E156 - Transforming Enterprise AI in the Data Cloud - Mats Stellwall

Join us for Episode #156 of the AIAW Podcast, where Snowflake EMEA’s Principal Architect AI/ML and Field CTO, Mats Stellwall, walks us from his SPSS beginnings in the 1990s to today’s cutting‑edge Generative AI, revealing why Snowflake’s Data Cloud is the premier platform for scaling AI/ML workloads. We’ll unpack game‑changing tools like Agentic AI, Cortex Analyst, and Cortex Search, dive into real‑world enterprise use cases, navigate regulatory and ethical challenges, and peer into the future of AI—from frontier vs. task‑specific models to the promise (and perils) of AGI—so you can stay ahead in the AI revolution. 

Follow us on youtube: https://www.youtube.com/@aiawpodcast

Chapters

• 0:00 Vibe Coding for Data Generation
• 7:15 Snowflake's AI Journey and Capabilities
• 17:07 Cortex Search and Document AI Functionality
• 29:39 Evolution of Snowflake Over Five Years
• 35:55 Enterprise AI Needs and Value Creation
• 45:38 EU AI Act and Global Regulations
• 57:47 Future of AI: From Foundation Models to AGI
• 1:05:29 Quantum Computing and Intelligence Tradeoffs
• 1:48:33 AGI and the Future of Humanity

Transcript

Mats Stellwall: 0:00

Keep it, not like now. I'm going to go there and pitch Snowflake, unless there's a pitch Snowflake session, which the technology in preview and so on is. But when you do the keynote, you have to talk about something.

Anders Arpteg: 0:12

Okay good, but getting back to what you said about it, you actually did some vibe coding or what was. You mentioned something about. You know you need to create some demo. What was that?

Mats Stellwall: 0:27

yeah, absolutely so. So I finally and I'm probably the last one on the ball when it comes to these type of things so I finally tried to use and in this case, use yemeni, uh, um to produce code for me, so build it. I was, or are, because I'm not done yet, but I'm building a demo for something called Wealth Management Customer 360. And part of that I need to generate data, because we need synthetic data because there's no customers that are going to provide the data to me. So in order to that, I need to generate some nice code that can actually create this data and I can rerun it and I can decide how many customers I'm going to have and then my new advisors and so on. I gave it a try. I had a colleague that posted in internal Slack to say, well, hey, if you're going to generate code, here are some tricks you can apply to the.

Anders Arpteg: 1:22

LLM. You need to generate data, but to generate data, you need to generate data, but to generate data you need to generate some code and then you use vibe coding to generate the code to generate the data.

Mats Stellwall: 1:30

Yeah, which I mean it's. I mean, in the end it's when it comes to generating data, it's it's trying to figure out how to make it as as realistic as possible, but also, how do I keep things together? So if I, if I generate uh, customers out of the blue, I also need to tie them to something, and and so on and so forth. So so the whole experience was quite fascinating in the sense of how we can resonate and give me this is the way how this is going to be approached. This is the structure I'm going to create.

Mats Stellwall: 2:04

Let's start with step one, generating this, and then we iterate through this. I test it, I say, well, this is great, but we need to change X, y and Z. We get through the whole code and then after a while, I realize, well, I want to do it a little bit differently. So I give it instructions and do that, and I don't need to be very precise, which I found. It was pretty much like talking with someone next to you to say well, it understands it rather well, right, it does.

Anders Arpteg: 2:30

It can even write some prompts. That is very well or not well written in some sense yeah, absolutely.

Mats Stellwall: 2:36

But but also I mean because I got paranoid after a while. I said can you give me, can you provide me with a prompt that I can use with you later on if things go?

Anders Arpteg: 2:45

bad. Oh, so you generated a prompt to use as a prompt.

Mats Stellwall: 2:49

So, like a history restart, to say, well, if we need to restart this, can you summarize what we have done and give enough information so we can restart? And it could do that. And I found it also fascinating because what I did also, once I had my first version of the data, I wanted it to generate a user interface for me in Python, and then I suddenly started jumping back and forth between, well, oh, I need to change something in the data, so let's go back to the data and do that, and it could keep track of it. Now, having said that, I'm doing things where I feel like now I've reached the limit of the number of tokens this can handle, because it starts forgetting things.

Anders Arpteg: 3:32

It's not Okay. So you had a number of iterations. You generated code to generate the data, but then you also generated code for the user interface. So then you went back and said please change how the code is generated, or for how the data is generated.

Mats Stellwall: 3:47

Yeah, and and then it can update the code to to follow that, because I could just say, well, now I want to have a user interface against this data that we generated. These are the different pages I want. This is what I want these pages to contain and you should use. In this case, we're using streamlit in snowflakes. I specified a couple of things and they said right, and it built it for me. And then, when we went back and and changed the data, I said, well, now I need to update these pages to reflect this and it just took the changes what kind of data was you, were you trying to to generate?

Mats Stellwall: 4:18

so the, the. So I generate um, um, so wealth management. So there's advisory data, it's customer, customer data, it's portfolio data, which is the most fascinating not fascinating, but the interesting part it's also interaction data. So I have it to generate prompts in order to ask another LLM to generate unstructured content. So call transcripts, emails, complaints, and make them as realistic as possible.

Anders Arpteg: 4:44

So it's not just the time series stock price candidata no, no, no.

Mats Stellwall: 4:47

It's interaction, so it also created logic to create interactions and generate it based on the customer, the wealth it had. And then it said well, now it's a monthly catch-up and it's done through a call, so generate a transcript. Well, a transcribed call, that's what I was asking for. So so it could also generate instructions for using other LLMs in order to produce this unstructured data that I can store and then use later on.

Anders Arpteg: 5:20

It's a lot of meta meta here.

Mats Stellwall: 5:22

It is which is fascinating me endlessly and it's, it's, it's after a while. It's a lot of meta meta here it is which is fascinating me, uh, endlessly, and it's, it's, it's after a while. It's also kind of a little bit scary how I interact with it. It's in in my mind.

Mats Stellwall: 5:34

It becomes like I'm talking with a colleague it is right yeah and and it's kind of scary in a way how how it responds and talks with me in that sense and you use gemini, right, or yeah, we're using so 2.5, which 2.5, uh, advanced. So here's the thing with with like, probably in in the organization, another organization in snowflake, we have kind of policies around what we can use for what, so, so, um, so gemini or or google gemini is one of the things we can actually use for producing things that we use in work, so otherwise I might have used something else.

Anders Arpteg: 6:11

Cool. I'm a super cool stuff and you were positively surprised.

Mats Stellwall: 6:17

Yeah, I was positively surprised. So I was a skeptic before. But in the end what I also realized is you need to understand what it's producing. So I mean, if I didn't understand Python code and the libraries it's using, it would be probably harder, because otherwise I cannot figure out if it's doing things wrong, if it's using the wrong parameters or the wrong libraries, and so on.

Anders Arpteg: 6:42

You're a technical person and I'm shortly going to have you also describe yourself a bit more, but I think also the general idea of how we can now for non-technical people not like yourself, but other people that are more business-oriented, that don't know how to generate data or code potentially in the future can actually do that surprisingly easily by speaking to AI and then having it generate both data and code in different ways and then find insights and build applications. I think that's a very beautiful future that will be exciting right.

Mats Stellwall: 7:19

Yeah, absolutely yeah, I agree, but I think I mean the thing we are still missing is that we still need a human between, because we need to verify. I think that is the next level to train new models that can actually handle that interactions better.

Anders Arpteg: 7:36

But yeah, In a couple of years, perhaps, who knows? With that you know. Very welcome here, mats Estelval. You're a principal architect right of AI and machine learning at Snowflake and please describe a bit yourself. You know what's your background. How did you come into the field of AI and the role that you have in Snowflake? Who is really Mats Estelleval?

Mats Stellwall: 8:01

Oh, that is a good question. So I mean, I've been around for a while so. So in my background is mechanical engineering, so so when I graduated in 91, that the market wasn't that great in the industry, so so there was nothing, there was no going from when I started my training when they say, well, you're going to have a job immediately after you leave to when I was leaving. There was nothing there. So I tried to figure out what I want to do. So it took me a couple of years and then I ended up in what we call ADB training, so it's a semester of just data science. And I thought that was fun because I've been playing around with computers before.

Mats Stellwall: 8:49

I ended up to do an internship as a very small consulting firm, so I was building client server applications. I found myself building more and more applications that was data centric, so building a lot of reports and so on and so forth. So I, I, I moved away to another company that was doing BI and data warehousing because I thought that was interesting. So I did that for a lot of years as a consultant and then I, I, I mean after a while I ended up in, in, in, up in Cognos for doing BI reporting. Then I ended up in IBM for doing some more consulting and part of that journey IBM acquired SPSS and my wife was working for SPSS at that point of time, so you met your wife at IBM.

Mats Stellwall: 9:42

No, I did not. I mean, if we're going to go that route I met her through. If you're old enough, you might remember something called the Lunar Storm, of course. Yeah, so that's how we met.

Anders Arpteg: 9:55

Okay, cool, I don't know it's the Tinder of the old days.

Mats Stellwall: 10:00

No, I wouldn't say, but not really. But that's how we met, but nevertheless she was working there. I found the predictive analytics quite interesting. So when IBM acquired SPSS, it's opened up some positions there and I was kind of tired to do the consulting thing. So I just took the chance and managed to get the role there and from there it really fascinated me. So starting from that, I've been working with data science or predictive analytics and machine learning in one way or another in various roles.

Mats Stellwall: 10:38

And then I mean, the way I ended up in Snowflake wasn't planned in any way. So I was working at Oracle at the time having a similar role that I have now, so kind of a data science SME, and I had gotten a new job at a company called Signavio to work with process intelligence, so process mining, which I found still find very interesting. But COVID hit, so I was let go and I've, after six weeks I find myself like, okay, I, there's some challenges ahead now I don't have work. And at the same time my wife had applied for a new job, so she had also put in her notice and the new job was canceled, so so she was also out of job. So I thought what the heck, let's just get something out there.

Mats Stellwall: 11:31

And Snowflake was very fast to reach out and my thinking was I've done data warehousing before I know it. Why not go back to basic and do that for a couple of years? And they were very open with me when I was talking with them and said we know, you have a data science, machine learning background. We don't have anything. We find it very interesting, but we don't have anything, so don't expect to work with it. And I said, fine, I can do this. And so I joined Snowflake. And when I joined I realized, well, there's not many people that is pushing data science, but Snowflake did have some value points around that. So I just grabbed that area for EMEA and became the data science guy in Snowflake.

