Snowflake Data Cloud Summit 2024 | Amanda Kelly | Snowflake & Chris Van Pelt | Weights & Biases

Clips
News
More from Snowflake Data Cloud Summit 2024

Amanda Kelly

Streamlit Product Director

Snowflake

Chris Van Pelt

Co-Founder & CISO

Weights & Biases

Snowflake Notebooks aims to streamline developer workflows

In this new era of generative artificial intelligence, big data companies are innovating, and new companies are working in novel ways. At the Data Cloud Summit, Snowflake Inc. has sought to outline how it plans to help businesses of all sizes with its AI tools, including with Snowflake Notebooks.Technology has been moving quickly, according to Chris Van Pelt (pictured, right), co-founder and chief information security officer of Weights & Biases Inc., which aims to build the best tools possible for AI developers.Snowflake’s Amanda Kelly and Weights & Biases’ Chris Van Pelt talks to theCUBE about Snowflake Notebooks

play_circle_outline Building Tools for AI Developers: A Visual Analysis of Model Comparison using Weights & Biases

play_circle_outline Benefits of Snowflake Notebooks in terms of data access, resource allocation, and integration

play_circle_outline Overview of new features and innovations in Snowflake Notebooks, including Streamlit integration

play_circle_outline Adrian and Lucas: Pioneering Neural Nets in the Woods Guided Weights & Biases and Streamlit to Empower ML Model Development and Sharing

play_circle_outline Bridging the Gap: Empowering Data Teams with Streamlit's UI Building Superpower in Python

Info

Amanda Kelly | Snowflake & Chris Van Pelt | Weights & Biases

Amanda Kelly

Streamlit Product Director Snowflake

Chris Van Pelt

Co-Founder & CISO Weights & Biases

search

Rebecca Knight

>> Good afternoon, everyone, and welcome back to theCUBE's Live coverage of Snowflake Data Cloud Summit. I'm your host, Rebecca Knight, and I'm sitting alongside my co-host, co-analyst, co-founder of theCUBE, Dave Vellante. Dave, this is a really cool conference and what's really becoming so clear is how much Snowflake is really arming developers with tools to accelerate AI.

>> Well, early days of theCUBE, we started with the Hadoop ecosystem, the big data ecosystem, and it dispersed.

Rebecca Knight

>> You're dating yourself.

>> Well, the cloud made it disperse. Well, I could really date myself, but the cloud dispersed that ecosystem into the cloud, into places like Snowflake. A lot of the legacy, if you will, big data companies are here, but they're innovating of course now, the AI era, and we're seeing a lot of new really exciting companies that do novel things.

Rebecca Knight

>> Indeed, we do. So with that, I'd like to introduce our next guests. They are Chris Van Pelt, co-founder and CISO of Weights & Biases, welcome Chris. And Amanda Kelly's Streamlit product director at Snowflake. Welcome, Amanda.

>> Streamlit co-founder.

Amanda Kelly

>> And Streamlet co-founder. Yeah.

Rebecca Knight

>> All right. Well, congratulations. I want to start with you, Chris, tell our viewers a little bit about weights and biases. Cool name.

Chris Van Pelt

>> Thank you. Yeah, so Weights & Biases' mission is to build the best tools we can for AI developers. We started about six years ago when the world looked a little different than it does today in terms of the AI stack, but have just been thrilled to watch this space grow and to work with really cool companies like Snowflake and products like Streamlit to help developers ultimately build better models.

>> So, who's got the best LLM?

Chris Van Pelt

>> Oh, yeah.

Amanda Kelly

>> It depends on what you're waiting on. A lot of ways to evaluate it.

Rebecca Knight

>> You're smooth. You're smooth, Amanda.

>> Amanda, yeah. But of course I'm half joking, but that's part of what you guys do is to help people understand the right fit for the right job. Is it not?

Chris Van Pelt

>> Yeah, so a big part of our interface is very visual. We make it easy to compare two different models on a variety of different factors. You can also dive in to say individual examples from a data warehouse like Snowflake and see how the model is performing at a case by case basis. So, it's a very visual product that helps people.

>> Are humans still better at picking the right model than machines, and will that change anytime soon?

Chris Van Pelt

>> Well, yeah. I mean, often when it comes to figuring out all the right settings you need to get the best model, machines are much better at that. You just let them run. So, Weights & Biases has a feature called Sweeps, which actually allows you to say, "Hey, just go find the best model." Because often as humans, we're just guessing. I mean, we have some intuition, but when you really to tweak the final settings to get the best model, a feature like Sweeps is where you go.

>> And Amanda, keynote this morning, great job. You brought a lot of energy.

Amanda Kelly

>> To wake them up. A couple hours in, they hadn't had coffee in a while

>> Well, how'd that feel being up there front of all those people?

