 And we're live on YouTube We're ready. All right, we're good to go Hello everyone good afternoon guys have made it this far first day of dev comp. That's pretty good So what we're gonna do today, we're gonna talk about the impact of AI and We have with us Jared Uli and Daniel and they will be our panelists. I am charard griffin the moderator I'm here to keep these guys honest and make sure that they tell you the truth and nothing but the truth So first off we're gonna do some introductions Jared, maybe you want to kick us off? Sure. I'm Jared play as your microphone. I think so There we go. Yeah, I'm Jared play. I'm in the office of CTO working on technology strategy Yeah, who did rapper work? Also for the CTO Chris writes No one knows what I'm doing. So I'm mostly looking at all kinds of compute things among them is machine learning Then you recall to office of CTO and I am managed the AI center for excellence working on retards AI strategy So we're gonna start off with Pretty generic question to get the ball rolling. What do we think about the impact of AI as it deals with software development? IT operations DevOps those types of things. So I'm curious Daniel. What are your thoughts on that? so well, I think that overall AI is this big transformation and In Many of the shiny topics that we we see in media like self-driving cars and all that are like hard problems to solve and there are a lot of low-hanging fruits for applying machine learning to improve IT itself software development operations embedding AI capability machine learning capabilities in our software Products in the platforms and you know, it's the picture used earlier with if you're driving a self-driving car You probably don't want to suss admin in the trunk But that that car is basically a data center on wheels And so in in order to get to that you want to automate things to a degree beyond what we have done so far With IT and I think that applies like generally to what we do with with software Jared in your line of work. What do you think the impact of AI will be? Well AI is interesting because it covers a lot of traditional analytics as well as machine learning techniques and I think I'll talk a little bit more about this Later on but I think that there's a very interesting set of patterns of newer applications that are built around data flows And bringing data into the system doing some sort of processing in it Which may include AI processes and then taking action on that and that may be taking action locally Or that may be taking action at a different point So Daniel just talked about autonomous driving as an example autonomous driving has the full suite of AI ML Techniques and at play but most of those aren't happening on the car. They're happening in a training process That's happening In a very large data center infrastructure So there's flow that's happening that's occurring where data is coming in from vehicles for Asynchronous training purposes. It's being analyzed and then that's creating updates that That are then delivered again asynchronously to the vehicle to improve its behavior. So managing how that data flow happens managing Security of that data flow the reliability of that overall application process The the general assurance that it's doing the right thing is going to be interesting because that same pattern occurs across use cases such as IOT internet of things now, they're kind of large vape, but Complex area use cases as well. So overall, I think we're going to see a shift to these data-centric applications of which AI is a core component for doing analysis So when you mentioned about Security and things like that you see a place in AI to maybe mitigate the risk for businesses or at least expose potential issues I think from a security perspective in that case. I was talking more about Making sure that you have Templates for what your application life cycle is so that you know that you've done appropriate integration testing You know, you've done appropriate regression testing before you put something in what can be a safety critical situation AI techniques also can apply for Security use cases that really just had a talk Two sessions ago about using AI techniques for anomaly Detection or intrusion detection all the things that we've traditionally built rules based systems around But then AI techniques allow us to create much more flexible much more reactive environments All right now given that Linux is you know the biggest open-source operating system What's interesting is there's a lot of Linux developers out there now Uli can you kind of speak to what you think AI means to a Linux and a systems developer and How that might change their paradigm or change their approach to development? So I don't necessarily think it's limited to Linux developers, but there are big opportunities when it comes to to developing any kind of complex System nowadays where we reached the point that Individual is not actually able to understand every single aspect of this It's not able to understand the signals which the system can actually bring out and recognize it as something positive or negative or something like this and therefore Not so even if they're sometimes recognized them there might be biased in a certain direction that they actually don't want to recognize the signal as being something good something bad or something in between and so on and getting mathematical logic in place to To do this kind of work to do the analysis of the work Will open up completely new venues. So just to give you an example Also a couple I gave earlier today