AI Pulse: August 2026

Rachel Roundy: Hello, and welcome to the Snowflake AI Pulse. This is our ongoing series where we bring together the most important AI innovations across our platform, and most importantly, what they actually unlock for you, our customers. Whether you've joined us from previous Pulses or are here for the first time, welcome. We are really glad to have you. Each episode here is designed to cut through the noise. We are really trying to connect our newest capabilities to your real workflows, whether you're building with data or using it to drive decisions or doing even cooler things with AI. Of course, I have to do a quick shout out to our legal team. Here are our safe harbor disclaimers at the top of all of our content. I'm your host, Rachel Roundy. I'm the product marketing lead for AI here at Snowflake. And we have an amazing agenda here today and incredible experts that I will introduce shortly. Just wanted to give a quick shout out on the agenda as a reminder to our audience, especially those of you who are brand new here today. We will be doing live Q and A throughout the presentation. So please use tool in the webinar platform to input your questions. Some of our team on the product and PMM side might be answering live there. Otherwise, we'll have dedicated Q and A sessions after each of the deep dives today. So as I mentioned, I am joined by an incredible crew here today. We have Paras Mehra, Umesh Bene Krishnan, Avanash Joshi, Efi Gonawan, and Hana Maikula. And without further ado, let's jump right in. So last month in July, we talked about how enterprise AI has moved beyond the prototype. A lot of you are doing really cool stuff. Right? And organizations have stopped asking whether AI can work. We've really been asking, how do we scale it? And that is still really true for everyone. But in really recent conversations with our customers, a sharper problem has really come into focus. And curious to hear how this resonates with you all. But really, what we're hearing is that models alone are good enough. The models are great. That's not the bottleneck. What's actually happening right now and what's stopping AI from leaving the development environment is everything around the model. So some of the questions we're hearing from our customers sound like, who controls what AI can actually touch? How do we keep costs predictable when one user can run up thousands of dollars in a couple of weeks or, honestly, like a couple of prompts? Right? How do we promote a model to production on evidence and not just faith? And how do we make sure that a great workflow doesn't just live on one person's laptop, right? How do we share that goodness? And most importantly, or critically, I will say not most importantly, all these things are important. How do we get AI to work across all of our data and not just text? How how is it able to understand that? So when agents are built in separate tools, they often live in silos. And that makes decisions really complicated because they're making decisions without shared context. I know this is something that we hear from you every single day. And that is the biggest reason why eighty eight percent of AI pilots never actually reach production. It's not because there was a failed demo or the prototype wasn't great, but the surrounding machinery to support that scaling wasn't there. And that's really the connective tissue across today's launches. We are putting more of that production machinery into place from governance and cost controls to ML ops, model choice, and the ability to actually just get work done, right, and have those deliverables done by your AI tools. And we are putting that machinery in place with a unified platform, one where data, models, apps, and governance all come together. I'm sure you've heard us talk about this before, but we really believe that the agentic enterprise is built on four things. The first is enterprise data and context. AI is really only as good as the data and business context behind it. Models, the flexibility to choose the right model for each workload most critically. Software and applications, obviously connecting to the systems where your work actually happens. And tying it all together is the agentic control plane, which is a single place to build, deploy, operate, and govern AI at enterprise scale. And what's really worth emphasizing today in particular is the bottom of this diagram, the single governance boundary. We have cost controls, RBAC, auditability. This is exactly the layer that lets you answer your CSOs questions before they ask them. We've heard from you over and over that you don't want a dozen point solutions that are bolted together, a coding agent over here, a model gateway there, a BI tool that's priced per seat. And so Snowflake really is addressing that by governing cost, access, behavior, execution from one control plane right next to the data and the agents. Alright. So today, every feature that we're gonna talk about was really purpose built to help you achieve these goals. First, we're going to talk about open weight models. Then we will talk about Snowflake ML. Third, AI functions. Next, CoWork, and then CoCo. So how these things are really helping you build that kind of surrounding machinery for your AI in production. This next slide is not one that I'll spend a lot of time on, but this is a great view of what everything is what the status of everything is. So with AI Pulse, we really want to show you what can be sandboxed today or what can be used in production today. So love to see what is available for you or coming soon. And then beyond that, every single one of these AI pulses has more innovation than we can ever cover in a single session. Our team is cranking on amazing features, and we don't have the ability to share all of them in this live format. So I wanted to take a moment to highlight a few additional launches that will not be in the deep dives today, but are definitely worth putting on your radar. The CoCo Slack app brings notes like CoCo into Slack so developers can interact with their data and AI workflows without leaving the tools they're already working in, which is fantastic news for anyone who's a Slack enthusiast. Cortex Sense is something we will touch on in multiple pulses. And Effie will speak to a little bit today. This is a proactive intelligence layer that surfaces insights before you act before you ask. And co CoWork dashboard gives knowledge workers a persistent home base in CoWork, also coming soon. And then on the interoperability side, we have Salesforce zero copy integration and data connectivity proxy across AWS, Azure, and GCP. Those are both GA today, making it easier than ever to connect Snowflake to the systems where your business already runs. So if any of these are relevant to you, we have a ton of additional resources on the webinar platform. Please go take a look at those and explore more. Now, as I mentioned, the main content of the Pulse is really focused on what is PUPR or GA, what can you sandbox or use today. But today, we're doing a special edition extension of this Beyond the Pulse section. We're going to take a couple extra minutes to talk about two really exciting things that we announced today. The first is open weight models in Cortex, and the next is Cortex AI gateway. These are still in private preview, but they represent one of the biggest shifts in AI economics that we've seen. And I want to make sure that you walk away with some information. So in the spirit of that sneak peek, I'm going to hand it over to Paras Mehra to speak about those brand new exciting features. Paras, over to you.

Paras Mehra: Thank you so much, and great to be here with you all today. I am very excited to talk about some of the innovations that we are bringing to you all with respect to open source models. Now as you've probably seen, it seems like almost every day, we are seeing a new open source model that is pushing the cost performance frontier further and further out in a way that is beneficial for the end customer. We're really excited to be bringing DeepSeek v four Flash in private preview within the Snowflake security perimeter and coming soon, GLM five three. Now we chose these models because we're seeing that they are just completely crushing it on a wide variety of different tasks that we think are very relevant and applicable for Snowflake customers. This is, like, software engineering type tasks, data engineering type tasks, data analysis. These models are very, very capable. And what this means is that in many cases, you are getting lower cost per task or token, but able to achieve the same level of performance that you get at the sort of the frontier spectrum today with some of these closed source models. The other thing I'd like to call out is is that these models are hosted within the security perimeter. We are hosting these on GPUs that we've provisioned and are ensuring that they are running with the utmost security bar, such that you can trust using these models with your most secure data. Now the other thing that I want to add is is that we're also layering in dynamic model routing capabilities within Snowflake. What does this mean? So as you see more and more of these models pop up, what we're finding is is that each of them are actually quite good at different things. Now navigating this space of which is the right model to pick at the right cost is very hard to do at scale and expensive. Running these evals is nontrivial, and you have to constantly keep up with every new model release. So what we're shipping is this dynamic routing capability such that you, the end customer, selects for either your first party applications like CoCo, CoWork, or your third party agents. Let's say, like, you know, whatever coding agent you're using today, you simply point it at the auto model, which is our router model. And we will, behind the scenes, route that model to or that that request or task to the best model for the job and pass on those savings to you and ensure that you're still hitting a very high quality or performance bar. So with that, I appreciate you guys' time, and I will turn it over.

