 Dynatrace is expanding its DVI engine to create the industry's first hypermodal AI, converging fact-based casual AI and predictive AI insights with new generative AI capabilities and to deep dive into this announcement. Today we have with us Alois Redbar, Chief Technology Strategist at Dynatrace. Alois, it's great to have you on the show. Thanks for inviting me. What you folks are announcing today? What we're announcing is extending our DVI, which has been around for a very long time with a new component, which is called DVI's co-pilot, a generative AI component. But the real announcement is not that we're just adding an LLM to it. It's really that entire platform that our customers have available, which we refer to as hypermodal AI. You might wonder what is actually hypermodal AI? I think we can describe it maybe in two, maybe four sentences here. First of all, hypermodal AI is a combination of different types of AI. We talk, as you mentioned, about predictive AI, which does forecasting, anomaly detection, has some basic understanding of what metrics means and causal AI, which can do cause and effect analysis. We do it a lot for problem analysis. What's the impact? What's the root cause? Instead of having raw alerts, you combine those. It uses a multitude of mechanisms to do this behind the scenes, using an understanding environment with what we call smartscape analysis, understanding topology of the environment, and so forth. And then there is obviously the LLM component, which is David's co-pilot, so that that's one way. We're combining multiple ways of AI together and make them interact with each other. The other way, what we're doing with the data is we take that raw input data, which is traces, metrics, logs, user sessions, make the data events and so forth, and they're semantically enriching them. So what does this mean? We take like these modalities of data and convert them into what I would call a hypermodality. Like you have a time series stream of data, and we say, wait, this time series stream has actually response time. It belongs to a certain service. It is linked to a certain service level objective based on the service level objective. It's behaving like this. This is its error budget. We also predict this metric to behave like this going forward based on past data on the predictive side. And we also know about this metric. That it was involved in a couple of problems in the past once it behaves in a certain way. For another one, though, our analysis might look entirely different. This is a conversion rate. Again, it has goals. This is how it behaved. This is the user sessions, user interactions. It actually depends on and so forth. And again, even linking that metric then maybe to a response time metric that we had before. So we're really changing that or augmenting the meaning and then like taking the input from the predictive, feeding it into the causal AI when we do the root cause analysis by detecting anomalies, see how things fit together, do the cause and effect analysis, not only telling you that 15 things in the systems are wrong, but telling you this is the actual reason it's related to a deployment, this changed in this deployment, and these are like the impacted users and so forth. And we can take all of this information. For example, to create a remediation proposal or to do like other types of automation based on it into an LLM. So obviously prompt engineering is this critical piece you do with LLMs that you have to tell it what you actually want it to do. We can create like this very specific, very context intense prompts to make it either generate something automatically like the remediation examples or we could go the other way around where the user asks something and we augment that query by the user so that we can automatically build notebooks, dashboards, workflows or even answer questions what they should be doing in their environment. Think of a very simple question. You're asking the system, okay, show me all services that will be relevant on Black Friday. If you would ask an LLM out of the box, the answer would be Black Friday is usually a day of very high intense load. All of your services that are subject to have erroneous or failure mode behavior on the high load are the ones that you should be monitoring. Yeah, thanks, but kind of like thanks for nothing. I knew this up front. What we would do is we take that input and that's where you see the different modalities playing together, take that initial input of the LLM, understand what it actually means, and we go back to causal EI, like show me all user interactions that have conversion rates. For all of those user transactions, show me all the systems or the services in the system that are related to those, then we go to predictive EI. Tell me which, what's the SLO and the target goals for these specific services, response time, failure rates and so forth. Then we again ask for the predictive EI. For those services, how would they most likely behave under a significantly high load? And once we have collected all of this information, we put it back into a very large prompt to tell the LLM and to build the dashboard that we initially wanted for us. And that's a very different output than just telling me that I should take care of the services that might have problems under load. So I think these are two good examples of what it does like on the analytics side and how it feels for a user to act with a hypermold EI. One question that I want to ask before we go deeper into especially what you folks are doing is that we always hear buzzwords that it came a few years ago or a few months ago NFT was a big thing. And now we are talking about Genitive AI. Of course, we have to dip our toes in all these technologies to see how, but do you feel that Genitive AI is the next Docker containers because that technology and it changed the world, then Kubernetes changed the world? Are you saying that, hey, this is blip? I mean, it's hard to predict, but since we are talking about predictive AI, so let's hear what you