Anders Arpteg: 12:10

Cool, before we go too far into your current role in Snowflake, perhaps people would like to know what is really Snowflake. For people that don't know it, how would you describe the company?

Mats Stellwall: 12:21

Oh, I mean, yeah, that's a tough one, but but I mean we are. We are a product and and and only thing, and we call ourselves the AI data cloud and that sounds very fancy, but in the end, we are a data platform for the cloud and we have a lot of fancy functionality around that, but, but I mean, the essence of what we do is managing data in a very, very, very clever way, uh, and we are very good at it I mean in sense, if you were to compare it to traditional relational databases, or how you can scale it up, how we can move to the cloud, I guess you have, you know, one of the strengths.

Mats Stellwall: 13:02

If I were to say it is, I mean, the scaling aspects of it, right, yeah, and and I think also, I mean, if I go so let me put it like this way so when I join snowflake, I you know part of the process you read through the website, you go through all the material and and I mean I'm coming from american companies, so working for org and ibm and everything is great by marketing, everything works. Uh, when you read the marketing stuff and then when you use it is it might not work as well as you think, yeah. So so I had the same thinking about snowflake when I joined. It's like okay, easy to use, it's okay, everything worked, okay, fine.

Mats Stellwall: 13:39

And I, when I started, I got myself an account and I was starting to do some, some, some training and I realized quickly it's like holy crap, it actually does what it says. It do deliver on on on on the marketing promise, which was very for me, it was a was, was, a hug thing and and these guys know what they're doing so. So I would say that that that is the thing, and these guys know what they're doing. So I would say that that is the thing. I mean it's a very well-architected product that actually does what it says it's going to do and does it very well, and so for me that is the fascinating part of being at Snowflake. I mean, I was very skeptical when I started.

Anders Arpteg: 14:24

I mean I, like you, came from a more data science and AI point of view and I also been looking into like Databricks and different like Spark solutions and big data tooling like that, and my view was a bit like Snowflake was more on the data warehousing side and then other solutions was more on the data science side, but I must say that Snowflake has really picked up on the AI side in recent years. Would you agree on that?

Mats Stellwall: 14:51

Oh, yeah, definitely. As I said, when I joined 2020, the most data science thing we had was a Python connector, which basically was a Python library that you could send SQL through Python and get back a Pandas data frame, and the whole you know us, both of us was we can manage the data very well and you have the marketplace and everything, and then we fast forward now where we have, you know we can. You can do model training, you can run containers. You can do a model registry. You can run a number of LLMs Well, we run a number of LLM foundation models that you can do a model registry. You can run a number of llms well, you, we run a number of foundation models that you can use, and so on and so forth.

Mats Stellwall: 15:29

So the last 16 months has been crazy. I I mean even I as a working there has a hard time to keeping up, and sometimes customers say so oh, you have this function, we do we, and, and then you go look at it oh, we have yeah, it's tough to just keep up with the general landscape of AI, but then in Snowflake it's most fast.

Anders Arpteg: 15:49

But would it be fair to say and sorry to bring up a competitor of yours but if we take Databricks, they were more in data science but really bad in data warehousing and working with structured data and then they're trying to move in that direction. You were really good in structured data and data warehousing and now moving into AI.

Mats Stellwall: 16:09

And would that be a fair kind of yeah you could say I mean, yeah, I mean there's no secret in that. I mean we came from, yeah, we come. Well, basically, we were focusing initially on data warehouse and semi-structured data. So I think the ambition was when they started was that they saw that you have relational databases is really good at relational data, and then you have Hadoop, which is very good at unstructured data, but once you're trying to combine it, it's very hard. So that was what they were focusing on. And, of course, for analytical and and having spark and and data bricks coming from from a data engineering, data science background where they try to solve these type of things, and now moving and and I mean, if you put us on top of each other, you will see we overlap 98 and then it's just different flavors of things I actually attended the last year's what's called snowflakeake World Tour or something.

Mats Stellwall: 17:05

Yeah.

Anders Arpteg: 17:06

I guess the theme then was like enterprise AI, was it?

Mats Stellwall: 17:09

Yeah.

Anders Arpteg: 17:11

It's fun to see how much you and I actually was not very knowledgeable, I must say, about Snowflake and their AI offerings, but now, when I actually do have a bit more experience with it, it's's surprising to see how actually it is yeah, it's a.

Mats Stellwall: 17:26

Sometimes it's a. It's kind of a. It's not a well-kept secret, but it feels like it sometimes and and I think we have been better to to to have a market message that that resonates a little bit better. I think it's also the the, our uh sridhar, our new well new. He's been for over a year as a CEO, has changed a lot of things. He's very active as well on the market and talking with people around. This has helped us a lot in this.

Anders Arpteg: 17:54

If you were to speak a bit more about your role at Snowflake as a principal architect. Is it specifically for AI then, or how would you describe your role?

Mats Stellwall: 18:03

Yeah, so my role, yeah, so I'm focusing on AI, ml, and actually I'm also focusing on specific on financial services. So the whole purpose I usually try to describe it as a customer advisory type of role so my job is to talk with customers and try to explain to them and discuss with them is how Snowflake fits into all this when they want to do AI in ML, where we are a good fit, where we're not a good fit, and so on and so forth. So it's not so much in, and I think using architect sometime is a kind of a misleading. I don't draw that many diagrams. It's quite easy with Snowflake. It's just one box and then you're done. Is it really Okay? Well, it depends on how much you want to do, but it's the same thing.

Mats Stellwall: 18:52

If you think of it like, how do you draw an architected diagram of Salesforce? It's just a box that does it. It's a soft solution. So the inner workings is not that interesting, to be honest, but we sometimes tend to do it a little bit more complex. Just to look, there's tables here, there's this here, there's this here, but you have all the Cortex services and whatnot. Yeah, but it still is functions. I mean, it's the same thing if you take a, you don't take a relational database, and then you start drawing like, okay, so in this Oracle installation I have, well, you can draw a database, and then I have this select function. I have this function and this function. This is just like we just try to make it more interesting. But I usually said just draw one box and say data in and data out.

Anders Arpteg: 19:36

It's funny you say it like that. I remember those Spotify days. We actually tried to say what is really the vision that we want to have for a music service and in reality it should be just a big button saying play. Yeah, that's the only thing it should be, and it should understand what you want to do. You shouldn't have to care about. You know how the songs are organized or what you want to listen to and whatever. It should just understand you. In some sense it sounds like you're.

Mats Stellwall: 20:01

I think. I mean, yeah, I think we internally, we talk about when we do it. We usually talk about the easy button, so making it as easy. I use it, as maybe I shouldn't say this, but I use it sometimes as an argument when I talk with development. I say, when they do something and I say, well, this is not the easy button.

Anders Arpteg: 20:20

Nice to hear so. Okay, so you work a bit more in finance services, and, uh, and as an architect though, or am I correct in saying you're trying to understand how they could implement it?

Mats Stellwall: 20:38

yeah, well, sort of. I mean, a lot of it is now is in in. A lot of discussion is like they can come to us and say well, we want to do this use case. So, so let's say we have, we have our portfolio data in Snowflake. We want to talk to that portfolio data, we want to do Texas SQL. How can we do that? What? How do we start? What is the you know advantages, what, what can you offer us? So there's a lot of these type of discussion and then there's other type of type of discussions, which is more we want to do. Well, we, we are doing ai. How can we use snowflake in this context? What do you provide and how do you provide it? So, so I mean, I'm I'm still part of the sales side, so a lot of it is positioning and you know.

Anders Arpteg: 21:27

But if you were to go a bit more technical, then and Snowflake, of course, is known for their data warehousing and cloud capabilities, that they do have. But what kind of AI offerings would you say Snowflake do have? Oh, start easy.

Mats Stellwall: 21:58

Is it just a box or is it something else? Like any other foundation models, you choose a model, you provide a prompt, you get an answer and then you can use them in various ways as part of a SQL state band, so you can run it over your table and so on, and then we have functionality to certain of these models. You can also fine-tune them in Snowflake. So taking Mistral Large 2 and then I fine-tune it on my own data and that model stays that. So that is the foundation.

Anders Arpteg: 22:26

So you manage the fine tuning as well, then.

Mats Stellwall: 22:28

Yeah, so we provide it as what we want to call a serverless feature. So basically, you just point to, this is my source data and this is the model I want to fine tune and this is the model name I want to have on the new model. So we provide that type of functionality. And then we also provide more task-specific functions, so LLM-based functions for doing things like summarize something. So instead of me writing a prompt for an LLM to summarize a text, I can just say summarize this. We have a function for sentiment scoring and now we also have a function for entity sentiment, so we can also provide which topics you want to get sentiments for. And then there's translate functions and a couple of others we can use for this. And then we have a more of a. You could call it high level services. So we also provide functionality around text to sql, which we call cortex analyst, which is a in. There is a rest service that you can use from outside Snowflake to talk with data you have in Snowflake based on a semantic model you have.

Mats Stellwall: 23:32

We have Cortex Search functions. That gives you the possibility to do content retrieval based on unstructured data, especially a RAG solution right, especially a RAG solution or enterprise search solution as well. So think of it. So Cortex search is quite an interesting aspect because it's a combination of similarities but also semantic search. So think of it as similarity search combined with Google search functionality. So the keyword ranking and these types of things are supplied as well. So the keyword ranking and these type of things is applied as well. But the main goal is for ag solutions. But we have customers that start using it for enterprise search type of solutions as well, and then we have on top of that, we now have what I use, cortex Analyst to retrieve structured data. Or should I use Cortex Search to retrieve unstructured data? In order to answer this question, Awesome.

Anders Arpteg: 24:32

I'd like to break down each one of these to describe a bit more, and we actually had a session last week speaking specifically about agentic kind of AI.

Mats Stellwall: 24:41

Ah nice.

Anders Arpteg: 24:42

So perhaps I'm thinking about the order here, but now let's go agent to cortex. Agent's solution and I actually liked the way you describe it and you said a bit about that it can choose a bit. You know what kind of other tooling if you call it that that it should use. How would you describe or define what? What is really a more agentic kind of AI solutions?

Mats Stellwall: 25:10

I think that is a very wide question. We have so many things of it, but I think I mean, for me, what I work with day to day is kind of limited. It's just data retreat, I mean around data agents. So we don't have today anything that can do anything above that. But I think it's the agenting part which I find interesting is or how I see it, is the possibilities to have multiple tools that do different things.