Umesh Unnikrishnan: Hello. Umesh Unnikrishnan, and I'll spend the next ten minutes talking about some of the innovations that we are planning for Snowflake CoCo. Before we go there, I wanted to double click on what Paras just said. You know, he just introduced us, to the amazing open source models that Snowflake will be hosting. Many of those models and automatic routing will be available in Snowflake CoCo, so within Cortex Code as well over the next few weeks. So with that out of the way, I thought we could spend a few minutes talking about governance and AI governance in general. So not features, not model quality, but the thing that I hear often from when I talk to customers is that the thing that's standing between having CoCo in a pilot environment and between rolling it out to thousands and tens of thousands of developers and users is that you want to be able to manage it at scale. You want to be able to manage security, costs, and all of those things. And in many ways, the way I think about it is AI is at this maturity level that cloud was at a while ago. So cloud gave us you know, that discipline didn't really exist until infrastructure spend became unpredictable at scale. And then suddenly every company wanted to figure out how to, like, optimize, how to govern, how to, protect and manage all of this at scale. And so, when I talk to, customers, the things that I'm hearing is, you know, the finance department wants to manage costs. The platform teams wanna make sure that, everything is governed, and they have the security team wants to make sure that there are controls in place, data is not going to the wrong place, the IUIs protected. And, of course, the developers actually want more context. So given those, you know, the key thing to remember is that CoCo is built on top of the Snowflake platform, and the Snowflake platform provides a lot of these capabilities built in. And as we advance the platform to support more AI capabilities, we want to bring along all of those governance and cost management kind of capabilities to that. Right? So, four features that we are announcing on the governance side of things. First of all, per user quotas, you know, we've had budgets for a while now for a couple of, months, I believe. And that's been very popular. Customers have been adopting it, and they've said, oh, yeah. This is great. Now I could set the budget for a team, but how do I prevent one person from just, like, blowing all of that up? Right? So so for that, you also wanna have controls at the user level. That's what per user quotas will enable. And then, and I'll talk a little bit more about this because a lot to unpack on the AI governance suite, but being able to manage organizational team and individual, deployments of Snowflake of of CoCo is super important. And I'll get into a little bit more detail in a in a slide that's coming up. And then we announced the AI gateway product today. And what that does is give you control access, give all of the AI products that you're hosting, not just the Snowflake ones, all of them control access, to various MCPs, tools, and so on. CoCo will support that out of the box so you can manage MCPs and you can, deploy them at scale, in a fully managed governed manner. And then finally, with the growth of AI coding tools in CoCo, what we've started to see is that everyone's now building their own skills. And we've also introduced the notion of plug ins, which is a group of skills and and MCP bundled together. And our customers, are are creating those skills and plug ins, and now they would need a way to govern and manage those. And that's where the, skills and plug in sharing comes in. I will not be talking about two upcoming features, but I did wanna tee this up, for a future session. And, you know, please reach out to your reps if you are interested in participating in the private previews of the upcoming Slack app and the ability to do automations, from the, CoCo command line. So those are coming up. I'll not be spending much time on any time on it today, but please reach out if you want to learn more. I think the slides are taking a few minutes to load. But yeah, the first thing I wanted to talk about was around, cost control with per user AI quotas. Right? And this is the situation that I mentioned earlier, which is being overshooting is fairly common. And there's a story where, you know, a customer came to us, and they're like, okay. One of our users has spent spent thousands of dollars, yesterday. They are blowing through the monthly budget that we have set. How do we set per user quotas, on a weekly and monthly basis? And now that's actually, available. And, you know, budgets at aggregate and resource level, they catch team wide overspend, not the single outlier. And then per user coders, you know, you can set them, at the monthly and at a daily credit limit, and they're independently evaluated. And then the scope defaults to every user of the account, no tagging or anything. Then you can also narrow it tags, which is all dynamically. So but otherwise, it should just work and manage your budget across CLI, desktop, Snowsight for CoCo, and but also AI functions, agents, and other things. Right? So that's the high level overview. Oh, one cool thing I wanted to highlight. The default is to actually just cut off access with your head to one hundred percent of the budget. But if you are interested in doing things like at ninety percent, I wanna send a a warning, that's possible as well. So that's AI Codas. Now let's dig a little bit more into the governance layers because this is something that we hear about a lot. Let's start with and the way we've built it is this there's a layered approach where you start with at the organizational level, what you want is for assigned OS protected config profile. That gets pushed out by whatever MDM tool you're using, like using Jamf, Intune, whatever. So there, you can do things like specify which skills are allowed or what models should be used and those kinds of things and just push it down at the company level, so that even before any, data gets sent to Snowflake, like, this is it's a best client side enforcement if you're using CoCo CLI or desktop. The next layer, and this is new, is the agent profile. And what this means is that now there's a config object, CortexCenterScopeCode dot config, that gets set at the team level so that you can grant a team access to a specific set of models or make the default for a team to be a different model. So for example, here at Snowflake, what we've done, as we've said, hey, everybody at the company gets a sonnet class model by default, but, you know, the data engineering team gets an Opus class model. So those kinds of things are now possible via agent profiles, and, again, follows the same RBAC and roles and all of that that Snowflake has. And then the next layer down, and and this is new, and it's very nuanced, so we'll have a lot of documentation explaining how this works, is what we call a restricted session scope. And this is at the query layer. Right? So, what this means is many of us operate with multiple accounts or with multiple roles. And I'm an account admin on some on a specific account, but I'm also using a product manager role or a data analyst role while working with things. What I don't want is for the agent to actually switch into an account admin role and go delete some databases or drop some databases or delete some users or things like that when I wanted it when I'm actually doing my data analyst role. Right? So this, what it does is actually allows you to scope some sessions down so that you can say, okay, for this specific session, I only want it to do selects on these tables, views, streams, stages, whatever. Right? Added, I wanna exclude stored products, or I don't want these roles to be enabled. So it's a way to, like, define a restricted session that the agent can run it. Right? So, obviously, agents are powerful. That power comes with a lot of risk. And the the here, these are the kinds of things that RSS will, protect you from. Again, we'll have tons of documentation on this because it's it it is a little complex, and we wanna make sure that people are able to use it effectively. Right? Cool. On the AI gateway side, I will just breeze over this because a lot of this is covered in other sessions. But the AI gateway product from Snowflake is now, available. I believe it's in private preview. This came as part of the acquisition of, Natoma that we did about a couple of months ago. The interesting part about this for Cortex Code is that MCPs can now be defined within the gateway and can be pushed down to Cortex Code on the desktop CLI and desktop CLI and soon within SnowSight. So this is extremely useful because as an organization, you can now define a group of MCPs and the roles that are able to access it. And it's available to every user of CoCo automagically. All right. And then finally, let's talk a little bit about skills and plug in sharing. You know, the last one, this is the one that take turns your individual productivity, the skills that you've built and things like that into organizational productivity, organizational leverage. So, yeah, if people are writing skills, it compounds fast. We've seen internally that, you know, there are something like ten thousand plus, I think it's now twelve thousand skills that Snowflake employees have written. So way more than one skill per person. A lot of those are duplicates. A lot of them are very similar to each other. And then we thought about to get and say, okay. We need a way to, like, govern these at scale. So we built skills and plug in sharing. So now what if I built a skill and I wanna share it with my team, what I do is, I publish that into the, into the Verizon catalog, and then anyone can now install that skill. If I just want people in my team to get that, I can restrict, it via our back so that only my team can use it. If you're an account admin, you can certify some skills, and you can only certify skills should be utilized by people in the organization. And that allows that. And that makes that kind of sharing possible as well. Cool. I think that is all that I have, And I think we'll take a bunch of questions.