Amanda Kelly

>> I mean, it feels amazing, mostly because I'm getting to show off some of the great work of my team. And so a lot of long weeks and hard nights going into making the Notebook really to integrating those AI features and things like dark mode, all of the things that we announced with Copilot, the new APIs. It's just amazing to start seeing it all come together.

>> So, you and I talked yesterday in the analyst one and one, so I kind of know the answer, but the answer was great. Help our audience understand-

Amanda Kelly

>> I got to repeat it.

>> Absolutely. Pressure's on why does the world need another notebook?

Amanda Kelly

>> So, really it's all about having the right tool for the job. And so there's a lot of great notebooks out there, and we have a lot of great partners like Hex that are being really next generation experiences. What we're trying to do with Snowflake is really make sure that when you have a question about your data, when you have something you need to get done today, you are really no more than one click right away from having that answer. And so that's why we've built a really integrated notebook. So, we're bringing in all the great features that notebooks have, the get integration, the scheduling, the Python and SQL cells mixing, but we're really optimizing for the types of tasks that people want to do on Snowflake. And then we're bringing the best of Snowflake, the integrations with Unistore, the integrations with Cortex, all of that together in one notebook. So, it's really easy to accomplish anything that you want to do with your Snowflake data.

>> So, ask the user here, is that what your experience has been with the new Snowflake Notebooks, and can you share those?

Chris Van Pelt

>> I mean, the Notebook interface is such a common paradigm for data scientists or people doing experimentation, working with models. I felt right at home in the Snowflake Notebooks. In many ways, it's a more polished experience, like just the look and feel of the interface feels shinier or newer model .

Rebecca Knight

>> So, how are you using it, Chris? How is Weights & Biases using the Notebook?

Chris Van Pelt

>> Yeah, so our product actually integrates with Notebooks so that when ML engineers are doing experimentation or trying things out, they can keep track of all of their experiments in Weights & Biases. So, I was actually using it as one of my users would use to make sure all of our functionality is deeply integrated and operating appropriately.

>> When you're testing a new product like a Notebook, what are you looking for? I mean, obviously the overall experience and how integrated it is, but what else from your standpoint are you looking for?

Chris Van Pelt

>> Yeah. Well, I mean, one, anytime you're doing ML modeling, it starts with the data. You need to be able to quickly get data from whatever sources there might be, and Snowflake Notebooks actually makes that really easy. It's directly integrated into the warehouse, and I can specify a worksheet and then just get a data frame out, which is the object that I'm used to working with as an ML engineer. I'd say the other thing that starts to get more and more complicated when you're doing AI or machine learning is being able to get the right accelerator for whatever task you might be doing. So, if I'm going to do deep learning or something that's really computationally expensive, I'm going to need a GPU or maybe mini GPUs and the Snowflake Notebooks make that easy for me to specify, "Hey, these are the resources I need." And magically behind the scenes everything gets scheduled and taken down for me.

>> So, does it pretty much replace your other notebook experiences, or are there use cases where you might say, "Okay, I'm going to use Snowflake's Notebook here, or other notebooks in other situations?" You mentioned Hex as a partner. When would you use each?

Chris Van Pelt

>> Well, I'd say the Snowflake Notebook is going to be for the things I need to get done with work data, maybe potentially more sensitive things that are really internal to an organization. I might still use a tool like CoLab if I'm trying to share a bit of research with someone out on the open internet or in certain extreme cases, just run a little notebook locally on hardware that I might have that's not in the cloud.

Rebecca Knight

>> Amanda, beyond the beautiful polished interface, can you walk us through some of what is new here and how it's really going to help streamline development?

Amanda Kelly

>> Yeah, so as you mentioned, a lot of what we want to do is we work very hard at Snowflake to make things are integrated down to the object layer. So, that can make it really, really easy for you to just open up and say, "What's going on with this pipeline? What's going on with this table?" And that's going to work just out of the box. Again, everything that you were seeing right here at Snowflake Summit are things that we are working to make sure only one line of code or a couple clicks away. Cortex, Copilot, we're going to be introducing Inline Copilot soon. Everything that we're talking about with Unistore, Polaris. All of that management in the Snowflake Notebook. But then we have a lot of other features and other features that are coming soon that just make it really easy kind of in the user interface to work with that data. So, the Streamlit integration is a really big part of this in terms of allowing a next level of interactivity so that you can understand and do things with the data. But things like scheduling, get integration, telemetry and debugging, these are all things that we're investing in to make a really crisp developer experience that's going to be as good as anything that you would get with any other notebook.

>> As peeps who hang out with... Your peeps are developers. So how has a developer's lives changed since November of 2022? You hear all productivity boom and it's questionable how much the broader society, but the one area that people say that can have a measurable impact is developers. From your perspective, how true is that and what kind of impact have you seen in the last two years?