Mats Stellwall: 25:38

So select data, look up something on the website, do something, look up something on the website, do something, do X or Y and having it to select between these which tools are needed in order to answer the questions I provide. And, to be honest, a lot of this is kind of I don't know, I don't want to use a scary word, but it's kind of a worrying thing in this that we have an LLM to choose and the choosing is not always so you logical or I can see the log. I like the new models. That has more reasoning because they can also provide you. This is how I was thinking. This is why I'm choosing this and so on, because I think that is an important aspect of this to understand. Okay, why did you pick this tool for this, or why did you submit if you allow it to do something? So why did you submit to this to x, y and z? I need to have the reasoning behind this.

Anders Arpteg: 26:33

So yeah, okay, so cool. Good, I like it, would you? I'd love you know these kind of super simplistic kind of definitions of things, and me and henrik sometimes argue a lot about this. But would you agree with a super simplistic three-word description of a genetic AI as the ability to choose action? If you take a normal chat, ept, you know it, just yeah, yeah.

Mats Stellwall: 27:01

Yeah, it kind of, but I mean it's yeah maybe. I mean in the end it has to choose between multiple things and then choose one or more of those and use those.

Anders Arpteg: 27:20

Because otherwise you could manually program basically a set of AI tools to run in some kind of sequence.

Mats Stellwall: 27:28

Yeah, and I mean the poor man's agentic framework.

Anders Arpteg: 27:32

Yes, another definition that we talked about, which is a bit harder, is to say that agentic AI is when it choose when to stop, meaning that it continues to run a number of steps until it comes to some kind of termination decision saying now I'm done and here is the result would. Would you prefer that, or do you think it's sufficient to say that when an AI system can choose what tool to use, what action to take, then it would be agentic?

Mats Stellwall: 28:09

Yeah, maybe I think I mean it's so because it can be both and it is both. And I think that is the challenge in this, because from a surface point, when you look into it, it looks very simple. I mean, I define the number of tools and then choose. But then when you start thinking of it, it's like but there's a complex process that is ongoing here. Yeah, and how do I explain that? Because I mean, if I program, there's logic that I can follow and I can step through and say, well, when this condition condition is met, this is going to happen.

Mats Stellwall: 28:43

But here it's kind of a yeah, you need to have a end goal to say, well, this is the end state. If that you need to achieve, otherwise, continue until you do so. So I, I think I like both of them, because because sometimes it's just enough to have the first one, because you don't need to know that much more, because that might be just confusing things. Uh, because in the I mean, I think we are moving also into an, into an aspect where we well, maybe not, but, but, but I but I hope so where we need to, the need to describe this is is is less needed, because we do accept how it works Because it's interesting.

Mats Stellwall: 29:22

You sort of trust it a bit more. Trust it. I mean, I don't need to explain how a relational database works because I have no clue in some cases. I know the basics but I don't know exactly. But I still use it every day and I don't. I mean, nobody comes to me and say you know, explain exactly how well actually they do sometimes, how snowflake works when it selects data, but but usually they don't, because they they, they accept it, it's it has, I guess it's like driving a car.

Anders Arpteg: 29:49

I mean, yeah, you don't need to explain all the details of an engine to be able to drive a car in some sense no, thank god you don't have to, otherwise I would never get my driver's license. Awesome but let's go through the other parts, both the Cortex Search and Cortex Analyst, and if we start a bit more with the Cortex Analyst, just describe briefly what does it do?

Mats Stellwall: 30:12

It does text to SQL with some add-ons, but the primary goal for it is taking a natural I mean a question I'm asking and turn it into Snowflake optimized SQL that I can execute.

Anders Arpteg: 30:29

Actually, to me it's a rather big thing Because in general, we want to have more and more people being able to find insights from data, more and more people being able to find insights from data. Now, the problem with that is that you normally are required to know either a lot about Python to be able to process the data and do all the ELT kind of processing, then building the pipelines and bringing the data models, and then in the end you need to probably do some SQL to be able to extract the data, and then you need to know how to visualize in different ways. But I guess if we can remove that constraint and have more people, at least for basic use cases, to just talk to an AI, either by writing or talking, to get the data you want, that would be a big thing.

Mats Stellwall: 31:18

Yeah, I agree In a way. I think so. I mean, in essence it's a I'm what do you say? I think it's a good thing, I think this is a really interesting product I mean not just Snowflakes, but the whole talk to your data type of aspect. But I also think that we might hype it a little bit too much now. And and if you think of it like, if you remember when self-service bi came, so when tableau step into the door, everybody, you know we should throw out the dashboards because now everybody can analyze their own data, they can ask the, they can just create whatever they want, and we did that a lot.

Mats Stellwall: 32:04

And then people started also realize, like, well, I don't want to start every day with a blank canvas, I actually want to have a starting point and then I want to analyze the data. So what I'm after is I think this is a really good tool in our toolbox. But I think when it is really getting interesting is I can get my initial data to say, well, here's the state of the world for me. And then I can say, well, based on this, tell me what's matter or what should I do about things, and I think we are heading. I mean we are not there yet. There's a while to get there, but I think this is the end game. I mean, sometimes I say I don't want to talk to the data, I want the data to talk to me. So I want to be notified, to say you know what I don't know? Sales has changed since yesterday and here's the reason why, or here's the key things that has happened in this data. Here's five recommendations that I I, that you need to do in order to change the outcome of this, because then it's getting really interesting, because then I get also the uh, what has happened? But also what I need to do.

Mats Stellwall: 33:14

I mean, sometimes, when you I mean it might have been changing, but I mean my, my experience has been at least when I did a lot of bi and data warehouse when you talk with c-level people and and they say I don't have time to read the freaking dashboard, I just want to have a green or a red light. If it's a green light, I don't have to do anything and if it's red, I need to call someone and and do do something. And I think we, we the end goal of this is kind of a similar thing If you start thinking of it like I. Just I want to get the insights pushed to me in some way or another, but I do think Texas SQL is quite interesting because it's it, it's, it's, it's the self service analytics next level, because the challenge we have with these other BI tools is you actually need to know how you can combine your data.

Mats Stellwall: 34:02

So you need to know, yes, what to drag to where and and how to connect things. What, what text to sql give you is well, we still need to have some kind of a semantic model behind the scenes. But, given the semantic models, what I need to know is just what I want to ask about, and then it can generate the correct sql for me. So, so I I think it is a step up on that ladder.

Anders Arpteg: 34:25

Let me disagree a bit here. I like what you said a bit about you know, in the end you just want the green and red lights. I certainly agree. But then if you get the red light, you won't be able to analyze it somehow, right, and then what do you do? I mean, then you need to be able to talk to the data in some way.

Mats Stellwall: 34:43

Yeah, but you have a starting point right, okay.

Anders Arpteg: 34:45

So starting point is data speaks to you, but then in the end you want to start speaking to the data to understand yeah, because then I can investigate.

Mats Stellwall: 34:52

I mean, I think also and I mean no offense but while people like you and me, we are analytical people, so so we like to, you know, dig into. So for us this is a very natural thing and it talks to us. Pun intended, look at this this angle, and we can, we can take this constraint. Many people are not. They just want to have here's the state and then they can start asking questions about it. But I think I think it I I'm not saying it has potential and I don't say it it's an important feature. I I think it's a very important feature. I think we might overplay it's significant in the state it is today.

Mats Stellwall: 35:45

I think it's I mean, if I'm plugging my own company or Snowflake in this, what I find interesting now we started with text-to-SQL and what we have added now is I can also say, based on this SQL and based on the executed data, it can also provide me an answer. So okay, so based on these numbers, this is some conclusions, and then it can now send me also a nice visualization of this data. And now it's starting to get interesting, because then it not just only give me a SQL, it doesn't just give me numbers in a table. It also gives me some kind of summarization of these numbers. So I think we are starting to getting there that it gives us, but it might be just me that wants to have that type of features in that I think there are.

Anders Arpteg: 36:36

even for experts in data engineering, it can be really hard to find insights from data and if you have thousands of data models and tables and data march and you then want to answer a question you get. It's hard to just know where to get started, absolutely.

Mats Stellwall: 36:58

And that is valid. Actually, we have a customer in uk and what they're using it for is basically for their data engineering to understand where to get the data for so the. So the business user comes to say, well, I need you to create this data product for me, or this, I need this data set, and they it's like, okay, I have no clue where to find it. So they use this functionality to ask the question used to get the sequel, because then they can see okay, is these five tables I need to use in order to build this pipeline.

Mats Stellwall: 37:29

So so they are actually using it for that purpose, uh, as a developer, uh helper tool in order to figure out in which table and columns do. Does this information actually exist?

Anders Arpteg: 37:40

so so yeah, so there are a lot of challenges, so to speak, with the current way of working I would say, with normal data engineering or analytics engineering in Snowflake that I think AI can help with right.

Mats Stellwall: 37:55

Yeah, definitely, I think I mean to be honest. I mean, we've been focusing a lot on the end-user side. I hope the next focus will be around this type how I can generate pipelines based on. If I describe what I want to do as saying I'm going to take data that comes into table A, I want it to do X, y, z, and then I want this and this and this. Please generate the code needed for that, and then we can start implementing. We don't have it. I mean, you can use other tools for it, but I would see that it's next evolution and I appreciate you saying that.

Anders Arpteg: 38:29

Okay, we, we have a certain functionality today. It's not perfect, doesn't do everything, it's a starting point and we shouldn't overhype. You know what the current functionality does, but I still have this kind of vision that I wish there will be a point when talking to data will be a reality. Yeah, yeah, yeah, right, and and I think you know if people can do that without knowing all you know the thousands of tables you have and how the fields are described and how to join the tables and how to write the sqs and how to visualize that in some kind of report and imagining a future where you simply say I now want to understand how the sales have progressed in this country over these products over this time period, and please visualize it. I mean that would be an amazing future, wouldn't it?

Mats Stellwall: 39:17

Yeah, absolutely, yeah, absolutely. I think, I think so. I mean some of it we can. I mean we can build today. I mean I don't know why, maybe we have I haven't talked to all customers but I mean what I would see is what you can, what I would like you know. You can do that in the multiple steps, like first you go and describe this is what I'm interesting to knowing every day. So please send me an email, an SMS or a text or a chat or whatever with this information which I can discuss further with you about in that sense, and we can definitely build that today. And then I think that is where you get this nice future, because you do want to have a starting point, right? You don't want to ask every day what was the sales previous week divided by X, y and Z? You want to have that, but then you want to say, okay, these numbers looks a little bit interesting, so why don't you drill down on this?