Rachel Roundy: Yes. Thank you so much, Umesh. That was a great overview of all the goodness in CoCo. We have a ton of questions. So we'll take a few live. If we don't answer all of them, everyone, we will answer you in the tool. We can also connect offline as well. And to answer a couple of questions in the chat, absolutely, this will be recorded as well. Okay. So you ready for the first one, Umesh? So do I need to set the secondary role as a default, or do I need to disable it?

Umesh Unnikrishnan: You do not need to do either of those. I mean, with restricted session scope, you can specify what you want to do, right? So you can essentially say that do not use account admin, do not use sysadmin. You can essentially block the high privilege roles by doing this.

Rachel Roundy: Awesome. Okay. The next one, are the monthly and daily credit limits limiting output tokens per agent, or does that mean something else?

Umesh Unnikrishnan: That is what it means. And you can set the, level at which they get applied. So you can say this applies to CoCo. You can say it applies to CoCo Snowsight, or you can you know, there there are various levels. Essentially, everything in Snowflake anything that's a Snowflake object, you can restrict the amount of credit spend that happens along with that. So it's not necessarily output tokens, but it's more on credits.

Rachel Roundy: Awesome. And do per user quotas only apply to human users or also to agents that might be using Snowflake?

Umesh Unnikrishnan: It can apply to anything that is a Snowflake object, which includes users, roles, anything like that.

Rachel Roundy: Awesome. Okay. And last one for CoCo. For skill sharing, what is the trust model? Can a shared skill make outbound network calls or access data the consumer hasn't explicitly granted?

Umesh Unnikrishnan: Yeah. It's a great question. So the thing to know about skill sharing is once you actually import the skill, it runs as you because it does it's it's it's as if you had built the skill or anything locally. So, you know, any restrictions that you have set up will still apply. And, of course, within CoCo, within Core Work, everything is controlled by RBAC, the things you do run as yourself. Awesome.

Rachel Roundy: Oh, and one more just because this is a really good one that popped in. We have per user quotas in CoCo, does that limit their ability to use agents within CoWork or a Streamlit app that is using Cortex functions? So does

Umesh Unnikrishnan: that have question? No, it should not. So you should be able to specify that this is only for CoCo. And if you you can see early versions of this right now. When you just go into your cost management page, you can see that there are different services, and you can essentially set these quotas on a per service basis.

Rachel Roundy: Amazing. Thank you so much, Umesh. Thanks for joining us to talk about Snowflake CoCo. Have a great rest of your day. All right. We are up. Next, we have machine learning. So we'll pass it over to Avanash. Avanash, over to you.