Chris Van Pelt

>> I'd say for me personally, it's dramatically changed my workflow. Historically, if I'd be working on a programming problem and scratch my head and say, "Hey, how do I do that?" The first place I go is Google. Now the first place I go is ChatGPT or Claude or whatever kind of large language model interface I might choose. There's also the Copilot integrations where I essentially get nice auto complete that is just as a developer, it's magical, when it works. And I'd say it works more often than it doesn't work, but just to see I'm typing, I'm thinking and suddenly, boom, what I would have typed in the next two minutes is there and I could just hit tab and it's done. I'd say another big change for me has been taking the parts of programming that maybe I don't like that much and outsourcing that to a large language model. So, if I've implemented some logic and now I want to write some code, some tests to make sure that that logic will continue to work, I can just tell an LLM, "Write me the tests," and for the most part, it'll get that done. I mean, I've got to do some cleanup and move things around, but it's a big productivity boost.

>> So testing, debugging, are you there yet or?

Chris Van Pelt

>> Well, yeah. I mean, I'd say a really common thing is I'll be running code, I'll get some weird error. My first step is copy the error, put it into an LLM and say, "Hey, do you got any ideas?" And it's not always going to be a silver bullet and get it right, but often it gets me going down the right path.

Rebecca Knight

>> It'll get your human brain thinking in new ways about different possible avenues.

Amanda Kelly

>> You run into fewer dead ends because now you have an assistant who can be coding along with you.

>> But as a developer you can scan the output the way you and I can scan the output of something we ask ChatGPT to write and say, "That's not right. I need to go in and fix it." Or, "Wow, that's really good. I'm going to change two words or line of code-

Rebecca Knight

>> It's really that good.

Chris Van Pelt

>> It's a bad idea if you don't understand what the code is doing and you're just copy and pasting code, you'll have weird issues-

>> I've tried that, actually. I'm not a developer.

Amanda Kelly

>> When you're developing things, there's a number of things that you're often trading off. Would I rather have this run faster, be cheaper, more accessible to different people working this way or that? So, there's a lot of business context and user context that you're weighing off and you're making a lot of decisions about when you're building things. And that's things that can be a little bit hard right now specifically to use LLMs for, but you can continue to use it to give you great ideas to work with you alongside and especially to build new experiences. And I think that's a big thing that we've seen, especially with Streamlit, is now being able to weave chat into so many more products just really opens up new ways that we can interact with data, but with UIs in general. And so it's a really interesting kind of new paradigm that we are arming developers with to make much better experiences because it is a much more natural way to just say, "I don't want to click through four things to find out what happened with cost. I just want to ask the question, what was my bill last year?"

>> Go, ahead please.

Rebecca Knight

>> Segueing into that, just talking about the idea of experience. How do the innovations that you've announced at this summit really fit into the broader context of Snowflake's vision and strategy?

Amanda Kelly

>> Well, I think Snowflake's vision at the end of the day, and one of the things I really love about Snowflake is that it's so simple. We really just want to put your data to work for you. We believe if you aggregate the data and we can make it very easy to govern, to access, to do all of that hard work for you, you're ultimately going to be more productive as a company because you can use that data in new and interesting ways. And I think this new kind of experience layer that we're getting with gen AI and also the new types of things we're offering with Notebooks and Streamlits and other things like that, just really take that up a notch in terms of what we can do. And I often talk about, you guys probably have a number of questions about data. I wonder how many people are watching this right now? How do they segment by this, by that? How does that compare to another broadcast that we did earlier? Very few of those I bet you ever get answered, because you've learned that it's hard and it's costly to go to someone and ask them, and they're going to say, "Ah, give me a week or whatever." But you got to make a decision now if you want to pivot before your next thing. Those are some of the types of things where all of a sudden we're unlocking a new wave of questions, especially coming from the end users who can talk to data, facilitated by great data engineering teams who can build these experiences, but also allowing them to build them a lot faster too, which makes for a much better iterative loop, I think, between the people who need to use the data and make decisions and the people who are working with the data.

>> Well, and it would be nice... You're right, we get those questions all the time, and it's a root canal to get them. We have developed our own little dashboard, but it would be nice if LinkedIn had an easier to use API. So, I want to ask you about your title. You have CISO in your title. How does that fit in to the whole Weights & Biases story? The whole governance and security piece of it.

Chris Van Pelt

>> Weights & Biases helps keep track of all of these really important and often very sensitive artifacts, such as a dump out of a Snowflake table that is going to train a model. And our enterprise customers often have a lot of very strict requirements around how that software is developed, how it's deployed, understanding that architecture. So, as someone who's been developing the software since the beginning and also goes out and talks a lot to customers, my role as CISO is to A, ensure that we're running a compliant and modern security practice. And then B, actually go and communicate with other enterprises and talk about how we do data security and deployment, which is I think something that Snowflake has actually done incredibly well. I look to them as an example of a cloud-based service that has been able to educate the market on how you can keep data segregated and secure in this new cloud-based world.

