data science teams. I remember the first data science team I was running years ago, our process would look like a bunch of executive stakeholders would give us a whole bunch of questions. We'd go and do a whole bunch of analysis. We prepare like a 40-slide, slide deck, with all of these charts, and answering all these questions. I swear, we would never get past like the second slide. There were always just so many more questions, right? That the first time that they saw that data, wait a second, why is this spiking here? Could we look at it across country? Could we look at it – all of these different types of questions that we're getting to these underlying needs that they had, as product and operations executives to understand where they could optimize, what different decisions they could make.

But it was so broken. We'd have to take all of those questions, then go back for another two weeks. And we come back, and it always felt like we were behind. I know a lot of data teams feel like this. We feel like we're the bad guys, that the product and marketing and sales could be running so much faster if we just had a better way to do it. And I think that the most pernicious thing is people stop asking questions. They just learned that it's not going to get answered. It's going to be a P2 and they stopped being curious and about how we do our jobs better.

[00:13:48] SF: Yes, I mean, I think that when you introduce a lot of friction into the process, or extended timelines, then essentially that creates a higher bar for when you want to put forth the ask because you're now, you know that like, “Okay, well, I'm like setting somebody off on a two-week bespoke request to get me the data.” And I've been there as well, a number of times, either when I was founder in a board meeting, or meeting with investors or whatever it is. And people ask a question, and then you don't have the answer readily available, or you haven't pulled the data. You have to essentially go pull that data later and come back. It leads to a lot of friction, but it also kind of makes you look bad as well, because you don't have access to it, and you hadn't thought to ask that question in the first place.

[00:14:32] AK: Absolutely. We end up self-censoring, and that means that the data team knows less about what we actually need. Because we're like, “I don't know if this is big enough. I don't know if anything's going to come of this.” For years, I remember talking to the A teams where I’ll be like, “Push back. Push back on your product partners and say, ‘Are you going to change your mind if I get you – if this is 1 versus 100?’ If I get you this answer, because it's going to cost me 10 hours to get to that. That's valid advice. But at the same time, I think that, again, it really stifles curiosity. And I think so many – I can go back to so many times where I changed my mind about what we want to do in the product or where we wanted to focus the company, because of some insight that we wouldn't have gotten to if somebody wasn't curious, about diving down and being like, “Huh, why is this area much bigger than this area?” And then one level down, and then one level down. Ask those five why's. But you just can't do it if everything takes 10 hours. Expensive engineers time to do.

[00:15:31] SF: Yes. I think the rigidity of whatever you're that you're using to investigate becomes a barrier, and essentially being curious, because there's just, again, too much friction. I think, one of the things that I think, historically we've struggled with is essentially, the existing tools are okay when you know what questions to answer. So, I spent the better part of a decade in the earlier part of my career, like building data apps, in particularly, in bioinformatics space, and there was tools like Tableau and Looker, that are great for sort of investigating, you know, maybe a well understood problem. But then there's many problems, especially in the world of bioinformatics when I was there, when I was a founder, that you don't even necessarily know what question to ask. You just want to be able to investigate.

So, can you maybe dive a little bit into how Streamlit can help somebody address something like that? How do they actually – how's it helped with this, essentially, someone being curious and diving into data so that we can figure out like, what questions should I be even asking?

[00:16:35] AK: Yes. It's a great question. We obviously say that Streamlit really attacks the longtail. That the real insight on Streamlit was that we really needed a frontend and an interactivity layer on top of Python, to enable this curiosity. So, Streamlit, in many ways, people say, “What can Streamlit do?” I say, “Well, what can Python do?” Right? Because our goal really is to just write on top of that, and write on top of the work that you're already doing. In Python, the questions that your team is already asking. So, what Streamlit in essence, and maybe we should say that. Is that Streamlit is an open source library. That that gives you a number of different out of the box primitives that makes it really easy to assemble the data. You can just drop and markdown, write a chart type, write a slider, a widget, right? You can put these all together. Honestly, you don't even really have to know Python to know Streamlit. Streamlit’s incredibly basic, and I see apps all the time that are 95% SQL, right? They're doing all of the hard work there. Then, at the end, they're writing it to a data frame, and then you're assembling kind of the app pieces around that.