a talk on micro architecture of CPUs and so on so this is something which if I would for the room with a hundred people and throw a stoner probably Could not hit the single one who actually knows anything about these kind of things like market architecture This is a very as a tech topic, but at the same time if anyone wants to do performance analysis using today's Performance counters, which I've ever been a city you you cannot do anything without actually understanding what they mean So but with some if we're writing some logic some mathematical logic around the analysis We can actually learn what makes up what measurements actually describe a good workload As opposed to a bad workload and have the AI actually learn this kind of things Without the user actually able to describe what makes it good about so we can learn from the systems directly and circumvent the lack of understanding which the Which the user that might not have what a program might not have and this extends to the set beyond just the Linux world So it's for everyone some a big advantage going forward. You just have to build these kind of systems Okay, Daniel. We have any thoughts on that as well. Well, I think this It's a What is this general pattern that we can use AI brought the sensors with buzzword To derive knowledge directly from data. We have the ability to generate a whole lot of data now the complexity of our systems are beyond the capability of humans to understand and AI can help us make sense of things or like an autonomously derive meaningful information from this very complex Set of vectors and amount of data, right? It's so the problem with the data is it's it's the Volume but also the complexity of data That just beyond like what anyone can still grasp as a human You just have a limit on how how many vectors can you consider when trying to understand? What is like what's the root cause for a problem and a machine can help us? And you know the other aspect is that something which has been worked on also is some of the tasks are mundane This is and therefore programmers are not likely to take them out voluntarily So monitoring is CI system of some sort or recognizing very simple faults like well Yeah, you didn't commit this message is this this check in here and son. It's missing something. So this is done And sure I can talk about this as well So these have done using work using bots oftentimes in some form where simple Tasks might not be even have to be handled by humans anymore And therefore we get much more reliable systems because these kind of mistakes if they pop up can be very quickly rectified So with the automates sounds like a little bit of the you know removing the manual tasks that a person might have to do I do have you guys experience any of those examples of things Maybe we're doing in red hat in that space or or other tools where you've seen a good job of it eliminating that manual mundane task I've certainly interacted just on I don't know the the project or product But I've interacted on get out the projects where I've reported bugs and immediately had a bot come back with you Might find these other ones relevant and it's a very simple Conceptual piece of code, but actually having it produce useful results was enormously helpful And yeah, we have examples like that that we're In the inflake analysis and see I for example Where you it tells you if there's like somewhere deep in the in the many As big of like how a system can fail They look they look random to a human observer because you don't see the depth of you know It was like on a full moon and what is some someone tortured a black Schrödinger's cat on on the graveyard And you don't you don't know that but but it's somewhere in the log and it's a in the metrics of the of the cluster where you run Run your test and after it happened five times You know there's a pattern there because somewhere deep in the vectors of the of the data It's a clustering of issues and so the system can identify that that's actually not a flake There is a problem somewhere hidden deep in the code that as a human It just doesn't happen often enough or it's too hard to understand what happens and we have that today It's improving our quality as we speak Is not a common build You'll be surprised you would be surprised But to the other side that this is actually also an interesting thing where we need some more quite a lot of research going on Most of the works in machine learning nowadays are built on the law-flush numbers where you need statistics to actually catch a relevant result So what is going on? To some extent nowadays and will be hopefully the focus of more work is that we Actually can work with very small data sets and derive patterns out of that in a reliable way So the thing which Daniel mentioned that we might get to the same error situation Five times in the sixth time. We can actually to do something about this ahead of time. That's very very useful We're trying to do some some work in that area. So I have my personal helper on that kind of thing to do the math for me and Looking forward to actually getting results in that area Jared I want to go back to something Daniel mentioned earlier and that was part of his Presentation earlier about data being the differentiator You know having all of this information all of this data and the way you use it how you make it available Both internally and externally to your team So how do you see with data being the differentiator? What does that mean in the AI space and how does that impact AI in general? Well, that's a very interesting