>> In your world, what's the hardest sort of security problem? Is it educating people in things like what not to do? Is it dealing with non-integrated platforms?

Chris Van Pelt

>> I'd say one of the biggest challenges for us, I think, especially in the machine learning space, is that there are still, there's probably even a growing number of companies that are actually operating their own infrastructure. So, instead of going to AWS or Google Cloud or Azure, they actually bought a data center and filled it with Nvidia GPUs, and now they want to have a tool like Weights & Biases running in that environment. That is something we support, but it's very challenging and tricky to get the software to work in one of those environments because everyone is slightly different. We much prefer the cloud. So, a big part of our education is saying, "All right, well, here's why the cloud is better. Here's how we've done it securely, and we can give you an isolated instance,: these types of things.

>> At RSA last year, '23, the conversation was about how bad guys are going to use LLMs to do better phishing emails. The conversation this year was all about the need to protect AI, to secure AI. Do you see that? Why was it such a gap that people didn't realize it was coming? Did you see it coming?

Chris Van Pelt

>> Well, yeah. I think early on as companies like OpenAI were releasing the models, they talked a lot about the red teaming and the work they've done to improve the safety of the models. That being said, it's a brand new technology and there are still unsolved problems. You take prompt injection or essentially getting the model to ignore what it was told to do and go off the rails is still a problem that is being researched and a number of things can be put in place to mitigate, but it's an unsolved problem, and I think it's true for any new technology. Everyone gets really excited." Oh, neat." And then you step back and think like, "Okay, how do we control against bad actors doing things with these systems we don't want them to?"

>> How do I manipulate it if they're the bad actor? Why did you start Streamlit and what's it been like to see Streamlit as part of Snowflake and the uptake? We see the data, we have some survey data on Streamlit, looks like continues to grow and penetrate Snowflake accounts. But take us back to the beginning and what the experience has been like here.

Amanda Kelly

>> I'll say, it's actually quite fun to be here with Weights & Biases because our origin story actually starts at the same point in time. About seven years ago, our co-founders, Adrian and Lucas, went off into the woods and were experimenting with neural nets and they ran into a number of problems and one of those insights to the founding of Weights & Biases and the other insight led to the founding of Streamlit.

>> Come on. .

Amanda Kelly

>> No. It's true. I've got the slides with the pictures to prove it and everything. But the insight that led to the generation of Streamlit was it was really, really hard, especially when you were developing these machine learning models to really understand, evaluate what was going on, share that, put it into production with someone else because you needed an interface on top of that. That's just generally the best way. Whether you're trying to tweak a GAN model or you're trying to chat with an LLM, you want an interface to do that. Unfortunately, that's not a skillset that tends to be shared by a lot of people working in data teams or machine learning engineers. So, what was happening was they'd have to go contact a tools team, hire a front-end engineer on their team, out sort this to somebody that they found online. Really, really slow development times. And as he was talking about, it's when you're at these early stages of technology, you want to be iterating very, very quickly. You need to be getting in there, understanding what's working, understanding what's valuable, building new things. And so the insight was really, we need to bring that superpower of building great UIs directly into the hands of data teams. And so we did that by making it really easy to build UI in Python, which is the native language of machine learning and a lot of data. Fast forward, we've been on this journey of saying, "How can we make it really, really fast and easy for you to build these applications, share them with others, do these kind of iterative cycles?" And we were out actually raising our series C when Snowflake approached us and they said, "We'd like to acquire you." And we said, "No, thank you. That's nice." But actually, Benoit one of the co-founders and he said, "I don't think you understand the vision here." Honestly, one of the things that we were struggling with at the time as a startup is we have this great UI layer. It could be fast to build apps, but everyone was like, "I need this in a secure environment. I don't want my data going out to your service. We got to lock it down." And Snowflake at the time, especially with launching its native app infrastructure, was looking for a really easy way that we could bring that capability natively. And so it really was a really ideal kind of match at the time. People have asked me, "How do you recreate that magic and acquisition?" I'm like, "There was a Streamlit sized hole in Snowflake's strategy and it worked out really well." And Snowflake has been a really fantastic steward of the open source project. So, we have a whole team focused just on open source. They're growing it. We're obviously doing a lot more with the Polaris catalog recently with the Snowflake Arctic launch. And the goal is just go, make great open source tools for developers. Then we will bring you a really great thing if you're a Snowflake customer, a great way to run that. Make sure your data never has to leave the premises, but you can build all of those experiences that you want.

>> Great story.

Rebecca Knight

>> Yes. Great way to, great note to end on. Thank you.

>> I got one more. Okay. Oh no, we're out of time. Good job.

Rebecca Knight

>> He could talk forever.

>> How GPU intensive is your business? Do you need to access a lot of GPU time?