Anders Arpteg: 40:13

So you have the normal dashboard? Yeah, when you see some kind of change, you won't understand why, right?

Mats Stellwall: 40:17

Yeah, absolutely, and I mean a golden future will give us some proactivity in that to say, hey, this looks strange, you should do X Y, z because of Z, y and these, these, these things. That has happened before.

Anders Arpteg: 40:32

I think that's you know. I know in Google Cloud, if you look in their security kind of functionality, they basically give you a text description of what's happened in recent weeks.

Mats Stellwall: 40:44

Yeah.

Anders Arpteg: 40:45

Saying okay, normally you have this kind of traffic, you have this many attacks and this kind of security vulnerabilities, and in last week you see an increase here and decrease there. So basically give you a summary of what's happened in recent times. Wouldn't that be nice to have from data in general in some sense?

Mats Stellwall: 41:02

Yeah, I mean yeah, absolutely, I agree. I mean, you can basically build that with existing functionality, but of course, if you could have it in a package solution or packet functionality, it would even be better. I think we I mean we're probably definitely heading in that direction with a lot of it.

Anders Arpteg: 41:21

And if we were to go to you know what is the functionality? I recognize of course Cortex Analyst is still in early stage and there's a lot of challenges still with it, but could you go? You know, for people that are a bit more technically interested, how does Cortex Analyst work today?

Mats Stellwall: 41:39

It is. I mean, in the end it's a very, not very simple, but from a user endpoint it's very simple. So what Cortex Analyst is requiring is that you create a semantic model, and today that is a YAML file. So think of it like I need to describe the tables I want to ask questions of. So what I need to tell about this table is, of course, I need to give them a description. What does this table contain? Which columns do I have? What does they contain? Is there any synonyms for the names? If we call it, what is the some examples, what is the relationships I have, and so on and so forth.

Mats Stellwall: 42:15

The nice thing is we do provide functionality. They can generate this for you and we're using LLMs to create table descriptions, columns descriptions, automatically for you. So it needs that and basically, based on that, I mean behind the scenes, there's an agentic application that does a lot of things. It uses your questions. I mean it checks for what is the intent of these questions? Do we understand it? If we don't understand what the user is asking for, we ask for clarification and send it back. Once we know that, we go and see can we generate a SQL based on this question. We know what it's asking about, we know how to go forward. We enter it to SQL and if it's a correct SQL, we send it back. So so there's, and I think it it's. It is very simplistic in the sense of for me to use. The only thing I need to focus us is to create my semantic model. Um so, and I mean, in the best of worlds, I already have descriptions of my tables, but if you don't, we can generate them for you based on the data.

Anders Arpteg: 43:23

Can you do the Jamf file automatically in some way?

Mats Stellwall: 43:27

Today there's a tool in Snowflake, so it's called well, it's called Cortex Analyst. You go in, you choose your database table and then you choose your tables you want to have, and then it generates the first version of the files and then you need to set up the next generation. That is soon there. You can create what we call semantic views instead, so you can describe it as a schema object, so doing more of a semantic modeling of the data, and then we will utilize that instead. But there's a manual step around this still.

Mats Stellwall: 44:03

But once you have that, then it's more of a, it's a REST API. So for every call, you provide your semantic file and then what is the question I want to ask from that? And then you get the SQL back and then it's up for you to execute that SQL. We are working on features that are actually going to allow you to automatically execute the SQL on our side directly. But now you get the SQL back and, based on that, you can execute it if you provide us with what we call a query ID, so you need to identify for that execution. We can also provide you a summarization of that data as an answer back as well.

Anders Arpteg: 44:46

So in some sense you can talk to the data in some sense.

Mats Stellwall: 44:49

Yeah, and it works quite well. And I mean there's a lot of things you might sometimes need to do in tweaking. You could provide custom instructions to say, like what I always have to do, since I have demo data, to say, well, when I ask for today, I actually mean this date, use this date always in the filtering so you can also instruct it how to treat things. Or when we say EU, we mean these countries, when I say North America, I mean these countries, and so on.

Anders Arpteg: 45:18

Nice If we were to move a bit into Cortex Search. Can you just describe briefly what does that do?

Mats Stellwall: 45:25

It's content retrieval based on queries, so from unstructured data. So what it does is it takes your text data and behind the scenes we are indexing it text data, and behind the scenes we are indexing it, we are creating embeddings and then we search on it, so based on similarity search and then also based on semantic search, and then we do re-ranking of the result and return that back to the user, which is also it's a REST API, so you can integrate it with whatever you want to integrate it with.

Anders Arpteg: 46:03

So it's enterprise search right.

Mats Stellwall: 46:06

Enterprise search. So you can say I usually when I talk to customers, so think of it the traditional well, what we call traditional RAG type of similarity search in combination with Google search on the same data which, according to the benchmark we've done and the studies we've done, we get a much, much higher accuracy compared to traditional ways of doing RAG.

Anders Arpteg: 46:30

So you provide a set of files I guess it could be PDF or Word documents or things like that right, and it shunks it up.

Mats Stellwall: 46:37

I guess in some way it doesn't do it automatically today, so you have to do that yourself. So we provide some helper functions. So there's a function called parse document that can read it as OCR reading. So you extract the text or we can do a layout reading. So we also create the markdown for the structure of it. And then we provide today just a recursive split character function for splitting it into chunks and then we feed it in. This Now what we do have if you use SharePoint, we can actually point it to SharePoint and it can automatically handle it itself. And further down the road we actually going to do all the chunking ourselves. So you can just point to where the documents are stored and we ingest them and create the indexes.

Anders Arpteg: 47:28

But let me see, I'm a bit confused here as well. I mean, you do have some kind of document AI functionality as well.

Mats Stellwall: 47:33

Ah, yeah, right we have, which is a slightly different thing. So and I no, I don't play marketing, but but we, we, we, we document ai is a acquisition we did, maybe two years ago, um, from a polish company, and what they do is they are extracting values from unstructured data, so from documents, pdfs or or pictures and so on and so forth, based on questions. But it's not a text extraction in the sense of extracting big paragraphs. Think of it, if you get things like invoices and you want to process these invoices in order to extract who is, who is it from, what is the date, what is the amount, what is the payment terms and these type of things. So what Document AI enables you is I'm just writing.

Mats Stellwall: 48:33

It's a process where you begin by defining the values I want to extract and how to extract them, and the way I extract them is saying what is the company name, what is the invoice date, what is the due date, what is the invoice amount, and based on that, it will identify these values and return them. Once I've defined that what we say, I publish it as a. It became some model in Snowflake, so it's an object I use and then I can use that on my data pipeline. So whenever I get new invoices, I run them through this function and I get extracted data that I can store in the table and use further down the road. Now, the way why it gets confusing is because we put it on the Cortex umbrella as an AI function, but it's more a data engineering, data pipeline function, to be honest.

Anders Arpteg: 49:22

Yeah, because you mentioned for the Cortex search that it does OCR and these kinds of things and it's something that doesn't do OCR.

Mats Stellwall: 49:30

but we have a function that is called parse document. That one does OCR and text extraction, but that function's whole purpose is just to extract everything from a document, where Document AI's purpose is to extract specific values or tables from a document.

Anders Arpteg: 49:47

Okay. So, in short, cortex Search is separate from Document AI, although they have some kind of overlap in the purpose.

Mats Stellwall: 49:56

In a way they might have but it doesn't really have it. It's different use cases for it. So In a way they might have but it doesn't really have it. It's different use cases for it. So even if we can say, well, you can use Document AI to feed a RAG type of application, absolutely. But see it more as I have a document onboarding process. So in finance it's very common. We know your customers and maybe not so much in the Nordics but outside the Nordics. You do have to send in a photocopy of your passport, you need to send in a payslip and these type of things and someone needs to extract the values from those and store it so they can use it in combination to really figure out that. Do I earn as much money? I say I earn to motivate why I move so much money around. And that's where Document AI comes in and shine, because it's very easy and it is very good at extracting values.

Anders Arpteg: 50:46

Let's see how technical we can get here, but I love tech, so if we could go a bit tech First on Cortex Search, you mentioned the RAG and I think RAG is something that a lot of people knows about retrieval, augmented generation and it has some kind of vector database, I guess in your case, and it tries to shank up some kind of documents, even though you have to do it potentially manually in some sense. Yeah, but then you get a set of paragraphs or a set of sentences that you can search through and then when you search for whatever query you have, it will find that through these kind of documents in some sense but you mentioned something about you have it will find that through these kind of documents in some sense. But you mentioned something about you know also Google search kind of functionality. Is that the ranking part?

Mats Stellwall: 51:26

Yeah, it's the ranking and the keyword search, so it combines. It also adds a keyword search or a semantic search as part of this. So it is a little bit language dependent in the sense of handling keywords and these type of things. So, and to be honest, I mean we need to bring in the research people to give you really the insights around that and to be honest, they are very happy to talk with people. But I mean what they saw was when you were searching only using like Cousin similarities or distance, it could be limited in the search space so it could miss some of the nuances between things which the search engine was quite good at capturing. So they saw and had research around. When you combine these two you get better accuracy. The whole Cortex search comes from one of the acquisitions we did, from which I can't remember where we pick our CEO from, so I can't remember the company name I should. But I mean what they were building was an LLM based search engine for them to compete with Google.

Anders Arpteg: 52:41

I recognize this completely. I remember from the old Peltorian days we also built kind of a search function like that that combines semantic search with keyword search, and I think you need both. And if you are unable to do a more syntactical kind of keyword search, people will get annoyed because they're so used to using Google search and just searching for keywords keyword search. People will get annoyed because they're so used to using Google search and they are searching for keywords, and if that doesn't work which semantic search rarely does then they will get annoyed.

Mats Stellwall: 53:08

Yeah, absolutely, and I think I mean we're quite unique with this feature in the market as it is and we do have published a lot of research around. Well, we publish it as a blog post, but it's actually research when they are testing and comparing and benchmarking and these types of things and, to be honest, the Snowflake research people are mind blowing in what they know and what they can do. They are very knowledgeable people.

Anders Arpteg: 53:38

I can imagine If we just go a bit more into documentary as well because, I think you know, for me personally and I think a lot of people, it would be interesting to just understand a bit more. So okay, given a certain document I guess a pdf, an invoice or something how does it actually extract the content from that?