question. I hopefully can give an answer that won't get me into trouble so One thing that's very important about not just AI, but AI is really driving this is driving value Further up the stack. We've already driven value from lower level software to higher level application infrastructures What AI techniques do is that they drive value from your AI Infrastructure to the data to the models that you're building with that data and so this can be very very challenging because You can ship software that's open source software Which is fully capable of solving a particular problem and may even have the hooks You know the input and output hooks to solve a particular problem But if you don't ship that with a working model then the software itself is useless And if you don't ship it with the data to train that model then then you'll be limited potentially in what What an open source user can do to enhance that Similarly shipping a model in and of itself can be challenging because it depends On what data has been used to train that model that if that's confidential proprietary data Your model can potentially leak information that you don't want to be leaking So I think that as ML grows and importance in the open source community that there are going to be some very interesting Conversations about what license do we make these models available under what about the data sets? Do we have open repositories for the data sets and the models that are training these newly critical systems? and a lot of the discussions that we had many years ago around Licensing and software we're going to have to have those same conversations Again as it comes to the data and the models in ML based Products that certainly there are data licenses that exist. There are content licenses that exist They may or may not be appropriate for these particular types of data or may not be Nuanced enough to describe the variance and licensing that we may want to allow That's interesting Daniel, what are your thoughts about the difference between open source data and these closed off black box systems? Obviously you have someone like an Amazon or Google where they've collected all of their own data Therefore they have the power with their own algorithms, but as you mentioned earlier, that's a little bit different with open source Can you kind of dive into that a little bit more? Well, it's that two sides of it So I think that for a lot of things where we're going to use AI we need Open source even more than we needed with the pure surface of it. So the the difference is that Traditionally we have it's all about code, right? We are in a we code centric it The source code is complete and describes the functionality of the software With ML suddenly you need to data describe the full functionality as a training data To describe the full functionality and depending on what you're doing like that might change like when you're running the software the one problem is that if You know if you have a machine that derives Information draws conclusion directly from data takes takes action based on that There's no human in between anymore. I think that even more You know drives a need for transparency that that only open source can provide It's that's one aspect of this Trusting a black box service is getting harder and harder you more you more Knowledge or intelligence. I because that's like in there's more intelligence isn't is in that service And then on the other hand, yeah, I think What Jared said and I'll be curious what would Uli's perspective is on that, you know when Open source Versus AI and like how do we keep that consistent? We've reached kind of a taunts with Firmware blocks, which are opaque binary data that's necessary to have a system operating Yeah, I don't think that we can stay static with that same level of acceptance as ML models become more critical to software Yeah, so the data aspect is certainly so it makes up part of the solution But beyond so David Penning mentioned also the implementation implementation is Important as well. So that it is actually freely available and can be inspected because these these the software which are built for from modeling, let's say a deep neural network or other other techniques and so on they Well, why would you trust them actually to do something? So that could be hidden somewhere in there if the user is Greek then spit out a thousand dollars. Otherwise, zero Good height is in there So it has to be inspectable and this also means that for the data itself even if the implementation is right And someone is delivering the the trained model with it. It has to be replicable. So in machine learning really is nothing but I bet the scientific process actually put in reality for everyone to use and one of the big aspects in science is that you have to Design an experiment and make it reproducible Because otherwise no one is trusting this in the same will now be true for every single program out there Therefore we have like Jared mentioned at some point we have to think about how we're delivering models along with this But also the description so things like notebooks which have been available for a long time so mathematical programs like this at them forever Notebooks are going to be the way how you actually describe how you arrived at the state that your model is right now And then people can say oh, yeah, I see this I can replicate this and therefore I can trust the service I think that the so firm whereas It's a compromise if you were willing to make because hardware is proprietary like hardware is physical, right? So it's It's