Chris Van Pelt

>> We don't actually manage GPUs ourselves. We partner with folks like Snowflake to actually provision the GPUs, but our customers use thousands, hundreds of thousands of GPUs, and we keep track of it all. So, we can look in our data warehouse and see how much money's going to NVIDIA right now.

>> But to create your product, you don't need a lot of GPU time. Is that correct?

Chris Van Pelt

>> Well, I mean, me personally, as I develop, I'm hitting the OpenAI APIs or the Copilot APIs, and they're burning GPUs, but now it's just -

>> My question is what are you doing with all that dough?

Chris Van Pelt

>> Yeah, good question.

Rebecca Knight

>> And we will end there. Chris and Amanda both, thank you both so much for coming on theCUBE. I'm Rebecca Knight for Dave Vellante, stay tuned for more of theCUBE's live coverage of the Snowflake Data Cloud Summit. You're watching theCUBE, the leader in enterprise tech news and analysis.

Amanda Kelly | Snowflake & Chris Van Pelt | Weights & Biases

search

Rebecca Knight

>> Good afternoon, everyone, and welcome back to theCUBE's Live coverage of Snowflake Data Cloud Summit. I'm your host, Rebecca Knight, and I'm sitting alongside my co-host, co-analyst, co-founder of theCUBE, Dave Vellante. Dave, this is a really cool conference and what's really becoming so clear is how much Snowflake is really arming developers with tools to accelerate AI.

>> Well, early days of theCUBE, we started with the Hadoop ecosystem, the big data ecosystem, and it dispersed.

Rebecca Knight

>> You're dating yourself.

>> Well, the cloud made it disperse. Well, I could really date myself, but the cloud dispersed that ecosystem into the cloud, into places like Snowflake. A lot of the legacy, if you will, big data companies are here, but they're innovating of course now, the AI era, and we're seeing a lot of new really exciting companies that do novel things.

Rebecca Knight

>> Indeed, we do. So with that, I'd like to introduce our next guests. They are Chris Van Pelt, co-founder and CISO of Weights & Biases, welcome Chris. And Amanda Kelly's Streamlit product director at Snowflake. Welcome, Amanda.

>> Streamlit co-founder.

Amanda Kelly

>> And Streamlet co-founder. Yeah.

Rebecca Knight

>> All right. Well, congratulations. I want to start with you, Chris, tell our viewers a little bit about weights and biases. Cool name.

Chris Van Pelt

>> Thank you. Yeah, so Weights & Biases' mission is to build the best tools we can for AI developers. We started about six years ago when the world looked a little different than it does today in terms of the AI stack, but have just been thrilled to watch this space grow and to work with really cool companies like Snowflake and products like Streamlit to help developers ultimately build better models.

>> So, who's got the best LLM?

Chris Van Pelt

>> Oh, yeah.

Amanda Kelly

>> It depends on what you're waiting on. A lot of ways to evaluate it.

Rebecca Knight

>> You're smooth. You're smooth, Amanda.

>> Amanda, yeah. But of course I'm half joking, but that's part of what you guys do is to help people understand the right fit for the right job. Is it not?

Chris Van Pelt

>> Yeah, so a big part of our interface is very visual. We make it easy to compare two different models on a variety of different factors. You can also dive in to say individual examples from a data warehouse like Snowflake and see how the model is performing at a case by case basis. So, it's a very visual product that helps people.

>> Are humans still better at picking the right model than machines, and will that change anytime soon?

Chris Van Pelt

>> Well, yeah. I mean, often when it comes to figuring out all the right settings you need to get the best model, machines are much better at that. You just let them run. So, Weights & Biases has a feature called Sweeps, which actually allows you to say, "Hey, just go find the best model." Because often as humans, we're just guessing. I mean, we have some intuition, but when you really to tweak the final settings to get the best model, a feature like Sweeps is where you go.

>> And Amanda, keynote this morning, great job. You brought a lot of energy.

Amanda Kelly

>> To wake them up. A couple hours in, they hadn't had coffee in a while

>> Well, how'd that feel being up there front of all those people?

Amanda Kelly

>> I mean, it feels amazing, mostly because I'm getting to show off some of the great work of my team. And so a lot of long weeks and hard nights going into making the Notebook really to integrating those AI features and things like dark mode, all of the things that we announced with Copilot, the new APIs. It's just amazing to start seeing it all come together.

>> So, you and I talked yesterday in the analyst one and one, so I kind of know the answer, but the answer was great. Help our audience understand-

Amanda Kelly

>> I got to repeat it.

>> Absolutely. Pressure's on why does the world need another notebook?

Amanda Kelly

>> So, really it's all about having the right tool for the job. And so there's a lot of great notebooks out there, and we have a lot of great partners like Hex that are being really next generation experiences. What we're trying to do with Snowflake is really make sure that when you have a question about your data, when you have something you need to get done today, you are really no more than one click right away from having that answer. And so that's why we've built a really integrated notebook. So, we're bringing in all the great features that notebooks have, the get integration, the scheduling, the Python and SQL cells mixing, but we're really optimizing for the types of tasks that people want to do on Snowflake. And then we're bringing the best of Snowflake, the integrations with Unistore, the integrations with Cortex, all of that together in one notebook. So, it's really easy to accomplish anything that you want to do with your Snowflake data.