Avinash Joshi: Hey. Thank you, Rachel. Good morning, good evening, and good evening, everybody. I'm, happy to be here and chat about machine learning and the updates that have happened in the last month. Slide. Sorry. Firstly, the Snowflake ML, as a just a quick refresher on what is Snowflake ML platform. Snowflake ML is our end to end platform where our customers can manage the entire lifecycle of their model right from data management to training their models to deploying their models and monitoring them. And it's currently it it it has been a huge success over the last few years where we have about six thousand plus enterprise customers who are our trusted ML partners, managing millions of models and trillions of predictions per month, structured by some of our some of the biggest names out there where we have we are the first national bank of Omaha who which who's who reported 10x increase in their productivity using using our agentic ML platform. PayPal has done about fourteen million real time transactions every day, automated through, with ninety seven to ninety nine percent accuracy. And Atentev has modernized their entire marketing personalization at a billion row scale, delivering sixty eight percent speed up over their legacy feature stores. So what is the platform? So quick look again at the end to end tech stack. Your Snowflake ML platform can be accessed from the using either the Snowflake CoCo, which which we've already heard a lot about in this session, or using Snowflake notebooks or external IDs. These are our usual interaction points for our ML features. And the features themselves span across the life cycle. Right? For for development and training your models, we have our Snowflake ML APIs, experiment tracking, and cortex training. Orchestration and automation can be done using ML jobs. And you can manage all these artifacts that you are generating using the model registry and the feature store. And you can deploy and serve these models and features in either in batch mode or in real time. And then you can, of course, anything in production has to be monitored and measured and explained. So you set up money model monitors, and you can when models go awry, you can, of course, use explainability to understand what went wrong. And all of this can be deployed on two of our compute workloads that we compute substrates that we have, the warehouse or the container on time on container services. So today, I'm specifically going the the the updates I'm going to chat about are specifically around the ML ops category. So we'll be talking on experiment tracking and monitoring pieces a lot more with a quick update on the batch inference. So first, let me start with the experiment tracking piece, which would fall in the development category. So what is experiment tracking? Experiment tracking is pretty straightforward. When you're developing your models, you are training you are trying multiple things. You're trying different model architectures. You're trying different sampling strategies. You're trying different features, etcetera. And as you're doing that, a ton of this data and the metrics and parameters that are generated need to be logged in catalog for your analysis later, which helps you find your best model, which helps you find your best helps you collaborate with your team members, and overall is a best practice for for for your m from an MLOps standpoint. Now, of course, this feature has been there in Snowflake for a while, but there are some cool new updates that have come out in the last month. Along with specifically, along with the met model metrics, parameters, and artifacts, now we automatically capture your logs that are generated when you are running when you are doing a training run and the source code files as well. So what that means is if you have started a especially to in today's day and age where the training runs are not short, they can run for hours or days, you require the logs stay with the experiment. You cannot have a different place where you want to go. I mean, let's say, you're training running crashes, or you want to figure out what hap what was what happened in a particular run that happened then five days ago. You shouldn't have to scramble around finding, hey. Where is my source file? Where are my logs? Where are my parameters? All of that needs to be available to you easily at one place. And that's exactly what we deliver with our experiment tracking. There is zero operational overhead. There's no servers to provision. You don't have to set up an ML flow separately. You don't have to or any other tool that you would like to use. You just create an experiment object, and then you get going. Or you can ask just ask CoCo to do it for you, and it'll just set it all up for you. And, of course, as everything that's within Snowflake, the built in governance controls automatically apply as an experiment is nothing but another Snowflake object. So whatever our bad policies you set up on the database and schema will apply to the on all the underlying objects. And we'll see more of this in a quick demo that I'll show in a minute. Next, I want to quickly touch up on the batch inference, piece that's, that the in that we didn't talk about. We're talking about benchmarks in the last couple of of Pulse events where we were talking about how Snowflakes benchmarked extremely well against competition across training, feature store, and inference. But last time when we spoke about it, we're primarily talking about online inference. Now we've gotten benchmarks on our batch inference as well that we would like to share. And on batch inference, we are even better. We are almost up to six point five x higher throughput and up to five point eight eight x lower compute cost per million rows for batch inference. And this is across our compete for across a major competition. Using this other that's and this feature can is not only unimodal, also works for multimodal models, and it can scale up to tens of billions of predictions every day, which has been proven across our customers. And again, as soon as it's an agentic ML platform, all the deployment, how to make sure that your the tuning is set up correctly, making sure that the mass your inference job is running most optimally so that it's getting the most out of the compute underlay underneath is automatically handled from you for you either through CoCo or to the platform itself. So for batch inference, all you're calling is calling one API that says run batch, or if you're using SQL equivalent execute inference job, and that's it. You don't have to sit and figure out anything underneath the computer, either at the compute either the hard, infrastructure layer where you don't have to figure out what should be the cluster, which machine to use, or you don't need to set up the Docker containers or figure out what dependencies to run, etcetera. All of that is taken care of by by by the by the platform for you. And finally, on the back to them on the MLOps team is, in the last event, I spoke about how the online ML has is a mature online ML product that we offer today, where you can set up your models to serve on real time inference along with integrated feature store, online feature store, where the features are coming in and these predictions are happening automatically. And these feature stores, of course, helps you all making sure that your training and your production data looks similar because you're using the same feature store and which brings down your round trip latency by a lot. So you can serve features under thirty milliseconds. This is and these are p ninety nine numbers and models under one hundred milliseconds, the p fifty numbers are much lower. And then the cool new update that has come out here is our ML observability doing AB testing on these models is going GA has gone GA this month. What what AB testing lets you do as a as a lot most of you must be familiar, AB testing is a way of making sure that when whenever you're upgrading your model in production or you're putting out a new model version, you need to ensure that your new model version is better than the one that's already exist. So that's like the champion challenger model. How do you ensure that when you're upgrading your model, you ensure that it it needs to be data driven. It cannot be based on your gut feel. And you how you do that by deploying both the models at once and then either splitting the traffic between the two or or doing it in shadow mode with the entire hundred percent of the traffic goes to both the model versions and then see how which model is doing better. You have a very easy way of doing a one click rollback, but save new model version is not doing well. You should have an easy way to dial back the traffic with the old model version that you already had in production and. It. And this gives you a unified observability where you have a single pane view of job health. How is your model doing? How is the ingestion throughput? How is this serving latency? All of that available to you in a single view in Snowsight. I'm gonna show that all of this in a demo. I'm gonna quickly grab the screen share. And then once it switches over, I will yeah. Here we go. So in this demo, I'm in the interest of time, I'm not going to go through the whole thing. I've already ran the notebook just to save time. But what I'm doing in this notebook is we're gonna use our classic fraud detection example where I'll be training a couple of models, like simple models just to keep keep them small and quick. Using experiment tracking to log these, I'll show you how we automatically capture everything without you have without the user having to do anything, deploy these two model versions behind the gateway, and then do a run and AB test on them. So first, I'm gonna jump directly to the training cell because the rest of the thing rest rest of the portions are primarily setting up the data. And once I have this function now ready, which can now train models based on as and when we want to switch as and when we want to sorry. As as it just takes inputs as your your training data as your input and then trains the model. Now instead of just directly training the model, what I'm going to do is I'm going to create an experiment and set in within that experiment's context, I'm going to run the training job. And right now, all I'm done is I've set the live logging status to true, but in future, we'll just enable this by or to prove by default as long as they're running it within software. Now once I have done that, I can, of course, do it a couple I can do it for the second world log model version, which would in which I have changed the some of the parameters and the and the sample and the sample that I'm using for training the data. Now once I have these two models trained, we can quickly switch over to the experiment tracking UI, which shows you which which you can be accessed using our Snowsight. On the left now, you should find it under the machine learning section. You'll have something called experiments. You go there. I'm gonna look at my fraud detection pulse demo experiment, and you see that there are two runs here already available. Now I can click into the run and then look into more graphs and charts that regarding the metrics that I'm using. Of course, it's a simple model with just a single metric, which is just showing you the f one score, ROC, accuracy for it. And I can, of course, come I can also compare the two runs and see which which run was a better run based on whichever metric I want to see. But this is existing functionality. The new things that have come out is you can also check the logs of this run. So for example, if I wanted to see, hey, for my train first training run, what are the different logs? It looks like there was some error while this was training. This sample input has some to which which is very handy for me to easily find out where say, I have done this training run. I go for a coffee, come back, and if I lose all the logs, or if I have to figure out my own logging mechanism, then that's that's suboptimal. Instead, all you do you're in Snowflake anyway. You're using a notebook. All you have to do is just say say with experiment on top and then run then line to write your training run, everything else is automatically logged for you. The other cool thing that you can see is in this experiment tracking, when you go to details, you will be able to see that, hey, for this for a particular run, which database which model which model was generated in this training run, and more importantly, which notebook was used as an entry point for creating this run, which is help which is super helpful to track. Hey. Once I have a model in production, it's not doing well. I want to get back to the source code which trained this model. You can track it complete back to your experiment and then back to the actual notebook that was used to build this model. A good question here would be, what if I'm not using a Snowflake notebook? What if I'm using it on my local machine or some or on on a Jupyter notebook? Well, as long as you're using, a Git project, we'll automatically capture the Git commit and the Git hash for it. And we so that you can always get to the source file that was used for training that run. Cool. So that was experiment tracking. Now let's move to the production side of things where these two models that I have trained, I'm going to deploy them for product as as online inference. For that, let's go to our models. In your in your models, we should be able to find our models here. And we can go to the inference services, and you can see that for this for the fraud model, I have these two inference services running. One was the one what I'm calling is one that's already deployed in fraud, and the new second version is what I'm calling, as a canary service. At this point, you could, of course, individually use these services if you want to. But what we recommend is you create a gateway on top of it. And you can, of course, create a gateway either using the UI or as everything in stuff like everything that can be done using a UI. You have a SQL equivalent command for it. So you can just create a gateway. You can pick the database and schema you want this to go in. I'm just gonna pick some database schema and give it a quick name, and then you can start adding services. So you can add whichever service you you the two services that we have created, you can just add them here. So in the interest of time, I'm just gonna use the one that I already created while prepping for this demo. So I already have a gateway called fraud detection x g boost gateway. And if I click into it, you will see that the two services that I was talking about, the fraud service and the canary service, are part of this gateway. So what this lets you do is instead of using the REST API endpoint of these individual services, you use the gateway service, which remains fixed. This is what the external application is going to see. So for this gateway, this is the URL that I use in my REST call. Or but you can set all these requests are coming in, you can set how do the traffic should split between these two. So as you can see, right now, it's set up such that the prod model version that's already there is getting seventy percent of the traffic, but the thirty percent of our traffic is being split to this new Canary service that we have set up. We can, of course, split make this into a shadow mode as well where all the hundred percent traffic goes to both of them. But if you're in when you're in traffic split mode, you need to ensure that this sums up to a hundred percent, which is enforced by the product. Now as and when this keeps seeing data, you can you can switch to the monitoring tab and check whether, hey, and start monitoring them. So let me quickly pick a time frame in which we already have data for this. Let's say June and July. And you can see that we get a quick snapshot view on top, which tells you, hey. For this time period that you've chosen, your prod services has a f one score of twenty one percent, but your Canvas service is slight doing slightly better of twenty five percent, which both the numbers are much lower than we would want in general, but, it's your Canvas service is better than your existing service. You can also see the trend. This is this is the full snapshot view, but you can also see with overall, are you doing across time? The new service is performing better than your existing service, which is the baseline. Now that you can now that you have the confidence that your new service is doing better, you can now go back to the overview and then change this, you can split this change this traffic to be zero like, you can dial the entire traffic back to the Canary service and make that your new production model. So you've done a quick model upgrade, but you have done that with the checks and balances that and a good MLOps practices dictate for for putting models in production. That's my demo. Thank you. We can switch take some questions.