somehow an anchor to the to something proprietary historically. We're just changing right and like I wouldn't say firmware is trusted So just look at the Intel's People it's only good people dealt with it like we're actively working to do away with the stuff So we actually lack the ability to be introspective to audit the hardware that we run on to the same extent So I think I think Daniel saying that that the firmware is abstracted into the same level that I don't decap my CPU and Use a scanning electron microscope to validate that it's doing what it says it does and there could be all sorts of No, but I think by like at the point where whole operating systems are embedded in the CPU Like and not the best operating systems, right? Like I think it was minix is obsolete or something was the version which they use with many questions behind so Like at that point, I think like I think you know, it's it's probably It's a bit of different it's time for open source hardware, and then we're getting there anybody said that's another thing so working on that as well and and I'll move Lee like It the same like the more the more autonomy machines get the more important this is gonna be because we what they take over the more we have to trust them so and then that becomes a very fundamental question for a business for for Society well, and I think there's a very related security Issue that that you raise That goes back to the training data for models and the necessity to have that be replicatable in an open source product because one of the big challenges with learning systems is that The ability to be introspective into the reasons why they make the decisions they do is a much younger area of research than the learning models themselves and so to be able to one ask You know on what basis for these decisions made and also to make sure that there isn't a Trojan force in the model that as as we suggested if you're the right person then you can take all the money out of the ATM Because it has an ML model deciding what it should give you which is not a very good use case for All right, so Talked a lot about the hardware part of it in the latter part and the hardware versus software what about software in terms of The data versus the source code when we start to talk about the entanglement of Needing to have enough data for the algorithms But then also having to have the right algorithms that can analyze the data. Daniel. What are your thoughts on that? That's a very broad question. No, but the main thing is there The other science machine learning AI some people call it. I don't know why Is not something which you should think you know you can apply if you just know that oh Yeah, I can't call this library and be done So you have to think about a model as it is produced by a machine learning algorithm like a function So if you put in good data, you get good results out perhaps, but you put in garbage You get garbage out. So how do you to be able to successfully use any of the machine learning techniques? you have to be able to Distinguish between the model being garbage or not and that's not that easy So this is actually something which requires you to understand what the math is You know what the weaknesses of the various techniques are Etc. So this is not not as easy There will never be black boxes where you can say well, here's my data do something with it and in the end Well, it will work always. That's never going to be the case I think and Therefore you have to be careful with what you're actually doing. So who is actually doing the work? And so anyone can produce a model, but whether the model is useful that depends on lots of quality factors Do you see that as a challenge for organizations? Maybe Jared? I don't know if you've come across this but Companies want to move towards AI and ML and they implement these But it does seem to require quite a bit of work to tune fine tune get the feedback from Whatever service is being used and make sure that that model is accurate our companies prepared for the workload that it takes to Actually tune these models on an ongoing basis Well, I think what Uli was getting at was actually just a microcosm of software development as a whole which is that Many businesses are very for it understanding what software they want to develop and therefore the software that they produce doesn't do what they want but that we don't have the ability to open up a text buffer and Write a short description of what the software should do and have a system automatically generated and as Uli points out we never will because half of the hard part of software or at least a large percentage of it is Deciding and determining what it's supposed to look like in the first place that I'm not suggesting that the right way to go about this is the big design of front multi thousand page specifications But understanding what problems you're trying to solve are critical to selecting the tools you're going to use to solve them Validating that you're solving the correct problems Validating you're solving the problems you've set out to solve and then deploying that software so AI and ML techniques are are just another set of tools in the box for solving business problems personal problems educational problems societal problems that we want to influence with with systems and so in the business Framework you have to know what your problem is before you know what tools to solve it You're going to use to solve it and you have to know what those tools are capable of before you can select them And so there's a whole spectrum of tools for