>> So, ask the user here, is that what your experience has been with the new Snowflake Notebooks, and can you share those?

Chris Van Pelt

>> I mean, the Notebook interface is such a common paradigm for data scientists or people doing experimentation, working with models. I felt right at home in the Snowflake Notebooks. In many ways, it's a more polished experience, like just the look and feel of the interface feels shinier or newer model .

Rebecca Knight

>> So, how are you using it, Chris? How is Weights & Biases using the Notebook?

Chris Van Pelt

>> Yeah, so our product actually integrates with Notebooks so that when ML engineers are doing experimentation or trying things out, they can keep track of all of their experiments in Weights & Biases. So, I was actually using it as one of my users would use to make sure all of our functionality is deeply integrated and operating appropriately.

>> When you're testing a new product like a Notebook, what are you looking for? I mean, obviously the overall experience and how integrated it is, but what else from your standpoint are you looking for?

Chris Van Pelt

>> Yeah. Well, I mean, one, anytime you're doing ML modeling, it starts with the data. You need to be able to quickly get data from whatever sources there might be, and Snowflake Notebooks actually makes that really easy. It's directly integrated into the warehouse, and I can specify a worksheet and then just get a data frame out, which is the object that I'm used to working with as an ML engineer. I'd say the other thing that starts to get more and more complicated when you're doing AI or machine learning is being able to get the right accelerator for whatever task you might be doing. So, if I'm going to do deep learning or something that's really computationally expensive, I'm going to need a GPU or maybe mini GPUs and the Snowflake Notebooks make that easy for me to specify, "Hey, these are the resources I need." And magically behind the scenes everything gets scheduled and taken down for me.

>> So, does it pretty much replace your other notebook experiences, or are there use cases where you might say, "Okay, I'm going to use Snowflake's Notebook here, or other notebooks in other situations?" You mentioned Hex as a partner. When would you use each?

Chris Van Pelt

>> Well, I'd say the Snowflake Notebook is going to be for the things I need to get done with work data, maybe potentially more sensitive things that are really internal to an organization. I might still use a tool like CoLab if I'm trying to share a bit of research with someone out on the open internet or in certain extreme cases, just run a little notebook locally on hardware that I might have that's not in the cloud.

Rebecca Knight

>> Amanda, beyond the beautiful polished interface, can you walk us through some of what is new here and how it's really going to help streamline development?

Amanda Kelly

>> Yeah, so as you mentioned, a lot of what we want to do is we work very hard at Snowflake to make things are integrated down to the object layer. So, that can make it really, really easy for you to just open up and say, "What's going on with this pipeline? What's going on with this table?" And that's going to work just out of the box. Again, everything that you were seeing right here at Snowflake Summit are things that we are working to make sure only one line of code or a couple clicks away. Cortex, Copilot, we're going to be introducing Inline Copilot soon. Everything that we're talking about with Unistore, Polaris. All of that management in the Snowflake Notebook. But then we have a lot of other features and other features that are coming soon that just make it really easy kind of in the user interface to work with that data. So, the Streamlit integration is a really big part of this in terms of allowing a next level of interactivity so that you can understand and do things with the data. But things like scheduling, get integration, telemetry and debugging, these are all things that we're investing in to make a really crisp developer experience that's going to be as good as anything that you would get with any other notebook.

>> As peeps who hang out with... Your peeps are developers. So how has a developer's lives changed since November of 2022? You hear all productivity boom and it's questionable how much the broader society, but the one area that people say that can have a measurable impact is developers. From your perspective, how true is that and what kind of impact have you seen in the last two years?

Chris Van Pelt

>> I'd say for me personally, it's dramatically changed my workflow. Historically, if I'd be working on a programming problem and scratch my head and say, "Hey, how do I do that?" The first place I go is Google. Now the first place I go is ChatGPT or Claude or whatever kind of large language model interface I might choose. There's also the Copilot integrations where I essentially get nice auto complete that is just as a developer, it's magical, when it works. And I'd say it works more often than it doesn't work, but just to see I'm typing, I'm thinking and suddenly, boom, what I would have typed in the next two minutes is there and I could just hit tab and it's done. I'd say another big change for me has been taking the parts of programming that maybe I don't like that much and outsourcing that to a large language model. So, if I've implemented some logic and now I want to write some code, some tests to make sure that that logic will continue to work, I can just tell an LLM, "Write me the tests," and for the most part, it'll get that done. I mean, I've got to do some cleanup and move things around, but it's a big productivity boost.

>> So testing, debugging, are you there yet or?