But what's really, really important, and getting back to your question of how do we enable curiosity, and figuring out how we get to these better answers, and make better decisions as a company, is that ultimately, what Streamlit is doing is they're making it – we're making it cheap and easy to build a data application. And we're making it really, really easy for it to be bespoke to whatever you want. You've got another question, just add another select box, right there. Drop another chart. It's easy to kind of flow along with you, kind of in that thought process.

One of our initial kind of vision statements was can we be generating visualizations, interactions, like at the speed of thought, right? As you’re asking a question, and you're able to type up the code, can we be generating it and putting the interactive UI on top of that. So, what that means is that it's no longer you and marketing, asked me as a data science question, and I've got to take two weeks, or go out to a separate tools team to do it. As long as I have the data basically, there, I can go ahead and start surfacing that to you. Then, it becomes a really interactive process where we can refine that together. 

I've been working in data for years, and rarely do I get the question right when I go to the team. I know things about our data that our team doesn't even know, because I set up some of the initial things. But still, often they send me something that I'm like, “Oh, wait a second. Actually, can we filter it this way? Can we look at it that way? I'd run love to see this on a bar chart. Can we see this on a scatterplot?” Really kind of trying to dive in and kind of be one with the data, and that's really, really important to me, as a product executive, to really understand what's going on the product. And obviously, for you in marketing and understand the product and the customers and sales.

I've been in meetings, I've seen this with not just our own team, but I've sat in on kind of customer meetings where you'll see a customer ask a question, and the data team literally in the background is able to go and change. They're like, “Oh, this is great. But I would love to – can we see this more stacked to the bar chart? Or, what if we segmented it by this type of thing?” Literally, by the end of the meeting, they're able to show that. That speed thing, it enables something different in the relationships you can have with customers and stakeholders, because normally it would be, well, let's get back to you in two weeks. In the meantime, you've gone ahead and made different decisions on what you're going to do.

Instead, we're having that conversation in the meeting, and saying, “Oh, wait a second. Now, that I see it this way, I have a different question. Or I have a different decision that I'd be able to make.” That's what we really see Streamlit is empowering, is this really fast, iterative loops, between data teams. Among themselves sometimes, but then also with customers is with stakeholders. What that does, giving them especially that interactivity layer, is it starting to unlock more of the data that's in the heads, of all of your operators in the company. That's some of the most important data that you've have as a company. It's not in your data lake. It is in the heads, of people, who know who just talked to a customer, and heard about some new competitor that's coming up that saw somebody, do a demo of something, and they understand they're using the product in a different way. Who have some knowledge based on a lot of things that they're seeing about where the economy is going, or where supply chains are going. And being able to kind of partner with the data team to kind of join those intuitions together, I think, just makes for much faster loops, and much better decision making in a company.

[00:21:01] SF: Yes. It makes a lot of sense. I think that simply by having this near real time, interactivity, you can make data exploration like a brainstorming exercise, isn't something where it's like, you'll ask a question, come back, so you can kind of like stay in it. You can be embodying the data with multiple people sort of contributing, asking questions. That overall helps the investigative process, because someone from one department maybe sees something a certain way, or has some insight that's maybe not even part of the data, that leads to a new type of question, or a new type of thing to investigate, and you just get this cascading effect where you're going to drive way more insights from the data in a shorter period of time.

[00:21:42] AK: Yes. I live Streamlit, obviously. But we're just kind of a small API library, and it really is the data teams, I think, who are the heroes here. They're the ones who are changing how companies are doing it. We're just helping provide a tool that makes it a lot easier to work with their stakeholders. I was talking with a CIO of a large Financial Group a while back, and he was telling me how their IT team had built up, took the time to build out all of their pipelines, kind of a snowflake, and then to add Streamlit on top, so that now they kind of auto generate apps, and have a lot of different app types for types of questions that they know get asked to them.