think is the reality there. I think it's going to do two things. It's going to make us rethink of how we interact with systems and how good we want systems to be and also as our main goal is like really increasing productivity by what LMS can do and to be fair, they can do certain very impressive things. There's also where we believe like why should anybody still build a dashboard, write a query and do those kind of things when you have capabilities that can do it. But at the same time, and that's also why we combine it with other types of AI technologies, we should be aware of that these systems are only as good as their input. The question shouldn't be like, is LLM the technology that's going to change anything? Which other technology does it need to be combined with? And we believe other forms of AI have to play together and even if you see other tools that are emerging and tool chains like Lung Chain and others, where these AI agents are now starting to emerge. It's not LLMs only. They play a crucial role, but they do not make like a full AI platform. And I think we will see that some disappointment on people like solely relying on LLMs. One of the favorite example we use over here is ask an LLM to tell you how much is 5,000 times 3,000. That's a very simple calculation, but that's not what an LLM was built for. What it could do is it could figure out that you wanted to do a calculation, use a third-party plugin to actually perform this, actually outsource it to another model, to another service. And I think that's where we will see a lot of development. I think it will really depend on what those architectures looks like. It's not just, okay, I have an LLM here. I'll ask you something. I'll get an answer. I might ask you a bit more clever things on the prompt. So we will see both. We will see certain systems getting more and more and more powerful over time. And others will kind of like stay where they have been for a very long time. Think of like Amazon Alexa. In the beginning, this was like amazing technology, our Alexa. And what are people using it today for? Playing some music and using it as a timer when you cook extra, put something into the oven. Because the technology and the backend services and the other capabilities that were actually needed did not involve us quickly. There were like no real proactive components in there. So it's actually a base technology. And JetGPT for sure has exposed it to a large model to almost everybody in the world who has a computer right now, maybe. But it's not the built-in product. It's a demo application. I think people have to come over this technology enthusiasm and learning how to build the systems on top of it. Very good point. And that's what my next question is that we are still in a very early phase where we are seeing a lot of excitement, but you folks are moving with a product which actually going to help businesses in real production. Can you talk a bit about, once again, going back to as you're giving the example of Alexa, how much work teams, organizations need to do themselves to once again take leverage of some of these new generative AI technologies, because that sometimes becomes a hurdle, roadblock, because things are already complicated, economies crazy, where you folks come in with these technologies, lower the barrier of entry. So they start taking advantage of these versus investing too many resources. Does this question make sense? I think if you got it right, how can people leverage the technology without having to invest and learn a lot? And I think that's key. For us, we also don't want people to become experts in prompt engineering for David's co-pilot. That's not our goal. Most of the AI on the causal and predictive side has been around for a very long time, almost a decade. It has been used for that long. So this is like ready-made customers are using it today. We believe a lot of the analytics is going to happen ubiquitously in the background. We don't expect the user to be very knowledgeable to put in information, understand how to use those systems. Our goal is really to make it as natural as the example that I shared before. Show me all my Black Friday relevant services. Most likely this output in a pure LLM-only scenario won't produce a good output. So we do all the heavy lifting and all the platform work in the back. And I think that's really the key here, where we invest our time is like really building this hyper-model approach where we decided to not just go for an LLM, but build an entire platform that can be used in a way that sometimes is interesting automatically for you like proposing, hey, it seems like you're trying to build this. Maybe you want this in this component as well. All the way that you can ask questions like right straight away. And for us, this was also a key driver. Like we have all these capabilities in the Dynatris platform. But or luckily we ended, we're glad that we added a capability to also cater more to business audiences with business events. But do we expect the business user to really understand how to write a DQL that's the Dynatris query language query? Do we want to teach them? No. That's why natural language input comes in. I think it's really for those AI systems that have to take the burden away to interact with them and provide value without people yet having to learn another technology. I think this will drive adoption going forward. How good do they get at understanding what we want and how we do it? And how big is the margin of error in prompts that we can put in there? And even if you look at what OpenAI states on their website, they say the quality of the prompt will impact the output. It will impact how much it should hallucinate. That's why I'm not just going for like one model, one technology, but you're going for an integrated system, platform type of approach that takes a burden away from you. I think those systems will eventually be the ones that will be competitively and will stay. I will not ask you to give any names or something like that, but I do want to