Rachel Roundy: Amazing. Thank you so much, Abhinash. Okay. We have some great questions from our audience here. The first one, does experiment tracking allow you to track track production model accuracy, compare predictions to data once available, or does it only track training model metrics?

Avinash Joshi: It's up I mean, experiment tracking is primarily used for training usually because you're manually logging and setting up. I mean, you you choose what metrics you care for when you're training the model. And, of course, the same metrics are typically tracked for production. But for production, you usually would use something like the model monitor or an AB test. If you want to do an AB test between the two, you would set up something like a gateway and a monitor on top of it. And that's how I would do it. You can, of course, do what you can use experiment tracking for also measuring these. But if you use the monitors and observability, if you see the metrics are automatically computed. I never wrote any logic to, hey, please compute this f one score for me or etcetera. It was automatically being computed. So that's the beauty of you. You get that out of the box. In experiment tracking, of course, you are computing the metrics or you are whatever because we can and experiment tracking is a generic product, so you could be training any kind of models. We can't really pick what metrics to compute.

Rachel Roundy: Makes sense. Awesome. Do we have any post deployment features available right now, like data drift, performance degradation, etcetera?

Avinash Joshi: Yeah. Excellent question. Like, actually, if you're in the UI, I think I didn't scroll it so so long enough. But if you look if we don't just do performance metrics, we also do drift metrics. We have counts. So if you in general, if you just want to see, hey. If my number of predictions are going below a certain threshold, I want to get alerted. You have all those functionalities as well, where you might want to set up, hey, some sort of hey, I have this I have this model, and I expect at least ten thousand predictions going to the service. And if it's not going there, then I would like to know. So write counts, performance, drift, you can set up data quality. Have we have Snowflake, so it's all kinds of metrics.

Rachel Roundy: Awesome. Thanks. Okay. Next one. Is it reasonable to leverage experiment tracking with Ensemble modeling approaches?

Avinash Joshi: You can. So an experiment is basically a container in which you have runs. So you can it's up to you how you what you what unit you want the run to represent. And where, for example, if you have each run representing a certain model type or a model strategy, you can do that. Or even given a run, a lot of times our customers do and when they're doing, say, something like an HPO, like a hyperparameter optimization, that can generate a ton of runs. So you might want to figure out, hey, I want to run the HPO, but not log every model, like just the top ten models. That's also an option.

Rachel Roundy: Awesome. Okay. Can I use experiment tracking if my training and production model is saved in a shared workspace?

Avinash Joshi: You can. So experiment objects are scope experiments are nothing but an object scope for database and schema. It doesn't really matter where is it originating from. Shared works versus a regular workspace. If you have local laptops, everything is fine.

Rachel Roundy: Awesome. Okay. So last question for you, Abhinash. I've heard about your AgenTik ML offerings. Does CoCo know how to configure ML ops features like experiment tracking and AB testing? I think we've already covered that, but I'll let you tackle it. Yes. Yes.

Avinash Joshi: We have specific skills for machine learning in which there are subskills for each of these features. So we have an experiment tracking subscale. There is a model registry subscale. There is an observability subscale. So all of and you don't need to, of course, invoke any of these just based on the prompts that you're giving it. Of course, the more explicit your prompt, the better. But you don't have to. The the the prompts themselves will trigger the appropriate skills and then help you set this all all of this up. You don't have to actually figure out the sequence and tags or any of that.

Rachel Roundy: Awesome. Well, thank you so much, Avanash, for sharing all of the amazing machine learning updates. We're grateful to have you here today. My pleasure. All right. Next, we're going to pass it over to Effie to talk to us about Snowflake CoWork. Effie, over to you.