Mats Stellwall: 53:54

it has a. It has an ocr functionality as part of it, so it, so what? And there's a, there's, and there's a nice research paper. It's a couple of years old around this. So what they created is they call the model a tilt model. So yeah, so it's as they explained on me. It's a combination of it's not a full-blown foundation, llm behind the scenes, but it has similarities in it, but it has an OCR part, so it it actually extracts all the texts, it gets it on the make sure it just it, so it gets correct. And then it's been trained on number of languages in order to understand and find values based on a question in that sense. So so it's. They do a lot of these things, uh, uh, out of the box in it directly but it doesn't use like the big you know GPT-4 kind of models.

Mats Stellwall: 54:46

No no, it's something homegrown that they built these Polish people. That was very clever from the university.

Anders Arpteg: 54:53

Oh, okay, okay, cool, awesome Fun. To get down to some technical details here.

Mats Stellwall: 55:00

I will do my best. I mean, some of this is just flying over my head and then, when they start talking to me about it, it's like I don't know what you're talking about. You can stop now, because I'm super confused.

Anders Arpteg: 55:13

Awesome. Perhaps we can go a bit more into more long-term and as the time progresses we get increasingly philosophical and long-term kind of questions. But given that you've been at snowflake for five years yeah, almost five years how does it compare? How would you say snowflake compares today to what it was five years ago?

Mats Stellwall: 55:35

yeah, that's a good question. I I mean, first of all, the scale. Scale of it is insane. I think, if I'm not wrong, when I joined we had about 1,500 customers globally and now I think we crossed 11,000. So just think of the scale of things and the number of people that have been hired. So so I mean when, when I joined, you could I mean you can still do a lot of things, but but it it was very easy to to you could.

Mats Stellwall: 56:12

As I said, I used to grab the data science area because I saw that nobody in me, I was doing anything and I, I, I like it, so that nobody in EMEA was doing anything and I like it. So I just took charge of it and became the guy that they can ask questions about data science and I could answer it and position things and these types of things. So it was very easy to just grab your part and do things. That has not changed that much. It's still possible.

Mats Stellwall: 56:39

But of course the number of areas that are open is getting fewer and fewer because we are covering more and more ground. But I think also, what I like is people are very reachable. I think that is what has kept me still there is that you can approach anyone for anything and they will answer you and be very happy to do that. So so there's an, there's a nice culture in the sense of knowledge sharing and and being helpful and and I mean I wouldn't say I go up to ask people at C-level questions, but I mean I can ask any product manager or anything and they will respond and if they are the wrong person they will point me to the right direction, and so on.

Anders Arpteg: 57:26

Sounds like a good culture.

Mats Stellwall: 57:27

Yeah, it is. How many employees do you have today?

Anders Arpteg: 57:32

I don't know.

Mats Stellwall: 57:34

A couple of thousand, I think I don't really know. It's a number I should know, but I think it's like there's a lot. I mean, every time I and I'm not visiting the Stockholm office enough, but every time I go there I think there's always a lot of people that's like are they new? I haven't seen them before. And it's also uh uh, I mean I'm not visiting there enough because I also get the credit oh, you're new. It's like no, I've been there for five years.

Anders Arpteg: 58:01

But you work mainly remotely.

Mats Stellwall: 58:02

I work mainly remotely. My role is in EMEA role, so I spend maybe 90% of my time with customers outside the Nordics and you travel a lot.

Anders Arpteg: 58:12

I guess it has been a lot which is interesting in these days, but it is what it is, and I I know a lot of startup people and they claim that they want to build, like ai, solutions for whatever kind of enterprise companies, but I think they're missing out a lot of you know what enterprise I really, really want. If you were to elaborate a bit more, perhaps you given that the last year's theme for the Snowflake World Tour was enterprise, ai what do you think? Enterprises?

Mats Stellwall: 58:50

needs the. We have kind of moved away from, or started to move away from, the trying out phase. Or, you know, let's do a POC and see what happens to. There's need to actually be an ROI around this.

Anders Arpteg: 59:11

And.

Mats Stellwall: 59:11

I think I mean you can say whatever you want about Gartner, but I was at the Gartner Synopsis this year and the only thing I took away from there that I found that was interesting is if you're not measuring something, why should you apply AI on it? And that spoke to me a little bit because in this sense, if you think of it and they had a quite good analogy around it it's like if you get 15 minutes back every day, will you do 15 minutes more work or will you just have another cup of coffee? And it's kind of, yeah, it's funny, but when you think of it, like, yeah, that's true, I would probably not work more, and so the gain of this that efficiency is not I'm not going to be more efficient from a company point of view because I will not do more work, I will just have another coffee. And what they were after which I agreed to is you need to measure things in order to make sure that you actually have the impact of AI that you think you're going to have.

Mats Stellwall: 1:00:12

Not saying that efficiency type of use cases are bad. They are good. They will help us and make us more profound in what we're doing. But I think for enterprises and I see that more and more, that they actually have a clear view that if we do this, we need to see a gain in some way. We need to either cost, cost, we need to increase revenue, but we need to have an effect of it, and many of them is now not as afraid as they were, let's say, 12 months ago, that they were missing out. I think we have passed that where everybody was panicking and say we need to do AI and you say, yes, fine, what do you want to do? We don't know. You tell us.

Mats Stellwall: 1:00:55

Okay, that is not a good starting point for talking about what you're going to do with AI Two, that we actually have some use cases that we think is valid for AI and we want to test it by you and see if it actually would work and how Snowflake would fit into that. So I think what I see with enterprises is they are looking for things that gives an ROI, they are passing by the tryout and let's see what happens, type of things. I think also, they realize that it's easy to build something quickly but it's very hard to deploy and run it in production and there's a lot of moving parts in a lot of these solutions. We're talking about Cortex Analyst. As I said, to use Cortex Analyst, you only need to create a YAML file that describes it. That is the only thing you need in order to do text-to-SQL need in order to do text-to-SQL, and I think Uber did a nice blog a couple of months ago about their journey in building a text-to-SQL and I think they did it was talking more about it in a meetup in the Netherlands a week ago or so and they spent seven months in order to get it up to 75% accuracy or something like that, and there was a lot of moving parts in this and I think what companies realize. It is very easy to do a quick POC or looking at the vendor demo and see, well, this is easy when you work with something like, okay, we have two tables, but when you have 15, 20, 50 tables and complex relationships, it's, it's started to get more complex and you don't want to maintain that. So what I see is that they they looking for simpler solutions that is easy to to to get in production, that actually stable and works, because they don't have the manpower to manage this.

Mats Stellwall: 1:02:54

Not everybody has the capacity of Google or these companies. It's the. I think it's not everybody has the capacity of google or these companies. It's the I think so it's the whole thing we had before with, uh, with hadoop and everything. Like everybody was going to do hadoop, they're going to do it themselves. And and and you ask why, well, you know spotify is doing that. It's like, yeah, but spotify has probably 200 more engineers that you have that actually been writing the source code. So of course, they can do it themselves, but do you have that? It's like, no, we don't. Okay, so maybe we should talk about more package solutions that do these things. So I think they see that sure there's drawbacks by having something that is more pre-built than you do everything by yourself, but you get some advantages as you get something going and you can interact, and once you have that in production you get the feedback. Well then, we catch up the next iteration. You get more and more functionality.

Anders Arpteg: 1:03:47

So I love the the agile kind of aspects you're bringing up here. I think also that you know so many companies fail with ai because they think you know, just having a prototype is enough and then putting it in production. How hard can it be.

Mats Stellwall: 1:03:59

Yeah, it is very hard.

Anders Arpteg: 1:04:01

It is very hard. It's probably a hundred times more difficult and more time than just building a prototype. So I love that you're saying that, but I also. You know enterprises have a lot of procedures. They have a lot of rules. They have a lot of policies.

Anders Arpteg: 1:04:17

They need to be compliant, they need to have, you know, security demands being fulfilled and they need to log stuff, and I think you know this is actually something that so many companies miss out on yeah definitely right and I think also this is something where companies need to understand putting things in production is so much more difficult than people think and to be able to comply with all the regulation, with all the security demands and being saying some people should not have access to all the data they could have in a big organization, in enterprise settings. It doesn't work like that and so many are failing on that anyway. Um, I also must say I think you know snowflake has some issues there as well, if I may say so in yeah, we probably have, I mean, but uh, I'm glad to see that you know.

Anders Arpteg: 1:05:13

of course, a lot of enterprises are using snowflake and it's very successful, yeah absolutely.

Mats Stellwall: 1:05:17

But I mean to be very clear. I mean we are not the solution for everything. I think we're one part of this and we fit a lot of use cases. But as I say I mean I usually say I mean we chase data and if data is a critical part of your solution, we have a natural place there. But if the data is not critical in any sense, we might not have and you might build it in some other solution or you use a point solution for it or whatever. I mean there's room for everyone.

Anders Arpteg: 1:05:54

Good, I would like to also move. I know you're going to speak at the Data Innovation Summit in a couple of weeks as well, which I'm looking forward to. What was your topic again?

Mats Stellwall: 1:06:05

talk or it's only me now because my co-presenter had had to drop off but but we're going to talk about, if I remember it right, is is about the coming regulations we have, or the regulations we have like euai act, how it impacts us and and what is the things we can do to handle it. I mean there's there's a couple of parts of it that we can probably manage already today in that sense. So hopefully I will bring some insights and something that people can take advantage of or take to them. I think regulation is we're going to handle it, whatever we want it or not. We can either fight it or we can kind of embrace it and do the best of it.

Mats Stellwall: 1:06:59

I think I think the later, because it's also open up probably some opportunities for us and and in the end it's it's a lot of it is about having control, so, so making sure that you have documented what you do, you know what data is used, how the data is used, uh, who access what, and so on and so forth. So I mean there's a lot of aspects to this and I'm no expert in EU AI Act, to be honest in that sense, but I mean my experience from most regulations is about documentation and proving that we are not doing the wrong things, and and the EU act has it has levels as well, and I mean, depending on what type of application you have, you get more and more requirements.

Anders Arpteg: 1:07:48

I mean, I think this is a core part of also enterprise kind of data and AI usage yeah, absolutely. And I'm very surprised that not more efforts are being placed in how to be compliant, because if more companies had a belief and confidence in that we can be compliant, I think they could use data and AI to a much larger extent than they can.

Mats Stellwall: 1:08:12

Yeah, I agree, and what it might happen, or what I can see in some companies, is that they kind of say, well, since we don't know, we don't dare to do anything, so they kind of freeze, and in some companies it might be the opposite.