data analysis from traditional statistical techniques, which is a lot of what these these data science workbench products allow you to work with To the newer ML techniques where it largely comes down to Again knowing what problem you're trying to solve Can you clearly define the inputs and the outputs that you want out of your process? Whether it's statistical or ML if you can't then you need to go back and figure that out because none of the tools are Going to solve the problem for you And then it's simply a matter of selecting business something where when looking at the problem Can I you can I describe what the rules look like you know is it that there's a very large space of Possible input scenarios, but I have a very easy way of describing What rules apply for what actions I want to take then neural network isn't the right tool necessarily that there are other Rules-based systems that you can use on the other hand is really easy for me to create a huge set of example Data that I can feed into an algorithm where I can define very clearly 10 or 100 or a thousand parameters from that input data set where I can clearly define what outputs I want But it's really hard for me to describe in relation to those inputs what rules need to be applied That's where more machine learning Returnal network approach makes sense. So these are all tools that you have and yes Lots of people are using them wrong and lots of people are using them because they're buzzwords You know, we can have our AI AI IoT blockchain talk after this file for an IPO for Startup is going to do blockchain neural networks, but I'm the case to do that with blockchain I think you can do it all about change. That's my understanding is that I mean the IPO or Ico So there's one other thing and this is confusing many people who don't necessarily have the insight in there. So often times People are using machine learning, especially the deep deep learning deep neural network learning mechanisms in what is called unsupervised learning situations This is the entire business model of likes of Google or Facebook, etc So they are trying to learn in an unsupervised way the massive amounts of data Something about you. They don't actually know about what it is. They are just looking at similarities of some sort It's on so in a completely unsupervised way. So there's no human doing an intervention. So why does it work? Why do I say? Well, this doesn't work in general. Why does it work for them? The reason is simply that the impact of them getting the model wrong is that you are seeing an ad Which you don't like Big who cares? If on the other hand you're doing these kind of things and you're writing some piece of software Which decides whether to shut down the cooling in a nuclear reactor? Guess what? It's a little bit more important that you don't want to have this So there's there's a last spec spectrum of course in the middle, but I would argue that Pretty much everyone has to be taken serious doesn't have the Freedom to just ignore the quality of the model so most of the time you really want to have something which is Which is really solid in its results. So in my previous life I worked for a financial company and so on and there they could not do these kind of things Automatically because the result might have been that you're losing a couple of hundred million dollars. So that's not doable It's not just human life with one might be a steak and therefore I would really argue that these these Nice-looking prospects of machine learning as which come from the use of Unsupervised learning might be good for things like marketing Is that fair for answering these kind of things but pretty much no other area of life where you actually depend on the result All right, so we're just about out of time. Do we have time for any questions? Okay, so we'll take a couple of questions anyone have any questions Sure over there Do you have another make? Just a second take that come with the mic Hi, hello, I just wanted to know your views about AI and gaming Well, they interact with human constantly and there is a lot of criticism, which isn't one Yeah, and gaming I wanted to know your views about AI in gaming gaming Who's a gamer here? No, so I did write so I did writes my own games for a long long lying time And what we call the AI back then doesn't really compare to what it is today. So it it's up But I think you're mostly referring to the open AI and so on trying to solve games Using a essence on this. This is a completely different thing. So they're using this mostly because it can be done 100% inside the computer without tiring anyone out This is just an example for the solution where the input is not a binary one zero and so on its visual input So in theory, so there in that there's a there's a similar experiment where you can actually if you own a GTA license you can actually Write your own a driving simulator using GTA on it on the machine Versus it then just does a screen grab which would be the equivalent of a camera which you have in the windshield So that's the only thing why these these companies are using games for that because it provides something where they can cheaply create situations with very extensively complicated inputs and Trying to write something which we thought we couldn't do before in this and so that it's games It's beside the point. This is just a coincidence Great, I think we're just about out of time here Want to thank the panelists for joining us and they did say you get a special door prize If you guess who's been at Red Hat the longest All right, thank you