Effie Goenawan: Yeah. Thanks, Rachel. Hi, everyone. I'm Effie. I'm product manager for Snowflake CoWork. I've seen some of you in previous AI Pulse. Super excited to be back. We have a lot of things that we are rolling out in the next few days in response to what we announced in the summit. So with that said, let's start with, like, how are we doing with Snowflake CoWork? We're doing really well when it comes to, like, the customer demand, addressing customers' needs. I talk to a lot of customers almost, like, four to five daily. I hear a lot of the excitements. And from CoWork itself, we have thousands of accounts. And just within AI, we have thirteen over thirteen thousand accounts using Snowflake AI features on a weekly basis, including top brand names like United Rentals and Wolfspeed, whom also present with us at Summit and Fanatics as well, among others. And next is we dog food our own product. So you've seen you've heard this also in previous polls and summit. We have Raven, which is our go to market agent for sales org. This is something that I use myself. So it's not just within SalesOrg, but we have about over six thousand people using this Raven agent. What it can do is it gives you go to market data. For me, I use it every day to prep for my customer meetings and then get the health of the customer and then how the product is going, the consumption is going. And the impact we've seen is amazing, and it continues to get better over time. What we've seen so far is over sixty million dollar overhead savings with about eighty FTE recoup productivity, and then we've seen over five x ROI based on the effort it takes to build Raven, the cost to run Raven compared against the time savings and ability empowering go to market employees to be able to be self serve and data informed. And as of today, we've addressed over a million questions with Raven. This is something that a lot of us use every day. And then Raven's capability has basically, like, ten x in the past few months. It went from, like, a reactive tool where you ask a question and you get an answer to scheduling, automation in even, like, admin brief where the manager can set up automation for the rest of the org and having, like, recurring schedules so the agent works before you even wake up in the morning. And Raven also has, like, beautiful charts and other abilities that has been rolled out on Snowflake cohort. And in Summit, we also give a major update to Snowflake cohort, previously known as Snowflake Intelligence. It's not about reactive q and a tool, talk to your data, but it's going to the one AI chief of staff that also gets work done for you based on your business context. We are very different than the general AI tool. We anchored on your business data, bringing context from your most precious Snowflake data, taking into account the governance and the security that you know and love from Snowflake ARPAC. Take bringing also additional context from MCP, whether that's your email, your team conversation in Slack, or your, enterprise context from your favorite apps. For us, it's Salesforce, Cheetah, but for you, there could be other MCP integration you can bring in to Snowflake co works. So it has all of these context, the valuable insights that make the AI a lot more useful. And we add a lot more abilities to Snowflake Corework to run code execution tool as well as generate documents like PowerPoint, PDF, and soon enough HTML, sending emails, filing tickets, actioning. So not just giving you an answer. It gets the work done end to end within the same products. Now if they CoWork, you don't have to take your data out to another platform or copy paste to other tools. We also bring on artifacts, which is allowing sharing and collaboration so that work happens within the team, not just yourself. Artifacts has seen since made generally available. So we're super excited, and we have started the private preview for a lot of these personal work agent capabilities. The most popular one being automatic routing. Back in the days, we suggested building multiple agents, whether that's sales, finance, ops, product to empower your data teams, domain experts to build an agent that you can eval and made observability possible. We heard you loud and clear that, hey. The business user do not wanna pick which agent to choose. Even if you wanna stitch multiple responses of different agents into an orchestrator, now we support that with the agent router or what we used to call main agent. The private preview was very popular. We had hundreds of requests, and we've heard feedback loud and clear. So we're working on that, which is optimizing the latency, the cost, as well as improving the admin experience. So that is work in progress, and it's coming in PUPR very soon. We are currently pausing the private preview by opening up once we address the key issues I mentioned, which is on latency, performance, and admin experience. And then that's the router capabilities. And what we have coming in public preview this week is PDF, PPT creation, as well as the user skills capability that will be unlocked with execution tool. And then all of these is also integrated with the MCP tools that are already available today. You can connect to your apps and services like Slack, Gmail, Jira, Salesforce inside your user's existing permission. So that's the one agent for every question and with the right tool. And, again, the same private preview for personal agent got a lot of really overwhelming, like, positive feedback and excitement. We are very thankful from our customers, including Whoop, who has been super active in this private preview. They're super excited about the code execution tool, which is coming in public preview this week. I'm super excited myself. Nowadays, I use CoWork to generate beautiful tech, beautiful doc. I don't have to copy paste the materials. I, myself, I dictate to CoWork what I wanna produce, and I connect it with my Snowflake data, and it creates this amazing beautiful PDF and PBT. I can even customize to my company's brand. So some of these collaterals for this presentation also generated on Snowflake work. And then user skills and memory. We've heard from many customers that people wanna codify the repeatable flow or what they call, hey, Effie, how can I save my prompt and have my prompt library so I can execute and run it repetitively? We're also making user skills available with code execution tool coming this week so you can codify with natural language your skills. Most popular one being prep me for my customer meetings. What happened is very much nuanced to my workflow, your workflow, and the team's workflow. Things like pull up my customer data from Snowflake, look at my email and Slack, summarize previous conversation with the same customer, and then pull up a specific sales or entry for the tickets that this for the use case for this specific customer. Look at the Jira ticket for tickets they filed, and then I come ready for the customer meeting. I use it every day. Super helpful. I even share it with my colleagues. And then user memory is something where a coworker can remember your preface and role. The fact that I'm Effie, I'm a product manager, and I work on Snowflake cohort. Snowflake cohort used to be called Snowflake Intelligence. So this is when I tell cohort once it remembers, so it personalized the experience for me, and it gets me more effective to do my job. So I spend less time on the AI tool, spending more job thinking about how I can accelerate the business growth. So this is what I mentioned before with user scale. What it does, it repeat workflows. I can codify my workflow with natural language. And then during private preview, we also had shared skills so people can create and collaborate on skills within a team team within the team. We still have shared skills in private preview, but we will turn the user skill into public preview in this week's release. And then user memory status is it's also in private preview. We heard from some customers that they want to have the triggering a lot more implicit, meaning you don't have to actually say, always remember I am FE instead based on my feedback about maybe my charts. I do not like pie chart. I prefer bar chart or line chart, or I prefer red if numbers are negative without having to explicitly dictate remember x y z or always do x y z. We want cohort to be a lot more smarter when it comes to remembering things on users' feedback. So we're working through that private preview is ongoing. We expect in a few weeks, preview will come. And then iOS app, this is a lot of like, our execs love this because people are always on the go and then also people in the field. You can easily unlock Snowflake cohort just with base ID. No need to type password, and then you can dictate. I use voice a lot because, typing on keyboard is pretty painful when you're on the go, and we are best in class when it comes to displaying beautiful charts, visuals, even dashboard on your phone. This is a huge deal because previously, to get all of these complex information, you have to go to your desktop, you have to go to your laptop, and then, like, enter a bunch of app. With smoothly CoWork mobile app, You can just unlock with voice can unlock with your face. You can just ask CoWork the question, how is the growth of my customer x y z? And then boom, right there, you can even create the PDF to prep for your next customer meeting with the stats from your Snowflake data about this customer. And then this app is already made generally available. You can download it today on the iOS app. And then my favorite, which is automation. So this is when we can get forward to do the work before you wake up. I get excited every time I can get help from a system that can work autonomously because it saves a lot of my time so I can focus on higher level, better things that only I can do. So automation combined with user skill unlocks this beautiful, robust capability, which is like you can have scheduled briefing sent to you every day in the morning to prep you for the day. Gone are the days when I spend thirty hour thirty minutes every morning going through Slack, email, looking at my tickets, looking at my dashboard. Instead, I can have automation to look for anomalies, look for patterns, identify opportunities, help me focus my day and even my week so I can get going ASAP instead of having to, like, go through different apps and services. Automation is already public preview. Documentations are live. You can search for it, and then you can set it up right there on Snowflake org. And dashboards, another customer favorite. When we talk to customers, nobody seems to be happy with their BI solution, the traditional BI. Right? The static charts that you go every morning, but we believe that there is some better way to do this than the static visuals, which is conversation agent first experience. We bring conversation and a genetic experience to both the creator and the consumer sites. On the creator, analysts can build with natural language on Snus site with CoCo, build beautiful dashboard, multi tile representation with scorecards, charts, tables, and easily publish and also make it govern the specific roles on CoWork for their business users. Their business users can easily find these dashboards from trusted analysts from Snowsight and CoCo then ask follow-up question. Now they do not need to file a ticket or ping somebody to get an answer to their questions. They can self serve these dashboards. This is currently in private preview for select customer, but the feedback we've heard is super amazing, especially on the creation creation site. We just hey. My data is already on Snowflake. I can just talk to CoCo, and I get this beautiful dashboard, easily publish it in a governed way to my business users on CoWork. Super excited. And even, like, from the business user side of things, people feel empowered when they can self serve. So we've seen really good feedback. So this is public preview coming soon to the overwhelming demand. So we're super excited that AI BI is coming to cohort for real very, very soon. And then Cortex Sense is our attempt to help accelerate agent deployment. We've heard feedback from some customers, some of you that, hey, Effie. It's a lot of work to build agent and deploy them for production grade with high accuracy. We've seen that some customers or some people also try to use their favorite coding agent, whether that's Codex or Clot using Snowflake MCP, but then they're seeing high latency, high cost, and low accuracy. This is not good. We wanna help people build agents and also get the most out of their Snowflake data, whether that's through CoWork, Cortex agents API. And we have Cortex Sense, which is our runtime managed context with the value prop of given just a few hours, not weeks, not days, you can get an agent working in production grade. This we've shown really promising results with, like, eighty six percent accuracy on hard enterprise questions versus twenty four percent just from Claude Code with Snowflake MCP. So we are really excited. We've seen, like, amazing adoption internally at Snowflake. We use Cortex Sense ourselves rather than building a lot of semantic views that can can be a lot of effort. Cortex Sense essentially is our attempt to give you some sort of, like, out of the box experience to get an agent in production. Great. This is PUPR coming soon. We are at private preview right now. So just to recap, a lot of things happening with Snowflake core. So just to remind you, there are many things already in generally available state. We just try it out today for everyone. Deep research to give you in-depth answer to your most complex why, and then artifacts for you to be able to get beautiful robust visual that you can save and share with your teammates. And then iOS mobile app, my favorite. If you have not already, please go to the iOS app iOS App Store and then download Snowflake Core app to access these amazing insights on the go with voice, with face. Right? The best natural interaction ever. And then coming in PUPR this week is code execution with PowerPoint and PDF generation and then followed by HTML very, very soon. And then also async agent API, allowing your agent to run more than the few minutes that is allowed today so agents can work in the background. And then also automation is in public preview, available widely today. Try it out. Get the agents to work on your behalf while you are asleep. And then also POPR soon is user memory for personalized experience, dashboard, which is the AI first BI, and then also Cortex Sense accelerate agent development for production grade agent. With that said, try it out today. Use our quick start with our blog post. Ask us any question. We're super eager to unlock this powerful ability for you. And then these are trusted industry leaders we've been working with. We're very thankful for them, and we are we cannot wait to unlock even more partners and customers. With that said, we will take it to q