Now, they have kind of after multiple months of doing this, a process where their partners come ask them a question, and within a day, they've got something that they say is about 80% correct. Then, they worked directly with that person to say, “Oh, can I change this? Can we do that?” Within days, they have something that team is off to the races and they're running. It's completely changed, how the IT group works right with their stakeholders, but also the IT team is seen. I mean, I think that if I even just say IT team to people or not, on IT, your mind is usually kind of like, well, slow. Tells me I can't do things. Says I have to go through a procurement process. It's not often a positive one.

I think it’s amazing that just by switching and giving people things faster, and giving them more and easier access to ask their own questions, it completely shifts how we work with this group. I think it also empowers the different groups to be more, I guess, like the decision making is more of a strategic decision-making process, than necessarily like a pattern recognition decision-making process or something like that.

[00:23:21] AK: Yes.

[00:23:22] SF: So, you mentioned that when you were describing Streamlit, that it's essentially this Python library. Even though you might be able to build a Streamlit app, not necessarily knowing a ton about Python. But what was sort of the choice behind making it essentially a library built on top of Python versus something like, I don't know, like a no-code tool, or like drag and drop, or something like that?

[00:23:47] AK: Yes. I mean, I think it was never a choice in the sense. It was just like, it had to be. I think, partly because it was coming, especially from Thiago and Adrien, who are some of the most amazing engineers I've ever worked with. They're like, “This is how I work.” And they're data people, they're working in Python. Python is the language of machine learning. It's becoming more and more the language of data. And that's how we prefer to work in terms of development. So, there's actually, exciting things that I think we're doing, especially at snowflake now that do take kind of the core of Streamlit and extend it, to something that's more no code, low code, and allows people who are not directly in the code to do it.

But ultimately, I just think it's so important that you have something at the Python layer, not only kind of work with developers, but to get that level of configurability, that you want. Because it doesn't matter that the best tools in the world, the ones that you spend millions of dollars a year on as a company, you still probably have a number of things where you’re like, “I wish I could do this, I wish I could do that.” And the only way that you can do that is by putting pressure on your sales team to put pressure on the product team, to build you another feature.

So, having that ability to program it in yourself, is so important. We can always add abstraction layers on it, but Streamlit will always be Python. It always has been. I think that the big choice was making it open source, and it wasn't always obvious to us at first, necessarily, that it would be or that we'd built a wrap around it. But it became increasingly obvious. If we want to be running with the best of them all, if you want to truly empower data developers, it had to be open source. Plus, all of our engineers told us that it had to be open source. So, it's always best when you listen to your engineers.

[00:25:24] SF: I mean, did that choice of going open source, I mean, how, usually that leads to sort of a community led, go to market? How did that sort of impact Streamlit’s growth? And also, maybe the decisions that you're making from a robotic sense?

[00:25:39] AK: Yes. I mean, it completely changed the trajectory of the company. So, we were about a year in to building Streamlit, when we opened it up and said, “Hey, there's this new open source library.” Really, it felt like overnight, the community jumped on this, and they were like, “Yes. This is what I have been waiting for.” I think the success of any kind of developer product, but especially in open source one, really comes back to the community, right? Because when you're developing, you're always – it doesn't matter how well designed the product is. Basically, within 20 minutes, you end up opening a Google tab, and you're like, “What's up with that? How do I do this thing? What's up with this error message?” And so much of developing good code and good applications is about being inspired, is about being kind of supported by that one person who's also seen that type of thing, right, and get you unstuck or can give you a better way, a faster way to do something.

So, I think that the real power of Streamlit, there's the library, but it really is in the community, and the fact that like to the Streamlit website, and you can search and see thousands public applications that people have shared back with the community, and they share back the code, and they write components. And then they share them openly all to kind of empower more developers to do more. That's what really, ultimately, accelerates everybody, right? By making it faster. So, it's not just about, you had a good idea, and we wrote a nice library for you. It's that you can go and you can start off prior work. You can copy in great bits of code that other people have written or find that perfectly tuned component that's going to work specifically for your biomedical 3D, molecule reviewing need.