understand if you can share either some of the use cases or industries where you do think that the David AI or co-pilot is going to help all these other industries which are like very, really quick to embrace and all these industries they should because that's going to help them a lot. We have a very wide customer base and actually everybody who's using Dynatris is using David's AI today. So they're predictive and the causal parts are used today and how does it help them? One use case is very clearly problem analysis like root cause analysis and driving automation there. So we know our systems are getting bigger, they're getting more complex, we run in multiple cloud infrastructures because we use different services from different clouds or ended up there for other reasons. Teams are getting bigger, have to take on more responsibilities, applications themselves get more complicated and how we help them is like really manage these systems from an observability and security point of view and being able to fast the recess problems and by automating this analysis of problems and how they work and then triggering automations based out of this to not just get pure alerts but saying okay this is the root cause, this is the impact, acting on the root cause, remediating also the impact is where we can help and by putting this into David's AI and doing it fully automatically we can bring down the meantime to repair massively and that's just not a bold claim but it makes very much sense if we look at it about it how it would have to do manually. You would have to look at a dashboard or multiple, you would have to make sense out of it, you would have to run a couple of queries then you have to think about like how all this data fits together. This will take you minutes or hours or in some cases even days if you don't know where to look for an automated AI system this can be done in milliseconds. So MTTR is massive there so when you look at the benefits for customers it's really twofold. They can really with a very small dedicated team manage very large infrastructures and again this is not about like that any jobs will go away. All of our customers are in the opposite situations. They have more work or more tasks and approaches that people should be working on that they have people and they can't hire them for different reasons and obviously the other reason is when we talk about MTTR especially our customers in the travel and e-commerce industries but also the financial industries now more and more as well downtime or problems with the applications can be related directly to revenue loss. If I talk to an e-commerce customer and ask them how much is like that's one minute downtime mean for you they can give me a dollar value and that dollar value is six, seven or eight figures depending on the size of the business and as we're mostly working with the Fortune 50 thousands expect a lot of them to be eight figures obviously for e-commerce it's obvious okay people cannot buy only a shop so using money but think of the travel industry you cannot board your airplane what do you think is it going to cost if your airplane has to leave like one hour later from an airport just because you couldn't find people's luggage or you couldn't board that plane which just recently had a case study with an airport on this topic as well or banking as you may have to do some critical transaction and the bank will tell you about our banking systems are in maintenance they might or might not be up tomorrow so that's where we can already help them today and that's where we're adding massive value now we're just taking this to the next level one obvious use case is what we call predictive operations there are a lot of operations tasks that in that exactly know when you have to do them simple thing like disc resizing automatically scaling infrastructure but how do you know when you have to do it this is again where you rely on the the predictive and causal capabilities of davis ei like tell me what my response times or my lowest most likely going to look like in half an hour from from now tell me which other services will be affected by certain behavior and then driving auto scaling in your environment or like tell me what my disconsumption will be which takes actually some time to resize this these are simple tasks that we are automating but i think even there we have reached the next level it's like they're not the simple simple tasks these are well for human simple tasks but these are tasks that require some cognitive analytics capabilities to properly drive them they might even require a couple of like steps of cognitive analysis until you really arrive at the actual goal what you want to do that's also why we opened up davis ei right now so that you can more just build your own analysis chains based on you what you want to do and again to your earlier question does it make it harder for people and we deliberately built it that way that you don't have to be a data scientist or machine learning expert to just tell you what you want to do i want to know what this metric will look like in half an hour from now i want to understand which other services this this service is interacting with so like very easy to understand easy to comprehend analysis steps that you can build and chain together and then based on top of it there's lots of other productivity enhancements like automatically creating dashboards ai-powered notebooks like an analysis notebooks where you only need to specify what you want to get at from that the result and then queries charts and so forth to build automatically for you automatically proposing remediation hints based on the root cause analysis that davis ei has done like you need to increase your memory settings here you need to tweak your kubernetes configuration over there also the same for creating workflows whenever a problem occurs routed to the responsible person which means you have to extract ownership information set up cheer attack getting and ticketing and align stuff but in a massive productivity