Rachel Roundy: and a. Thank you so much, Effie. There was so much goodness there, we have a ton of questions. So the first one for you is, does the PowerPoint generation allow for use of templates?

Effie Goenawan: Yes. Two ways to do this. The user can upload the template and then ask the question and code directly. This is what I normally do. And as an admin, you can also create an agent skill with a template called using Snowflake's skill. So that those are two ways to unlock. Awesome. Perfect.

Rachel Roundy: Thank you. The next question is, can the skills and workflow tools be invoked if using the Cortex API rather than the CoWork front end?

Effie Goenawan: We support skills for the court the agent's API as well. So yes. And you can search for documentation. It's also available in GA right now. Right.

Rachel Roundy: Okay. This next question, can we add orchestration instructions or response instructions as a skill in CoWork Agent?

Effie Goenawan: It's not as a skill per se. It's more like the agent instruction itself. So, yes, you can do that with the the automated router that we mentioned, what we used to call main agent during the private preview. So you can specify at the top level agent the instruction on how you want the routing to be. For instance, if you have multiple, let's say, three agents, you can specify for revenue question directed to the finance agent, but for product question directed to the product agent. And then for book of business, use the finance and the sales agent. So you can specify that in the agent instruction. You can also create agent skill to do specific workflow, like how to do book of business. For instance, a, get the product data from the product data science agent, b, get the revenue from finance, etcetera, etcetera. So many ways to do it. One is specify the routing wallet routing instruction in the agent instruction or description. Description. The other is doing skills, which is more like codify workflow to do specific workflows for your business. Makes sense. Awesome. And I

Rachel Roundy: think you've gotten this one a lot of times, Zaffy, about when will the mobile app of CoWork be available in Android?

Effie Goenawan: It's coming soon in private preview. We have an early build available internally. We want to make sure we sort out all of the bugs and issues and get them ready for you in private preview probably in a few weeks. Alright. That's exciting news. Okay. How long are are

Rachel Roundy: artifacts saved in CoWork? And is there a retention period? Can you, as an admin, affect that retention period, if so?

Effie Goenawan: An artifact saved in cohort doesn't is not does not get auto deleted, so it's based. But then as an admin, you have the control to disable sharing or to monitor these artifacts.

Rachel Roundy: Awesome. Okay. Let's see. We've got one more CoWork question for you. Can the agent router be configured with custom routing rules, or is

Effie Goenawan: it fully automatic? You can customize based on the routing instruction that I mentioned earlier. And, also, what we do automatically is we look at the agents that it routes to, look at the the agent description as well as the capability. So there is some work that we do on your behalf, so making sure that the the router agent is aware of what all of the sub agents are capable of. That's what we do on our end, but then you can also add customization on the instruction for the router. Amazing.

Rachel Roundy: Well, thank you so much, Effie, for joining us today and sharing all of these incredible CoWork updates. We appreciate you being here. Thank you.

Effie Goenawan: Yeah, thanks, Lejio.

Rachel Roundy: All right. And last but not least, we have an amazing update from Hanna Mykula on AI functions. Hanna Mykula, I will pass it over to you.