[00:27:17] SF: Then, how is – since essentially, being acquired by Snowflake, how is this, I don't know, marriage between open source, and then, what historically had not necessarily been an open source focus company come together? I feel like Snowflake has leaned a lot more into community and builders, most recently, in the last few years, which is, I think, fantastic. But historically, it wasn't necessarily there. So, I'm kind of curious about how Streamlit has maybe played a role in that transformation, or what was that experience like?

[00:27:51] AK: Yes. So, Snowflake really has made a big shift over the past couple of years, and we're still in the midst of it, but really wanting to be a developer focused company. So, I think we've always focused especially on like SQL analysts, and data engineers and making life great for them, really extending into broader types of developers, ML, data scientists, people who are more on the code, and saying, “What does it mean to be great for you, on a platform layer, but all the other tools that we need to enable?” A part of that is being really well connected in the ecosystem and making sure that people can bring write their favorite tools in development.

So, we were lucky when Snowflake approached us a year and a half ago, right? And they were like, “Well, we see this kind of amazing synergy, that Streamlit could have with Snowflake.” So, oftentimes, when I talk about Streamlit, I talk really about the pain of building a data application, and why it's frustrating, and why you might have to learn JavaScript, or use the tools team. But for most data scientists, I sit right next to our data team. Their frustration really is in the data side. It is in the fact of how long it takes to run a query or the fact that we're missing telemetry doing that.

There is this kind of amazing partnership that we have now, to build on top of, this amazing data cloud platform, to make data developers and everybody's life a lot better. And as part of that, Snowflake really is embracing open source. So, they've thrown a huge amount of resources towards the Streamlit team, empowering us to do more than we ever would have been able to do, as kind of a standalone startup, and we're making investments and more and more things – every week, it feels like I'm in a conversation where somebody's like, “Can I open source this? Let's open source that.” Including Benioff, he's wanting to open source things left and right.

So, I think it's really exciting times ahead, both for Snowflake and Streamlit, but I think also for the open source community where we're going to be able to participate a lot more broadly.

[00:29:43] SF: So, I think clearly, there's obviously a lot going on in the world of AI and Generative AI, and Snowflake’s making huge, I think, like strategic moves in that space as well, which they did a lot of announcements and stuff summit, and there's been a continuous number of announcements that they’ve made. As we've been talking, you talked a lot about, like empowering people the speed of like access to the data, and a lot of these things, essentially, you can even continue to reduce friction, continue to reduce the barrier to entry by leveraging some of these Generative AI technology. So, maybe before we talk about the Streamlit’s role in the world of Gen AI? How do you see, Generative AI change or impacting the work that like data teams do?

[00:30:32] AK: I think there's no end, honestly, to the number of changes that we're going to see. I mean, we really are riding a giant new wave, I think that's going to change everything, honestly, and I think for the better, over the coming years. I work with a lot of people who go way back, and kind of competing technology, and they're like, “We've never seen anything like this before.” It's really exciting. I think, the first and most obvious one is how I think it's going to speed up just data development. The history of a lot of technological changes, is that it takes away a lot of kind of the rote work, the boring things, that we don't like to do right now in terms of what we need to do to clean the data, monitor our pipelines, things like that. So, just being a lot smarter than that. I think, to me, personally, one of the most exciting things is how well ChatGPT and a lot of these other kinds of LLMs write code, right? Which means that, again, you can start kind of from prior art.

So, one of the most exciting things, just personally on Streamlit, was just finding out how well you can go to ChatGPT and say, “Can you write me a Streamlit app that uses these libraries and does these things?” And it writes perfect Streamlit code, which means that now again, you're not starting from, I need to write 50 lines of code to get to a start of something, or you can start from something that already works, and you can be more of an editor. Someone who can operate kind of at a higher level, and then doing kind of the basic things that you want to do.