scale so i think as we have seeing more and more platform teams today where we have decentralized observability and security offerings at the same time people want to customize it to their use case and what this will also allow to do this mass customization so that everybody in the enterprise gets exactly the view they want without having to learn another tool how do you see that generative AI and this kind of you know technology solution that dynamically says to me how this is going to further you know evolve observability how you see the evolution of observability through these technologies yeah i think if you look at observability and i've been in this field for quite some time it really moved from like this almost niche for like very large enterprises highly specialized expert technology which is also where we started like it has like an observability back then we called them apm tools it was like this expert tools what we saw the more accessible we made the platform with dynatrace either by the way we structured the product the way we explained things and more we saw people using it like a wider adoption in the enterprise not just the expert teams but the individual teams as a general pattern we saw like that the more easier accessible you make the platform the more people are going to use it the more people have access to the data they can work with it and they're going to use it today observability is i think widely accepted way more than it was at the very beginning when dynatrace started to to enter that field like it's something that everybody needs to do so we need to make it more accessible to even more people and i think that's where especially the productivity side the getting started side generative AI can help a lot because you don't have to be that expert you have again a lower entry barrier and you can still go into the full power of all those tools that they actually provide with all of their capabilities because people might feel overwhelmed they're now i have all this data what should i do with it and the the actual question that you wouldn't want to ask the system is like what should i look at now uh tell me what's like what's like important and again this is a good example of where the other than itself wouldn't be able to provide a good answer and that's why we believe in hypermold AI because if you would just ask you about should i look at it yeah response times are important error rates are important um resource consumption is important so you should look at all of those while versus versus combining it with causal and predictive and while these services used to cause some problems in the past these are your most critical ones which again might be for example close to running out of an error budget on your slo's or these are other connected services that should be relevant these recently changed so the answer quality will be much better i think the this this entry barrier and even like helping people to use tools properly and the observability we often talk about the unknowns unknown the unknown unknowns like well we can't explicitly ask for what we want uh because we don't know yet i think especially there it can help like tell me what's important tell me what's going on in the system which is nothing you can immediately write into a query where the system wouldn't have an answer unless we have like this whole uh combination which we refer to as hypermold AI um to provide you these answers so i think we will see a commoditization of the usage of observability tools even more as we move in that direction i think it's commoditization in a good way it's not that the tools lose value and the people don't see the value but they become way easier to use way easier to adopt it requires less expert knowledge the tools will explain even more what they're showing why they are showing it i also want to ask as you earlier said you know we have been leveraging AI for a very long time talk a bit about how this approach is a bit different from if you use the term traditional legacy approach of leveraging AI our approach always was to use the best technology for the best drop so we always take machine learning for a lot of areas we work in doesn't really work like for the cause effect analysis it's not the ideal choice for the predict if it was not the ideal choice that you want to have obviously different algorithms that you're using and again today we see some people experimenting using lm's for cause and effect analysis which to be fair for very simply use cases work but as you understand how lm's actually work internally and how they were trained it actually doesn't work that way so i think our approach is different that we always agree that we need to combine different types of AI together and also that we need to augment the data we need to have interconnected data what like the second dimension of the hypermodal approach to do more qualitative reasoning over the data that is not just a metric no it's a response time it belongs to a service it's related to a login we're currently users are having problem so our approach was really building like this powerful integrated architecture on how to work with AI like rather looking at the individual pieces and just applying them on some data which like where the first type of AI was as well like just do some machine learning in a response time and even having the user decide like we also take this out of the user's responsibility like we take make these choices for them and ensure that they get the right data without even having them to become an expert in indian analysis and also yes you don't have to open up the whole to understand how the mode is working you just can get in and drive that's that's our approach and how we how we work with these technologies lice thank you so much for taking time out today talk about of course your announcement also going deeper into whole generative AI approach how it will also kind of transform uh observatory thanks for those insights and i would love to chat with you folks again thank you thank you