Mats Stellwall: 1:08:29

It's like, well, let's see what happens and just go for it, and I think both approaches is not the right ones to do. But I mean, the whole thing here is I don, I don't see it as a as a kind of a new thing. I mean we, as as long as we've been using ML or AI, we have always have this notion well, notion, but the need of we need to explain what we're doing, we need to have control. We I mean we've been talking about responsible, responsible AI for a long time. Before we had all this AI part of it, we had all these worries about. You know, when deep learning and object detections I mean these type of things was very hyped about. You know, let's identify potential criminals based on videos when they're moving around, and you know we have the issues with selecting the right candidate based on HR data. That didn't go that well.

Mats Stellwall: 1:09:24

Yeah, and so on and so forth, and it boils down to you need to know what data I use, how biased is my data, and so on and so forth. So I mean the EU AI act notched up a level, and of course there's a lot of things that is very vague in it, so it's kind of hard to know exactly what it means until someone challenges it. But I think we can go a long way by just having control. I mean lineage and these type of things.

Anders Arpteg: 1:09:58

If you were still to. I know you're not like a lawyer by trade, but still for people that have no idea what the eu ai act is, how would you describe it? I?

Mats Stellwall: 1:10:09

I mean it it regulates or it it sets level. So, depending on how you use AI, for what type of applications, from you know having a very simple AI chatbot that I can use to I don't know generate simple text for something that it doesn't use any personal data, it doesn't do anything. It just gives me a text based on the input, which is hardly there's no regulations around that because it doesn't do anything. It just gives me a text based on the input, which is hardly there's no regulations around that because it doesn't do anything. Up to a level where we have applications that in real time, identify my sentiment or my emotions, or emotions and so on and so forth, which is on the highest level. That is very regulated and requires a lot of things. Even prohibited, right? Yeah, even prohibited, it needs to be in very specific situations requires a lot of things. Even prohibited, right yeah, even prohibited, it needs to be in very specific situations where you can do this. And then there's a number of levels between. If I remember correctly, we have six levels that we need to navigate between and, depending on what you're building, you end up in one of these levels.

Mats Stellwall: 1:11:20

I think one of the challenges we have now is, we don't really know where to put the foundation models we have, in which area there they should end up in, because they are.

Mats Stellwall: 1:11:27

They are very general in what they can do, depending on how I instruct them. So do we say that that, um, they are responsible for making sure that they cannot be used for certain things, that they can be used, or is it the one that is building the application that is using the model that has the responsibility for it? So I think one thing this will open up definitely is more domain-specific models that is trained to do a very specific task, because then you get away from things. If I just want to summarize legal documents, I might just need a model that can do that and it cannot do anything else, and as long as I can prove it cannot do anything else, I should be in the clear for that thing. So it might also open up a new area of domains for companies to create models that are more specific, and it might be easier to do that if we look at what DeepSeek could do with less. If that's really the case, it might open up these type of things that we can do.

Anders Arpteg: 1:12:34

I mean that moves also to a question that I'd love to discuss shortly, which is you know, what are the future going to look like with super general models versus more specific and potentially open source models?

Mats Stellwall: 1:12:53

And yeah, it's interesting to hear your thoughts there. I think we're going to have less and less of these general models because they're getting so big and so on and cover a lot. I think there's the need for more specific models.

Anders Arpteg: 1:13:09

Let's wait a bit with that discussion because that's a big topic.

Mats Stellwall: 1:13:13

It is yeah.

Anders Arpteg: 1:13:14

We have a lot of thoughts about that. But let's get back to the EU Act. And then we have these levels. I think it's four levels, but I'm not sure. Maybe it's four, yeah, but still. Then we have this kind of risk level. We have to in some way assess and say, okay, what kind of level is this application on? And then we have to do.

Anders Arpteg: 1:13:41

If it's a more higher risk, we have to do a lot of compliance work to be compliant. What's your sense of what the impact of the AI Act will be? And just to give you some more context here, I actually heard from a member in the Swedish AI Commission who said something rather controversial which I was surprised but still positively surprised with saying we need to deregulate. And I think we hear more from France as well Macron and others who say as well that if we just keep adding more regulation time and time again, it will cause a lot of problems for europe. What's your thinking about this? Do we need to deregulate?

Mats Stellwall: 1:14:23

yeah, that is a good question, I think, as I mean, everything needs to be on the right level. I think there's challenges with letting these type of things roam completely free without any control over it. I, I, I mean, I, I there's potentials in, I mean, the. The challenge we have now, with all the deep fakes and everything is, is it's harder and harder to manage to where things comes from and in if it's actually real or if it's something that someone has created. So so we need to have some sort of control over things.

Mats Stellwall: 1:15:02

The level of regulations, I don't, I don't really know what. What is a good regulation? But I, I think, is it too much today or too little today? I don't, I don't know, because I don't think we have really. Everybody is just holding, a lot of companies are just holding back, so nobody is challenging the regulations in order to get them more clarified and we don't know. Until they start challenging these things, that's when we get what it actually means and we can start to see on it. But I'm not completely for a totally unregulated I don't think anyone really is.

Anders Arpteg: 1:15:43

I mean some regulation, I think everyone.

Mats Stellwall: 1:15:45

I think that but, but maybe we should talk more about responsibility, uh, and these type of things, and transparency. I think that is more important, because that will also help us in that sense that if we have transparency in the data we're using and how we're using it, and if we have also clear responsibility of things in the sense of, well, who is responsible for it, and so on and so forth, I think we will come a long way with use these type of things.

Anders Arpteg: 1:16:15

I think now but we can also see effects. We recently had Meta releasing the Lama 4 models last week and they were not allowed to be used in Europe, and we've seen a number of other models that are from OpenAI and Google that are not allowed to be used in Europe.

Mats Stellwall: 1:16:32

But is that I mean I haven't read fully. Isn't that more that they have chosen to not release? Or have they been told not to release? Because that's the difference?

Anders Arpteg: 1:16:44

The reason for them, not, I mean. It's possible for you and me to download these models from. Meta but according to the license and the fear of regulation, you're not allowed to use them.

Mats Stellwall: 1:16:57

Yeah, but that is because they don't want to take the fight, which I think is kind of a yeah, it's a little bit, I don't know the right word I want to use for it, but I think they have the muscles and they should definitely take the fight. I mean it cannot be worse than it. I mean the worst case scenario is they are not allowed to use them in the EU. That is the worst outcome of fighting the regulation about this, and the best outcome is they can't, and I mean so there's nothing for sure. There's some money in doing it. But they can question the regulations and say, well, we don't agree to this, or we see we can, we are fulfilling it by X, y, z, but I think they kind of take the easy way.

Anders Arpteg: 1:17:48

It's a different way here and I'm, of course, pro-regulation. We should have regulation. I think no one really want to have a completely unregulated market. It would be horrible have regulation. I think no one really wants to have a completely unregulated market. It would be horrible. Just the impact of cybersecurity creating a new vaccine or a new coronavirus, using AI or for war reasons or whatnot. I mean it would be horrible, of course, but then still thinking more about where the level should be. And if we have the biggest tech companies in the world they are the richest companies in the world, they have the biggest legal departments in the world of any kind of company and if they choose not to take the fight with EU regulation, what other company in EU would do so? Yeah, that's a good question.

Mats Stellwall: 1:18:40

I don't know, I don't have a good answer to that and I don't really know, but I mean, yeah, and maybe we have too much or maybe it's just like there's a weight, kind of a waiting game. Currently it's with all the things going on in the sense of what what they want to do or what they don't want to do, and I think it's also, I mean, down to it. I a lot of this around is probably around the data that is used for training. That is the sensitive part of it as well, and I think there's needs to be some some managing of that or handling of that, in the sense of we need to have some control on what data they're using. And I think otherwise it's. I mean, yeah, I don't know, I don't have a good answer on that. I wish I had.

Anders Arpteg: 1:19:34

But Me too. I don't have a good answer for it, but I think it's an interesting question.

Mats Stellwall: 1:19:40

It definitely is, and I think I mean yeah, I don't know, I mean working with finance, and finance is a very regulated area. I don't think they see it as challenging as maybe other areas as well, because they're used to regulations in a way where everything is very tightly controlled. But of course, if it stops us from innovating and we will lag behind and these type of things, then maybe we need to change things.

Anders Arpteg: 1:20:17

Going back a bit to open source and I would like to hear your thinking here a bit and, uh, um, we, if we go you mentioned deep seek, for example, and the impact that could have. Do, do you think?

Anders Arpteg: 1:20:35

okay, let me phrase it like this we had a discussion on this part a number of times saying what the future will be in terms of frontier models versus more bespoke models, more specifically designed models that you spoke of, like for legal reasons or whatnot.

Anders Arpteg: 1:20:50

Yeah, and if I just say a bit what we have said before and please let me know what you think about that, what we have said before and please let me know what you think about that we could imagine a future where we have a set of a few set of really large-scale generic frontier models like the Geminis, like the GPTs from OpenAI and Groks from XAI and so forth.

Anders Arpteg: 1:21:15

They will be really really expensive to train, they will even be really expensive to use to serve because they are so big and that means that potentially a lot of companies will have problems using them. But then we can think of having a huge number of more smaller models, like the DeepSeek started to do, and now that we're seeing OpenAI with the 03 and 04, which is going down in size more and more and more, would you agree that we will probably have a future where we'll have a few set like a 510 kind of frontier models which are really huge and then we'll have like thousands or even millions of more bespoke models that companies do use for real, that is potentially fine-tuned for more bespoke purposes. Is that the future you would agree with?

Mats Stellwall: 1:22:12

Yeah, I think so. I mean, sooner or later, a lot of this is just going to be functionality, so, so, so task specific models that, then, is good at doing, I mean generating uh, as we were talking a data engineering pipeline. We don't need a big foundation model for doing that, we just need a model that that understands the platform it's going to generate it for, and so on and so forth. So, definitely, I think so because I think that once the technology is going to be more embedded into applications in a way or in solutions as traditional functionality specific in areas where you might not need to, there's not that much that is changing. It's pretty static in what it does and you might tweak it sometimes because we have a new release of something, and so on. I definitely see that because I think we don't need these full frontier general models because they can do everything, but we don't need to do everything. I mean, I don't know how about others, but I can spend a lot of time just getting a foundation model to do what I want, because I need to get the prompt right and it takes a lot of time to.