In general, I mean, it's going to make so many hard things cheaper. Just so much, of I think just general work that we do, not even as data teams, and companies. It’s just retrieval work. I need to find that document that mentions this team. I need to find this table or things like that. One of the things that LLMs are just fantastic at, is kind of parsing information, getting us to that point, that we have. Then, there's so much more, I think, promised out there, especially for iterative work.

Chat, for example, is just the perfect way to kind of do a lot of this iterative work. I often think about what we're building with LLMs. Now, these LLMs, is just kind of – it can be my data buddy, kind of in a sense. It's somebody that it's helping me refine, and I'm asking questions, and being able to do things and ask more questions faster. It's not always necessarily getting the right answer just enough, because my data team doesn't always get the right answer either. But those are things that I think are just getting better faster.

Then, from there, there's just so many more new things, that we can do. So, you probably know from marketing, and I worked in marketing a long time ago. But so much of what's costly is just like, we got to put on a blog post, and then what's the tweet for that? What's the image that you attach to that? What's the clip of the video, right? These are all things that you can just have an LLM generate for you now. Read the text, pick out an obvious image, and you can select from three. You can summarize it. You can take a tweet, and you can say, “Hey, look at my top users. Do you think this will resonate with them based on past tweets?” And what has been done and things like that. So, I'm really excited for what this is going to empower.

[00:33:32] SF: Yes. I like your framing around how it's transforming some of the work that we might do into being an editing process, and necessarily like a generation process. So, even in the world of content, that it helps solve sort of the zero-page problem. A lot of times, it's easier to edit something into what you need, than to essentially start from scratch, especially if you're not feeling particularly creative. Then, I think, the chat becomes your brainstorming buddy. It can actually, I think, unlock or help you unlock a new way of thinking, because I find if you're working in the same space for a long time, your views or the way that you frame things, right things start to become more and more rigid. And if you can leverage something like an LLM, and might actually be able to like steer you in a way that you wouldn't have normally done because your way of thinking or your point of view is become a little bit hardened, which I think is really exciting as well.

In terms of Streamlit, specifically, what are some of the things that the Streamlit’s focused on to help people in the generative – essentially working Gen AI workloads and applications?

[00:34:40] AK: Yes. So, I would say, it's kind of three layers, that we're thinking about kind of strategically. The first is certain things that we want to do specifically in the library that are going to help enable the types of applications that we see coming up. So, we released a number of primitives already this year, in terms of chat, and status and things like that we see evolving with how people want to interact with LLMs. There's a lot more that we're planning to do there, right to make really kind of interactive chat interfaces, the way that you can weave in charts and data and be making decisions within a chat, that that lead and help refine and go to something else.

But also kind of beyond that, some of the stuff that, especially, one of the trends that we see, even outside of Generative AI is just how much people are doing more with images, and media, and videos, kind of in their work. So, making sure all of that's woven in, in very basic ways, like good image pickers, and things like that. So, again, the goal is that for whatever application that you're dreaming up in your head, or that you asked the LLM to dream up for you, we have those basic out of the box primitives that are going to be easy for you to assemble, and put at least a first version of your dream out there.

[00:35:49] SF: One of the first applications I built with Streamlit, was building, actually, like a privacy safe LLM based chatbot. So, I use the new chat components, and it was super easy. Kudos to the team that built all that stuff. 

[00:36:03] AK: I will convey that to the team, but that's always great feedback. And I think feedback is so important in this too. I mean, we're constantly asking some of our power users, but also looking on the forum and saying, “What do you need? What's the newest and latest thing that you're doing?” It's so important for us to be on at the forefront of what the community is doing. But also, the broader Python and Generative AI ecosystem. So, we're partnering deeply with a lot of the libraries that are emerging around right Generative AI. LangChain, LlamaIndex, AssemblyAI, making sure that all of these things just work really well out of the box together, for whatever choice that you're making.

So, we're going to continue to invest in that, as well as with other partners that are kind of adopting Streamlit into the workflows to make various AI products easier. Clarifai, Databutton, Hugging Face, are all amazing companies that are using Streamlit in various ways to kind of empower tools and interfaces for what they're doing.