Chris Van Pelt

>> Well, yeah. I mean, I'd say a really common thing is I'll be running code, I'll get some weird error. My first step is copy the error, put it into an LLM and say, "Hey, do you got any ideas?" And it's not always going to be a silver bullet and get it right, but often it gets me going down the right path.

Rebecca Knight

>> It'll get your human brain thinking in new ways about different possible avenues.

Amanda Kelly

>> You run into fewer dead ends because now you have an assistant who can be coding along with you.

>> But as a developer you can scan the output the way you and I can scan the output of something we ask ChatGPT to write and say, "That's not right. I need to go in and fix it." Or, "Wow, that's really good. I'm going to change two words or line of code-

Rebecca Knight

>> It's really that good.

Chris Van Pelt

>> It's a bad idea if you don't understand what the code is doing and you're just copy and pasting code, you'll have weird issues-

>> I've tried that, actually. I'm not a developer.

Amanda Kelly

>> When you're developing things, there's a number of things that you're often trading off. Would I rather have this run faster, be cheaper, more accessible to different people working this way or that? So, there's a lot of business context and user context that you're weighing off and you're making a lot of decisions about when you're building things. And that's things that can be a little bit hard right now specifically to use LLMs for, but you can continue to use it to give you great ideas to work with you alongside and especially to build new experiences. And I think that's a big thing that we've seen, especially with Streamlit, is now being able to weave chat into so many more products just really opens up new ways that we can interact with data, but with UIs in general. And so it's a really interesting kind of new paradigm that we are arming developers with to make much better experiences because it is a much more natural way to just say, "I don't want to click through four things to find out what happened with cost. I just want to ask the question, what was my bill last year?"

>> Go, ahead please.

Rebecca Knight

>> Segueing into that, just talking about the idea of experience. How do the innovations that you've announced at this summit really fit into the broader context of Snowflake's vision and strategy?

Amanda Kelly

>> Well, I think Snowflake's vision at the end of the day, and one of the things I really love about Snowflake is that it's so simple. We really just want to put your data to work for you. We believe if you aggregate the data and we can make it very easy to govern, to access, to do all of that hard work for you, you're ultimately going to be more productive as a company because you can use that data in new and interesting ways. And I think this new kind of experience layer that we're getting with gen AI and also the new types of things we're offering with Notebooks and Streamlits and other things like that, just really take that up a notch in terms of what we can do. And I often talk about, you guys probably have a number of questions about data. I wonder how many people are watching this right now? How do they segment by this, by that? How does that compare to another broadcast that we did earlier? Very few of those I bet you ever get answered, because you've learned that it's hard and it's costly to go to someone and ask them, and they're going to say, "Ah, give me a week or whatever." But you got to make a decision now if you want to pivot before your next thing. Those are some of the types of things where all of a sudden we're unlocking a new wave of questions, especially coming from the end users who can talk to data, facilitated by great data engineering teams who can build these experiences, but also allowing them to build them a lot faster too, which makes for a much better iterative loop, I think, between the people who need to use the data and make decisions and the people who are working with the data.

>> Well, and it would be nice... You're right, we get those questions all the time, and it's a root canal to get them. We have developed our own little dashboard, but it would be nice if LinkedIn had an easier to use API. So, I want to ask you about your title. You have CISO in your title. How does that fit in to the whole Weights & Biases story? The whole governance and security piece of it.

Chris Van Pelt

>> Weights & Biases helps keep track of all of these really important and often very sensitive artifacts, such as a dump out of a Snowflake table that is going to train a model. And our enterprise customers often have a lot of very strict requirements around how that software is developed, how it's deployed, understanding that architecture. So, as someone who's been developing the software since the beginning and also goes out and talks a lot to customers, my role as CISO is to A, ensure that we're running a compliant and modern security practice. And then B, actually go and communicate with other enterprises and talk about how we do data security and deployment, which is I think something that Snowflake has actually done incredibly well. I look to them as an example of a cloud-based service that has been able to educate the market on how you can keep data segregated and secure in this new cloud-based world.

>> In your world, what's the hardest sort of security problem? Is it educating people in things like what not to do? Is it dealing with non-integrated platforms?

Chris Van Pelt

>> I'd say one of the biggest challenges for us, I think, especially in the machine learning space, is that there are still, there's probably even a growing number of companies that are actually operating their own infrastructure. So, instead of going to AWS or Google Cloud or Azure, they actually bought a data center and filled it with Nvidia GPUs, and now they want to have a tool like Weights & Biases running in that environment. That is something we support, but it's very challenging and tricky to get the software to work in one of those environments because everyone is slightly different. We much prefer the cloud. So, a big part of our education is saying, "All right, well, here's why the cloud is better. Here's how we've done it securely, and we can give you an isolated instance,: these types of things.