Mats Stellwall: 1:23:24

I mean, if I wanted to say, well, take this document that I have. I can give you a good example I was doing a demo, for in Snowflake you can share data. So we have data providers that provides data so FactSet, bloomberg and so on and sometimes these data providers, they are not good at using Snowflake features to document the data, so they don't provide descriptions on tables and columns, they provide a PDF. And having for analysts, in order to figure out which data to use, they need to read the PDFs. So what I was doing, or actually a colleague built a base demo and then I tweaked. It was okay, let's extract all this data, these PDFs, and then use an LLM to extract the data dictionaries and provide that in a nice JSON format that I can ingest.

Mats Stellwall: 1:24:20

And I don't know how many hours I spent to get that prompt right in order to get the JSON as I wanted it, because every now and then it gave me a different response. You know, just to get it to not say yes, I can generate a JSON, here is the JSON. Even if you say don't say anything else, then just give me the JSONON. So and and having a function that only can do that, and and it's an LLM, but but it's just specific and I can say here's the document, give me the data dictionary. I can, of course, fine tune myself in order to get it to do it, but I think these types of things we're going to have these bespoke small models that does that functionality, especially if we can get them down in small sizes. We don't need to have the same type of infrastructure behind it. It's got chips to run them and so on. Why not?

Anders Arpteg: 1:25:08

I mean both for economical purposes, because they are quicker to run right and quicker to fine tune, but I guess also for regulatory purposes. It's easier to say that you know these can't be abused in the same way as a monochrome.

Mats Stellwall: 1:25:21

Yeah, because it cannot do anything else. But it will give me a JSON with this and if I ask for something else, it will say I don't understand what you're asking for. And yeah, absolutely, and I think that what I'm meaning. We will get more domain specific models as an output of this act. Probably, hopefully, and I think maybe we well, maybe we go away a little bit from the foundation model race as well.

Anders Arpteg: 1:25:46

Awesome. Let me try a bit more futuristic kind of idea. I have this kind of a three layer thinking of how AI can provide value for companies. Of course we can have data science teams or data engineers working with Snowflake and other tools like that to find insights and build AI applications on top of the data, but it's rather limited to just have these kind of super experts to build it.

Anders Arpteg: 1:26:16

And then you can think, oh, we can potentially have a second step, saying people that may not know how to program Python but at least know perhaps some SQL can at least be trained to write more simplistic or tweak SQL as they are being seen. If we do that, then we have a second layer where more people like business analysts can start to find insights from data. And then you know the value from data scales in a rather large extent by upskilling them in how to work with data. Right, yeah, Then moving to the third layer, and then it's back to the topic that we spoke about a bit in the beginning, talking to the data, and I would argue that if we actually get to the point where we can talk to the data using LLMs or frontier models or bespoke snowflake models or whatnot, that would be a significant value provider, since the non-technical people could start to speak to data.

Anders Arpteg: 1:27:23

Would, you agree with this? Kind of three-layer structure of how the future can enable more people to find value from data.

Mats Stellwall: 1:27:37

Yeah, absolutely. I mean for your layers. It will enhance each layer in different ways because, depending on skill levels we have and I mean to go to myself, I'm not a Python expert at all If I'm going to go back to a language for analytics, I will prefer the R. But that's me, but that's also how old I am.

Anders Arpteg: 1:28:02

I'd love to get into that discussion, but let's not do that.

Mats Stellwall: 1:28:07

But yeah, so I can see that, as I mentioned in the beginning, that having LLMs help me in coding will make me more efficient and get me quicker, and then we can add more layers where I need to know less and less details to do something.

Mats Stellwall: 1:28:24

And I think absolutely we are heading there and I think and I it might be as we discussed, I mean bespoke models might be the answer, because we will minimize the risk of hallucinations, doing wrong things, which means we can trust them more and we can test them, because I mean, the higher up we go, the more controlled well, controlled is maybe the wrong word, but it needs to be verified that it actually doesn't do things.

Mats Stellwall: 1:28:51

If I ask what the sales figures was last day, it shouldn't make that up. If it cannot find data for last day, it should say I don't have these figures, you need to ask a different questions, or here's questions you can ask. So each level will require more of the functionality than we might have today, which also speaks to more bespoke or smaller models. That is very specific for that, because we cannot have it make mistakes or that many mistakes. The higher, the lower we are, the more mistakes it can do, because I can verify, control and correct it, but if I get too high up in the knowledge level it's getting hard. In that sense, Mats.

Anders Arpteg: 1:29:39

What's up next for Snowflake Anything you can share about? What can we look forward to in terms of AI when it comes to Snowflake coming years?

Mats Stellwall: 1:29:50

Oh, that is a good question, I think I mean, like everyone else, we are finding new ways of using this and providing functionality. A lot of it is, of course, driven by what our customers are asking for and talking with us about. So I think, as we talked a couple of times and briefly around, I think we will push some of the functionality further down or hopefully adding more functionality around, not just only talking to the data from an analytical standpoint, but also talking to the data from a data engineering standpoint. I think that that would be a logical step, but I don't really know. But I mean, we are expanding the way how we are building our agent API. Of course, we see a lot of potential there in doing more with these type of things.

Anders Arpteg: 1:30:42

So you basically ask some kind of initial LLM or frontier model or foundation model.

Mats Stellwall: 1:30:47

Yeah, it's the whole orchestration part of it that we will. I mean short term. Today we basically have only two tools the Cortex Analyst, the Cortex Search, and we will, of course, add more types of tools that you can add to that um and and, and I mean we, um, we. We acquired a company called data volo a while back. That opens up for us to ingest more different types of data easily, which will also help these type of things in that sense.

Mats Stellwall: 1:31:18

But I think a lot of focus is, I mean we do a lot of things from the core product still you know, making it more, faster and better in that sense, but I think also in the AI space there's a lot of things that can be added in making things easier, more accurate, more faster and and so on and so forth. So I'm cool.

Anders Arpteg: 1:31:46

Looking forward to that, I have to ask you a bit more sensitive topic, perhaps a more political topic as well. We will see. Yeah, because I think a lot of companies and I know myself and others are speaking about this, but companies are becoming a bit afraid about the current Trump era that we are seeing and that they are adding tariffs to whatever kind of companies that we are seeing and requiring, you know, companies to pay, potentially, some tariffs for things that we haven't done in the past and that, potentially, is leading to fear of American companies in some sense. Do you see what I mean? Yeah, absolutely. Have you discussed this in any way in Snowflake or what can you say potentially for European companies to feel safe that we're not going to see 100% tariffs when working with Snowflake?

Mats Stellwall: 1:32:39

I mean, to be honest, I cannot probably say anything in the sense of I don't know, and I don't know our stand in this and I'm not particularly the right person to talk about it and I'm not particularly the right person to talk about it, but I mean, the way how we do things internally and communicate hasn't changed anything at all.

Mats Stellwall: 1:32:58

We still talk about the same values that we had before the elections and so on and so forth. I mean, snowflake works a lot with inclusion and these type of things. The discussion about trusting an American company has been ongoing as long as I've been working at Snowflake and a lot of discussions have been around well, how can you guarantee that my data is not going into the US and these type of things. So we have had it's not the same discussion, but similar discussion in the sense of, well, we run Snowflake in a certain way, but of course we are using the cloud providers, which is American companies as well. So who knows in that sense? So I mean I don't have anything that I can particularly say because I don't know, and I mean currently we don't have tariffs on services, so for some reason that we might know or not, know it's excluded.

Mats Stellwall: 1:33:53

I don't know why, but we, we can all guess around that and I, I, I don't really know, I think it's, uh, we are in times that we, we will have to see what happens and we are running things like we did before the election, in the sense of that.

Anders Arpteg: 1:34:08

But yeah's what you call exciting times.

Mats Stellwall: 1:34:15

It is in many ways, but I could wish for them to be less exciting to be honest, agreed, or it could be exciting. On other things, yes, true.

Anders Arpteg: 1:34:29

So, Mats, if we were to move to a bit more futuristic kind of questions and I'm not sure, do you have any thoughts about AGI? Do you think we will reach a point where we will have AI systems that supersedes humans in any kind of aspect? I think we all agree that it doesn't do that today, but do you believe it will come to?

Mats Stellwall: 1:34:51

that point.

Mats Stellwall: 1:34:53

Well, I mean, if we look at it from a, if we don't have an end time when it happened, I think definitely yes, but if it's going to be near time, in five years, 10 years, 50 years, I don't really know.

Mats Stellwall: 1:35:09

I think if you would have asked me a year ago or one and a half year, I would say definitely no, and and then we have the whole gen ai, and then you can always argue like gen ai is very simple in the, in the way it's just predicting what the next letter is based on the previous letter, and so on and so forth. But the complexity of of it is is kind of mind-blowing, uh, and and how we do it. So I don't know, I and I, I I tend to realize every prediction I do about this is is wrong, but I'm not quite sure because I think still the, the, the, the general way of handling things, and and we haven't, as to my knowledge, really figured out how, how we work and and why we are so good at generalizing and I think, and and I do believe that the current way of how our models are built is not gonna handle that.

Mats Stellwall: 1:36:03

Yeah, we need some more, it needs some more, but who knows what, what the whole quant, what, what that will do, because we haven't seen what that is going to do in this area yet, and I'm sure it will change a lot of things for us once they start figuring out that, okay, we might be creating some AI models based on quant technology.

Anders Arpteg: 1:36:25

And then, who knows, Did you have to go there? Oh, it's one of my rabbit holes you have to go there.

Mats Stellwall: 1:36:34

Oh, it's one of my rabbit holes. Well, to be honest, I quantis. Quantis is an area where once I every now and then I read up and I said, okay, I understand it. And then I read something else and it's like I have no freaking clue what is going on here. I I was a couple of years ago. I was at the at the german event, and there was this guy that was talking about quant and how he was working at university there and they were testing using quant compute for training traditional machine learning models and if it was giving any gains and it could for certain tasks, but for certain tasks not, and so on. And he explained it in a good way and when I was watching it I was like, okay, now I understand. And then I was asked for my colleague to say what it was about was about and like I don't have a clue anymore. But it can be, it can be the same thing twice.

Anders Arpteg: 1:37:22

It's a hobby topic of mine, it is. Perhaps we could take five minutes, if it's okay.

Mats Stellwall: 1:37:30

Let's see how I can can manage.