Then the last bits, I mean, it's really about the community. So, I mean, again, we invest in these primitives. We make these kinds of partnerships integrations easy. But it's really about the beautiful and amazing things that our community members build. I mean, first off, that's inspiration always for things we get better. We always say that the first step is can we make something possible, right? Then, we see, can we take 50 lines of code and make it one, right? And constantly just making it better and more accessible and easier for people to build. But you can go on our website today and just see the amazing amount of LLM powered apps and components that people have made and shared back with the community. I think that's what really kind of increases the pace of innovation, is seeing and being inspired by something else that somebody else has done, and then not having to recreate yourself, but immediately grabbing the code and going and adapting it to your use case.

[00:37:43] SF: Yes, that’s amazing. One of the things I was thinking about, is well, so this week, Dreamforce is going on in San Francisco, and is a huge, of course, focus on AI. But one of the things they talked about at Dreamforce was that the AI revolution is really like a data revolution. Essentially, you don't have Gen AI without the data. But additionally, we wouldn't have all the buzz around AI right now without the application frontend. It's one thing to have a foundation model or something like that. But if it was only an API entry point, there was no UI to it, my dad would not know about it. It was ChatGPT that brought Generative AI, essentially, to my parents, and to a whole other generation of people who aren't deep in the world of data and engineering.

So, a few months ago, I was at a Vercel and OpenAI meet up in San Francisco, and one of the things they talked about there was more like, Vercel trying to position themselves as the frontend for AI applications. But one of the things, as we've been talking, that, I think is interesting is that, in many ways, like Streamlit is the front-end application for certain types of AI applications, essentially the ones where you're empowering these data users, or the people who need the data within the application. It actually creates, I think, if you can essentially build these applications using Gen AI without having to actually write Python code, then you're also creating a whole new builder community that has access to building, essentially, AI powered data applications.

[00:39:18] AK: Yes. I think, Vercel is great. There's a lot of great companies that are doing things and, this is a large and growing space. I think, what’s always kind of been true for Streamlit is, we're just laser focused on data teams, and really kind of specifically in what they need. So, I think there's a lot of people that it would be great to help them make their own kind of applications and things like that. But what gets us out of bed in the morning is really thinking about that, data scientist who's working in SQL and Python today, or that ML ops engineer who's optimizing their model serving, and how we can make their lives better, and how we can bring LLM and Generative AI right to them.

So, we do aim to be the best library for those types of people, when they want to build – when they want to bring LLM and Generative AI into the data applications they're building, we definitely hope that they turn to Streamlit, and we're continuing to invest to make that a great experience for them.

[00:40:17] SF: I guesslooking ahead, what are some of the trends that you see in the data application space, that maybe even beyond Gen AI that Streamlit is trying to sort of stay in front of and adapt to? 

[00:40:30] AK: Yes. I mean, Generative AI is obviously the really big one right now. That just continues on a trend of ML, going deeper and deeper, into different functions. I think that, the bigger trend and a lot of data and a lot of companies is about ML powering marketing, right? ML powering sales, right? And how we move from a world of graph the past. I know exactly what question I want to answer to kind of predict the future, and get curious and kind of dive deeper.

I don't think we really had good methods, a decade ago, in order to say, “Hey, here's all of the data, and here's how you could access it, Sean, and that we feel comfortable, maybe giving you access or empowering them in that way.” So, that I think, has been a big trend just in the underlying data, in terms of making that more accessible. That is obviously been a big trend, and then data applications right on top of that, because we're able to do that.

Obviously, unstructured data, is a huge thing, and it continues to be a huge thing. Media, video, audio. All of these types of things, sensor data, that's coming in. There are new ways that we need to visualize that, new ways that we need to understand that, that's different from how you need to just understand financial data, or time series data, things like that. So, really, figuring out and making sure that we're adapting well for that unstructured work. It's very different, when you're doing something on a time series, there's ways that you're looking at that data, and you're correlating it, and you're graphing it. That's very different than when you're trying to understand a whole bunch of different media files, and you need to potentially draw bounding boxes on them, and other things. There's so much really, really interesting stuff that's especially coming out of biotech right now. Or in the ways that they need to visualize and understand molecules and RNA, and things like that, and that type of data, and how we kind of help them do that work.