>> At RSA last year, '23, the conversation was about how bad guys are going to use LLMs to do better phishing emails. The conversation this year was all about the need to protect AI, to secure AI. Do you see that? Why was it such a gap that people didn't realize it was coming? Did you see it coming?

Chris Van Pelt

>> Well, yeah. I think early on as companies like OpenAI were releasing the models, they talked a lot about the red teaming and the work they've done to improve the safety of the models. That being said, it's a brand new technology and there are still unsolved problems. You take prompt injection or essentially getting the model to ignore what it was told to do and go off the rails is still a problem that is being researched and a number of things can be put in place to mitigate, but it's an unsolved problem, and I think it's true for any new technology. Everyone gets really excited." Oh, neat." And then you step back and think like, "Okay, how do we control against bad actors doing things with these systems we don't want them to?"

>> How do I manipulate it if they're the bad actor? Why did you start Streamlit and what's it been like to see Streamlit as part of Snowflake and the uptake? We see the data, we have some survey data on Streamlit, looks like continues to grow and penetrate Snowflake accounts. But take us back to the beginning and what the experience has been like here.

Amanda Kelly

>> I'll say, it's actually quite fun to be here with Weights & Biases because our origin story actually starts at the same point in time. About seven years ago, our co-founders, Adrian and Lucas, went off into the woods and were experimenting with neural nets and they ran into a number of problems and one of those insights to the founding of Weights & Biases and the other insight led to the founding of Streamlit.

>> Come on. .

Amanda Kelly

>> No. It's true. I've got the slides with the pictures to prove it and everything. But the insight that led to the generation of Streamlit was it was really, really hard, especially when you were developing these machine learning models to really understand, evaluate what was going on, share that, put it into production with someone else because you needed an interface on top of that. That's just generally the best way. Whether you're trying to tweak a GAN model or you're trying to chat with an LLM, you want an interface to do that. Unfortunately, that's not a skillset that tends to be shared by a lot of people working in data teams or machine learning engineers. So, what was happening was they'd have to go contact a tools team, hire a front-end engineer on their team, out sort this to somebody that they found online. Really, really slow development times. And as he was talking about, it's when you're at these early stages of technology, you want to be iterating very, very quickly. You need to be getting in there, understanding what's working, understanding what's valuable, building new things. And so the insight was really, we need to bring that superpower of building great UIs directly into the hands of data teams. And so we did that by making it really easy to build UI in Python, which is the native language of machine learning and a lot of data. Fast forward, we've been on this journey of saying, "How can we make it really, really fast and easy for you to build these applications, share them with others, do these kind of iterative cycles?" And we were out actually raising our series C when Snowflake approached us and they said, "We'd like to acquire you." And we said, "No, thank you. That's nice." But actually, Benoit one of the co-founders and he said, "I don't think you understand the vision here." Honestly, one of the things that we were struggling with at the time as a startup is we have this great UI layer. It could be fast to build apps, but everyone was like, "I need this in a secure environment. I don't want my data going out to your service. We got to lock it down." And Snowflake at the time, especially with launching its native app infrastructure, was looking for a really easy way that we could bring that capability natively. And so it really was a really ideal kind of match at the time. People have asked me, "How do you recreate that magic and acquisition?" I'm like, "There was a Streamlit sized hole in Snowflake's strategy and it worked out really well." And Snowflake has been a really fantastic steward of the open source project. So, we have a whole team focused just on open source. They're growing it. We're obviously doing a lot more with the Polaris catalog recently with the Snowflake Arctic launch. And the goal is just go, make great open source tools for developers. Then we will bring you a really great thing if you're a Snowflake customer, a great way to run that. Make sure your data never has to leave the premises, but you can build all of those experiences that you want.

>> Great story.

Rebecca Knight

>> Yes. Great way to, great note to end on. Thank you.

>> I got one more. Okay. Oh no, we're out of time. Good job.

Rebecca Knight

>> He could talk forever.

>> How GPU intensive is your business? Do you need to access a lot of GPU time?

Chris Van Pelt

>> We don't actually manage GPUs ourselves. We partner with folks like Snowflake to actually provision the GPUs, but our customers use thousands, hundreds of thousands of GPUs, and we keep track of it all. So, we can look in our data warehouse and see how much money's going to NVIDIA right now.

>> But to create your product, you don't need a lot of GPU time. Is that correct?

Chris Van Pelt

>> Well, I mean, me personally, as I develop, I'm hitting the OpenAI APIs or the Copilot APIs, and they're burning GPUs, but now it's just -

>> My question is what are you doing with all that dough?

Chris Van Pelt

>> Yeah, good question.

Rebecca Knight

>> And we will end there. Chris and Amanda both, thank you both so much for coming on theCUBE. I'm Rebecca Knight for Dave Vellante, stay tuned for more of theCUBE's live coverage of the Snowflake Data Cloud Summit. You're watching theCUBE, the leader in enterprise tech news and analysis.