Hanna Mykula: Thanks, Rachel. Before we go into the new feature of AI functions that we have launched, I just wanna do a quick recap of what AI functions are so that those of you who have not used it yet, maybe you will will sneak a peek afterwards. So, yeah, functions are the foundation of transforming unstructured data into insights natively in SQL. So you can let the functions easily analyze all of your data regardless of the modality, so this is very important. They're not just text based. We also support image, audio, documents, and now video as well. And instead of moving the data, AI functions allow you to analyze your data in place directly where it lives already on Snowflake. Now to the new things. I think there is a bit of a delay. Alright. Every media company owns thousands of hours of videos, and most of them most of them is a big library of videos. However, you can right now search the title, you can search the description, maybe some metadata, but you cannot right now easily search what is inside the video. Let's take maybe one concrete example. For example, advertisement of energy drink or maybe of a trail mix. When the advertiser is buying a place in the film, they're buying not just a slot. They're actually buying an experience so that people could people are more likely to buy the product because it's associated to the moment that this actually belongs to. And we're gonna see that later on in demo. Today, finding that exact moment where the ad fits in the video means there is a person who is watching a two hour long film. And, of course, this process is very slow, it's very expensive, and it doesn't scale. And the reason is simple because video is never becoming an actual structured data that you can search over, and this changes with the with the function that we are launching. But before we dive into the function, I just want to bring up something really important. And this is the file data type feature, which we have g eight since September twenty twenty five. Because most often when people think about AI functions, they, of course, think about text. But as I mentioned before, we support documents, support audio image files, and the reason of that is that we have native support for the file data type in Snowflake. File is just a simple Snowflake com type, which holds a reference to the file on stage plus its metadata. So that's why multi model analytics using AI functions become super easy. And now to the announcement. We've introduced last week a video search using AI Multimed. So AI Multimed automatically segments your video, scenes, and returns visual audio and transcription embeddings for each video segment. So here you can see the example of the AI Multimed call where you just need to point to the right file. You can specify the options which embeddings you want. If you really just wanna focus on the visual or audio, you can do that here, and you can also specify the length of the clip. The cool thing is that it's powered by the twelve Labs Marengo three point zero model, and it supports videos up to six gigabytes and up to four hours in length, which means you can really do a proper video analytics and not just some some POCs and shorter videos. The way AI multi embed fits in our portfolio is that you can easily pair it with the rest of the functions. For example, AI complete has been supporting now video for a while and it allows you to process, for example, the whole video asset to write the summary and extract any attributes. Well, with multi embed, you really get the searchable corpus of embeddings over all the aspects of the video. And that combined with Cortex search, for example, where you can index all the generated embeddings using a multi embed, and you can feed it up all the way to Cortex agents and use it through CoCo or CoWork, for example. What are the actual three most useful use cases that we see customers using AI mobile team ads for. So the first one and is kinda the most the most obvious and simple one is the contextual advertising, which answers the question, where does an ad actual actually belongs? Because advertisers want to buy a moment, not just a slot in the video. And today, as I mentioned, someone just watches the footage to find this. And with the AI multi embed, you can actually get a proper recommendation associated with the with the moment that your ad should be placed at. The second one is scene retrieval. Using AI Multimad, you can take a large library of videos. For example, it could be a video of the past advertisements for your brand, and you can use the Cortex search in the AI Multimed to find the exact moment maybe when your brand was mentioned. Or in this case, we searched for every ad that referenced recovery. And if you zoom in a little bit, you will see that the output format is the top matches with the rank and with the second in the video when they have. And the third one is the segment of the safety and compliance. So we know, for example, lots of our customers have facility cameras that record everything that's happening, let's say, on the shop floor or on the factory. So with the AI Multivat, you can also search for the moments where workers, for example, are not wearing helmets just to flag that and to see which moments should be reviewed and addressed. Few of the benefits of a m multi embed. Well, first of all, you get the segmentation in the smaller clips, which makes the search of the fine granular moments in your videos really easy. You can embed it once, put it in the cardiac search, and reuse it across multiple multiple downstream scenarios, which of course means cost saving. You need to need to reindex every single time. Your data, of course, never leaves Snowflake. It's governed by design. And you can combine unstructured data with structured data and use it up for your downstream use cases. With that, I really would like to show you a demo of how AI multi embed works in practice on the example of the ad placement and contextual advertisement. Alright. So at the beginning here, we have the dataset that we actually get from the advertisers with two company names and with six products, and also we attach the product description with the keywords. In this query, we're now embedding the video file, which is open source cartoon of around eleven minutes long, and we're also placing this in the Cortex search so that we could reuse it for any other use case later on and not have to reindex it again. And the step number three is gonna be where the whole magic actually happens, where we're gonna run the search using the index video to find the top hits where we should place our advertisement. And here we have a bunch of advertisements from, for example, Camp Lantern, Trail Mix, and a few other ones. We're gonna see the output of what it get in a little bit. So now you will see that we actually get the product alongside with the recommendation where we should place the ad and the segment, and also a brief explanation why that is which keywords actually matched to the moments from the product to the video. And just to see that this is all real, let's check the best retrieve scene for the for for the camp planter at placement. And you see here, for example, it's a group of friends camping around the the campfire, which is actually a pretty good fit in my opinion. And on the second example, the trail mix, I think the segment was retrieved at minute seven, eleven seconds, and you see the girl is actually having an open backpack of of snacks and opening a pack of peanuts, which is also a great fit for the trail mix ad placement. With that, that will be it from my side. This was AI Multimedia from AI functions that will help you with your video analytics. Amazing. Thank you so much, Hanna Mykula. And thanks for joining

Rachel Roundy: us for your very first Snowflake AI Pulse and not your last, I am It's great to have you here. Okay. We have a few questions. The first one I want to address, somebody had asked in the chat, they want to pass the model to AI complete and other AI functions as a variable from a query. And currently, only a hard coded string works. So he's flagging that he has hundreds of different areas calling AI complete to make model updates as they end their life cycle and new models become available. And he's wondering if we will ever have a feature that allows those to be stored in a central table.

Hanna Mykula: One second. So currently passing the file as the hard coded string or as a model? I think the model. The model. Well, we have something that we're actually working on right now, but I I cannot tell you tell you more, unfortunately.

Rachel Roundy: Yeah, no worries. So that was a tough one right out the gate. But we will respond to that offline. Okay, the next question, are there any cost controls for AI functions?

Hanna Mykula: Yes. We have user quotas, which is actually GA since, I think, last week, if I'm not mistaken. I'm gonna try to pull up some docs and share this with you as well. Because, yeah, user quotas is new. Feel free to use it because it gives you actually pretty good control of not having overspends in the

Rachel Roundy: functions. Awesome. Okay. Next question. How are video inputs chunked? Are they chunked by frame, by scene, or by fixed time interval? And is that configurable?

Hanna Mykula: That's completely configurable. So you can choose if you wanna segment based on clips, and then you can give the the length of the clip that you wanna have, the minimal length. I think the default is four, but that can be any value, I think, from from two to five. But that's completely configurable.

Rachel Roundy: Great. Okay. Next question is, which embedding models are supported for video and audio? Is there a model choice, or do we just have a single Snowflake option?

Hanna Mykula: For now, we support for the video embeddings the twelve flaps Marengo model. That's currently the only model that we use for embeddings.

Rachel Roundy: All right. Okay. And last question for you, Hanna Mykula. Can multi embed outputs be stored directly in a vector store within Snowflake? And does Cortex Search support cross modal retrieval natively?

Hanna Mykula: Yes. Yes. So that's exactly the example that I that I showed you with the multi index and Cortex search. You can actually place your embeddings that's output of the AI multi embed in the Cortex search. Amazing. That's actually our way to to do that. Oh,

Rachel Roundy: great. Awesome. Well, thank you so much, Hanna Mykula, for joining us to talk about AI functions. We appreciate you being here.

Hanna Mykula: Cool. Thank you so much. See you.

Rachel Roundy: All right. Thank you. And thank you all, everyone, for joining us today for our AI Pulse. We had a ton of updates. Thank you all for participating so much in the Q and A. Anything that was not answered live, we will answer offline. And just to, again, reiterate, the recording of this will be available immediately after we stop this live event in the same link and platform. All of the related materials are also linked there in the platform. So please make sure to check those out, including the press releases today about our open weight model and AI gateway announcements. Again, this is a monthly series. We will always bring you the latest and greatest from stuff like AI machine learning. And thank you again for being here. Enjoy the rest of your day. And we'll see you next time on the polls.