Then, this doesn't, you know, maybe sound like, quite as sexy as all of those other things that the media files, and the 3D molecules. But really, just more workflow automation and self-serve, right? An really making it a lot easier for more people across the company to participate in that that data work. I think, everybody needs to and is becoming kind of more data informed, understanding how data works and how data impacts their area. So, extending, right kind of tooling to that, but then that brings new requirements, in terms of – and that, to me, is sometimes the most exciting stuff, is like when I get to go and talk to a sales engineer at a chemical manufacturing company, and understanding kind of what they're doing and how they're evolving. Then, the new types of things that they're doing to ship faster to their partners and have better visibility, for people internally into the ecosystem. So, those are other types of things that we're focusing on as well.

[00:43:16] SF: Absolutely. You mentioned biotech. There's this whole world now, I think around using generative AI for drug discovery as well. A lot of that is, how do you shortcut this process that takes like 5 to 10 years, basically, to discover a new drug. If you can cut that down by even 10%, that is massive in terms of the potential, like good impact that can have on people that are maybe suffering from a particular disease. But also massive in terms of reduction of cost to the companies that are actually doing all this stuff?

[00:43:49] AK: Yes. I am fascinated by what we’re doing there, and I am incredibly excited as someone who really does not understand chemistry, but understands that I'm going to need drugs and people in my family are going to need them and medical advances, over the coming decades. Yes, it really does feel like we're off to the races with a lot of things that have been developed now, and even in the past few years, some of the innovations that came out of COVID-19, and really kind of emerging just there. Then, yes, and then what we can do with Generative AI, what we can do with a lot – I mean, we have so much healthcare data, right? We do. It's just, that I think, is in some ways, for good reasons, one of the bigger remaining frontiers of how we join that and understand that right? But there's a lot of data that's already out there, that even beyond drug discovery, I think is going to fuel new insights and innovation to understanding very specifically tailored to you, Sean. You have that data, probably in your Fitbit and things like that, that could help your doctor. But how do we figure out in a good safe and governance secured way to do that and give access to people of insights? I mean, that's a really interesting problem to solve.

[00:44:53] SF: Yes, I mean, I think we're really only scratching the surface, was a lot of the stuff even as powerful as why are these tools are, we're really in the toddler phase of Gen AI. Wait until – I think Marc Benioff said that this week, like, wait until these things are teenagers. It's going to be a whole another world. The 56k baud modem days of the Internet, versus, essentially fiber optic to the world of Netflix streaming and so forth that we live in today. So, massive, massive change.

My husband and I often play this game of like, what are the things that we think are absolutely true today, that our kids are going to just like mock us for, 30 years from now? They'll be like, “I can't believe you thought that was true, or that was a good idea.” But it is kind of – I mean, I'm an optimist at heart. I can I can find a silver lining anything. But I am incredibly optimistic about the way that the data and Gen AI, and just kind of empowering more people to do more with all of this, is going to help us solve and just understand so much more in science and technology.

[00:45:50] SF: Amazing. I think that's a good place to leave it on a positive note. Amanda, I want to thank you so much for being here. It was really, really enjoyable, and best of luck with everything that you're doing over at Snowflake and Streamlit. I'm looking forward to seeing the continued announcements.

[00:46:04] AK: Yes. Thank you for having me. Please go try Streamlit, you will find out it's not great for everything, but it's great for a certain set of things and you will find out pretty quickly if you like it or not. If you do or if you don't, please give us feedback. Come to our forum. We love to hear it. Please help drive the community forward.

[00:46:19] SF: Awesome. Thank you so much. Cheers.

Right stop recording. It'll take a minute or so for you to upload on your end.

[00:46:30] AK: All good says 99% uploaded.

[00:46:32] SF: Ninety-eight percent, yes. That’s right.