Anders Arpteg: 1:37:32

Let's see okay, but let me give some kind of intro and then please disagree with me for any kind of statement I made, but you know I've been interested in that for at least 10 plus years. I actually wrote my first quantum computer program in 2014 and being part of a d waves as a's beta program, but anyway. So if we believe that quantum computers is basically this kind of entanglement and superpositions that can happen between qubits and then potentially we could have exponential speedup for computation in these qubits, that could be really cool stuff, right. But then I have a number of dots for this and you know, google have had their willow ship and microsoft recently had that myorama ship, and and they try to fight the big problem of decoherence. Decoherence means that as soon as you connect more than a couple of qubits together, they get disturbance from the noise surrounding it and then these kind of potential exponential speedups that you see start to degrade. So claim number one and please disagree is we have no single practical use for quantum computers today. Would you agree with that?

Mats Stellwall: 1:38:51

Oh, I mean, I don't know enough, but I I've seen to read that that they managed to do some nice not nice, but uh cracking some encryption algorithms with it. So so I guess that is uh no way they haven't.

Anders Arpteg: 1:39:06

No, no, no. I mean, I can say with certainty there is no single practical use for quantum computing oh, there is not. Yeah, interesting. So the only thing they've done is they have. You know, the Willow ship was about 100 qubits and with 100 qubits you can do nothing.

Anders Arpteg: 1:39:21

Ah they have 1,000 qubits, for I think IBM has something like that, but still you require at least 100,000 plus cubits before you can do anything practical in terms of breaking rsa kind of cryptation or anything like that. I still have um. Should I go there? Let me, let me do it. I have two kind of problems with quantum computers. For one, it won't scale. Yeah, because as soon as you add more qubits together, it requires them to be entangled with each other, and that will be increasingly hard for every single qubit being added, and that's exactly what we have seen so far. No one has been able to scale up quantum computers, so I don't think they will scale. Secondly, as Demis Vesavis said and he said it well and I hadn't thought about that before, but he said something If you take something like AlphaFold AlphaFold was this AI system from DeepMind which are able to predict the 3D structure of protein amino acid sequences, amino acid sequences Now, this is a problem that is exponentially in complexity.

Anders Arpteg: 1:40:36

Meaning, if you want to explore all the potential 3D structure that this kind of sequence of amino acids have, it will take forever to try them all out to see which one actually works. But that is not what the AI system did so. The ai system could actually predict of like a microscopic kind of selection of them. So you can use intelligence instead to say let's just explore a few of them to see will this 3d structure actually work. Now you can compare that with alpha go as well. Alpha go is also a game yeah, right and go and it has exponential explosion of complexity, meaning every kind of step you takes is impossible to try them all, but ai could still easily do it and it can now today compete and win over any human. So humans have no chance in playing Go or chess or these kind of games. It doesn't do so by brute, forcing every kind of potential step forward. It can find an intelligent way to just find the right path forward. So this is really what AI can do, both in Alpha Fold and in Alpha Go and these kind of systems.

Anders Arpteg: 1:41:46

Meaning it's idiotic to say that was a strong term. Meaning it's idiotic to say that was a strong term. It's not perhaps adequate to say that classical computers needs to explore all kind of exponential explosion of complexity when solving a problem. You can find smart ways to do it. You can find smart ways to do it. So AI could be a solution to find solutions to like if you take weather forecasting or if you take more exploring how you combine molecules for more biological purposes or whatnot. You don't need a quantum computer for that. Potentially, you can instead do what we've seen in the past. So even if you were able to scale up, you probably wouldn't be able to do it faster than AI could. Oh, that might be. Yeah, that was Demis Isabis and not my words.

Mats Stellwall: 1:42:44

Yeah.

Anders Arpteg: 1:42:45

But I agree with it. So for one, you can't scale them up. Secondly, if you could, you still could use an AI better than quantum computer could.

Mats Stellwall: 1:42:54

There you go Right. Yeah, it's interesting. I have a friend who works with encryption and they are working very hard now to make their encryptions.

Anders Arpteg: 1:43:06

Of course it's safe. It could be that we find a way to actually build and scale up quantum computers. It could be that we could actually break encryption with quantum computers going forward. So of course we should be safe rather than sorry and safe for it. And we actually do it. Yeah, anyway, we should. So I think it does.

Mats Stellwall: 1:43:22

Yeah, but it's an interesting aspect and I mean you certainly know more than me and I think it's. I mean we might just abandon that path and find a new path further down the road. But I mean, if we go back to Asia, I it's hard to say because we nobody thought that we're gonna do well. Few people thought that the DNA I step that would come that fast and be so, you know, advanced in what they could do and how it can do things. So so the leaps we do now is kind of so big, so fast.

Mats Stellwall: 1:43:57

But it might also be that we have reached the end because we don't really know yet. So so it's hard to say, but I think maybe not in my lifetime, but I think definitely we will probably crack it, because we always quantum no, no, crack data. No, no, crack the AGI. That's what happened in your lifetime, of course, right, well, I don't know. But I mean, what's speaking for humans and speaking against us is that we kind of we don't give up, even if we should, and we just go for it, and I think that is also maybe become our doom, but nevertheless, I think it will happen in some way or another. Do?

Anders Arpteg: 1:44:33

you have a favorite definition of ADI.

Mats Stellwall: 1:44:38

Not really. I think it's about generalizing. I usually, when I've talked with my children, we talked a little bit about AI, and you know that it's not that intelligent. And you know that it's not that intelligent and usually what I say is a very simple generalization is you, as a child, can learn very quickly that, okay, I can drink water from a glass and then I can apply that to everything that has a similar shape. I can say, well, I can drink water from that bowl as well, because I can see it's the same shape. And today's AI cannot do that distinction or do that generalization, because it's just, yeah, they haven't, they've never been exposed to the ball and I think once we crack that, well then things is going to happen. But but for me that if, if once it can do these type of things, that's, that's when we start seeing the more real maybe not intelligence, but generally being able to generalize in a way that it can actually do new things without me having to teach it.

Anders Arpteg: 1:45:37

I like Sam Altman's definition actually, and I like this kind of small, very simplistic definition. So he basically defined AGI as when an AI systems beats the performance of an average coworker yeah, that's a good one, right, an average one as well. And if you think about that, if you take whatever kind of co-worker you have, he may not be a top co-worker but he does average good. And you think, can I just take Chativity or another AI system and just replace him and say you can go home, I will just ask Chativity to do it or whatever ai system you have. There is no way you can do it today.

Anders Arpteg: 1:46:17

Oh yeah, I agree, because the only I like the, the metaphor of a autonomous car or autonomous driving saying you know, one part of autonomous driving is being able to perceive just the cameras around the car.

Anders Arpteg: 1:46:32

Second one is just trying to plan or find a way forward, both short-term and long-term, and then you have to take action in the control part. You know, should I turn left or right, or gas or brake, or just continue? I think perception-wise. And that, basically, is the knowledge that LLMs can have today and these kind of models, and that's better than humans, I would say, but the reasoning, the planning worse than humans, and control definitely worse than humans. And the control is basically the agentic part that we are still working on, but still very far behind, and certainly in reasoning as well. So once we do supersede humans in also reasoning and also in control and agentic aspects, perhaps we are getting close to performance of an AI system that could do things like an average co-worker, but we're certainly far away still, and especially in the physical world. It's one thing in the digital world, right, but in the physical world, in being able to do the dishes or whatnot.

Mats Stellwall: 1:47:31

That's far away. That's what we want. I mean, do the dishes right. That would be awesome. That would be awesome. Yeah, definitely, yeah, I agree, I think. I mean, as long as we need to instruct it so extensively to do things, it's not going to happen. I mean.

Anders Arpteg: 1:47:48

Okay, final question, mats. Yeah, Imagine that AGI do happen at some point in time. Who knows when, but it does happen. We can think about two extremes. Either it's the horrible, dystopian kind of future where we have the Matrix, the Terminators, the machine trying to kill all humans, or it could be the opposite, as Nick Bostrom wrote in his book Deep Utopia. It could be more of a utopian future where AI have solved the societal challenges we have. It's fixed cancer, it's solved the climate challenges, that we have Fixed the energy crisis and whatnot, and we potentially live in a world of abundance. I guess both are very extreme. Yeah, do you have any thoughts on where we will end up? Will it be the dystopian, terminator future or a world of abundance?

Mats Stellwall: 1:48:42

Well, I mean, I certainly hope for the later, and I mean well, given how we humans are, we're probably going to end up in the first one, because we tend to mess things up. But I think in the end I think we will end somewhere between, because I don't think we. Well, maybe in the end, but I don't think it will take. Even if we start getting an AGI it will. I mean, it will take a while for it to solve a lot of things for us and it will take longer than people think.

Mats Stellwall: 1:49:12

I think so too. It's not like well, suddenly it figured out how to cure cancer. I mean, we do have gotten some advantage there and we know a lot of things, and it needs to. I mean it can only base things on the knowledge that we have. So first it needs to create new knowledge before it can do something new. But who knows? I mean it would be. I don't know. I don't want to live in a utopia either. I mean I'm not quite sure. What are we humans supposed to do? I mean just walk, I mean.

Anders Arpteg: 1:49:45

That's an interesting question. You know, nick Bostrom wrote about that and he said you know, people are so afraid about not having a meaning to live or something. But then, on the other hand, if you take children, for example, yeah, that's true, they, they are living a very utopian kind of life, in some sense that they have been taken care of, and even like retired people, you know they have their actions coming in and uh, yeah, so some people, or even pets, I guess, or a cat or dog that you have, they've been cared for, they, they don't seem to hurt that much no, probably not, but but I mean they also.

Mats Stellwall: 1:50:22

I mean, if we I mean children they do it, for I mean a a limited part of their life and and the more they grow they, we are decreasing the amount we provide for them as that. But I mean it's one thing that a small amount of people, I mean, doesn't do it. But when everyone is supposed to figure out what they want to do apart from working, and suddenly you're responsible for, you know, for planning what you're going to do, not just for afternoon and evenings, you also have to take care of the whole day and you have to do that every day for a foreseen future, I mean it will put a lot of requirements of what type of things we can do in the sense of keeping ourselves busy. Keeping ourselves busy and interest doing interesting things. I don't know, but maybe that's me.

Anders Arpteg: 1:51:16

Who knows? I think it will be exciting years to come, for sure.

Mats Stellwall: 1:51:21

Definitely for sure. Definitely for sure. I, I, yeah, I would say that With that.

Anders Arpteg: 1:51:26

Thank you so much, mats Stelval, for coming here and the amazing discussions, and I hope you can stay on for a short after work as well to discuss even more topics.

Mats Stellwall: 1:51:36

Thank you for having me.

Anders Arpteg: 1:51:37

It has been a pleasure.