AI Architect Interview Questions (Ace Your Next AI Architect Interview)

are you looking for generative AI architect questions if so this video is for you hi my name is Michael Gibbs and I'm here with David lithicum and today we're going to do a roleplay interview on generative AI interview questions because I know a lot of you are going to be looking for your first generative AI architect job and in this video I will be interviewing David lithicum a little bit about us before we begin so you know where we're coming from I have been a network architect and an Enterprise architect for over 25 years David lithicum has been an architect of all kinds and a chief technology officer for about 35 years he's an expert on generative Ai and he is the author of our generative AI architect training program so I'm G to interview the expert here so you get to hear it the way it should be so you know how to win on these interviews David could you describe a recent project that you worked on that included generative AI what was your role and what technologies did you use yeah excited to um to talk about that that was a project uh worked on for a large retailer where we created a recommendation engine uh based on uh Gathering data on people who did not register for the websites in other words you go to a website you start looking around and based on your behavior we can discern whether you're a male or a female demographics income uh any number of things you know based on uh huge uh amount of training data that we have uh from past customers who uh are you know pres basically doing the same pattern so the idea is and lots of retailers use this is we're able to take this engine and recommend things to the customer that they may not have known that they need it so in other words we see somebody is you know interested in uh in me in men's clothing um we know that there's a probability that they also may be interested in camping or hunting Sports things like that your ability to kind of test it and see if that's something that they're looking to do then based on that response your ability to recommend other things uh and so the engine that we created uh was able to increase sales by 30% uh and this is is all without uh customers having to register opt into technology this is just about them using the site and our ability to upsell them just kind of based on the behaviors uh other thing before that I you know I built a uh uh generative AI driven supply chain integration system uh during covid the supply chains were absolutely a mess we couldn't get anything moving people realized that uh they were under automi under automated and uh under optimized so I built a gener AI system that would in essence provide complete Logistics uh planning and training uh based on where the thing was being built and manufactured versus where it needed to go dealing with the supply chain including the suppliers for the manufacturers building the product and looking at very complex parameters uh that are very difficult to manage normally and you consider tens of thousands of data points at the same time and your ability to make decisions based on what those data points are is to the best optimized way to build something to ship something to make sure that the supply uh suppliers are producing uh the raw materials that are needed for manufacturing just the ability to kind of get to Soup To Nuts in terms of how you manage something from a customer saying I need something to the building uh the selling the shipping uh the sales order entry all these sorts of things and where it's a complete automated system system where you're relatively assured obviously things go wrong from time to time but it's able to operate around those things where you're going to be as optimize as you can and in that case uh a manufacturer of tires was able to uh optimize their supply chain and it generated uh another $30 million in Revenue just in the quarter which in which it was implemented so those are the two uh that I've done done others but those are the ones I like to talk about David really great examples plus I really love that you came back and you talked about the impact of the business that's great and thank you good experiences there now David how would you approach data set preparation and preparation for training generative AI models well the core thing is that the more clean and understood that the data is uh the better it's going to be used within a generative AI system generative AI systems rely on clean and hygienic data in other words it's able to do provide the information that you need which is Paramount but the ability to make sure it's in good order and we we know what it is and how it's represented so we're Gathering you know very diverse data sets uh in many instances could be images and PDF files it could be customer data could be inventory data sales data all these sorts of things your ability to gather the data in such a way or it's organized to to be Pres presented to the generative AI model where it's able to learn and train from these data in a very clean way and this by the way is where probably 90% of generative AI projects go off the rails is they don't spend the time in investing in the existing data sets and understanding the data and getting the metadata correctly getting the data hygiene in good working order all these things are fairly uh easy to implement they're not overly complex and overly orous processes but um it's figuring out where everything's coming from what it means and your ability to present it properly to the models that are learning from the data so they can build the knowledge models around the data so it's the normalization augmentation the ability to uh you know reduce dimensionality of the data the ability to remove some of the complexity all these sorts of things the more that you can do that where it's in an understandable state if you can understand as a human being then your generative AI system can understand it so that's really kind of the metrics that I use obviously tools data hygiene tools metadata man man agement uh data integration systems you know all these things kind of come into play but the core value in this is your ability to get these things right your ability to make sure the data is in a state where it can be consumed where it is going to be the source of truth that the engine is built upon you know people consider generative AI as some sort of a magic trick where it's able to uh deal with bad data and it's able to find bad data remove bad data some of that's in there but not a lot so the core thing if you're it's it's garbage in garbage out with generative AI data and in fact it's worse because if it's building a knowledge model based on erroneous data unclean data things like that that knowledge model is going to have that uh is going to have that issue built into it that's very difficult to remove so start with clean data data hygiene's in the way understand where the data is how it's consumed and that's a foundation for generative AI success D I love that you said if the data is going to be the source of Truth it has to be clean expert answer there thank you so walk me through how you would design and train a generative adversarial Network for image generation what metrics would you use to evaluate its performance yeah basically at the end of the day um we're looking at core parameters um that are able to teach it how to create images um from random noise and discriminators that learn to distinguish between uh things that are real and things that are not real are fake so the process basically Loops through it so the idea being we're going to get it wrong getting get it wrong getting get it wrong getting get it wrong um based on these images that we use and until we get something that is moving to something that is the desired state so it continues until the gener generator becomes proficient at producing you know the images and you only do that through practice and trial and error so the great thing about generative adversarial networks is that it's almost like two robots looking across each other at a conference table and so they're producing ideas in this case images it can be text it can be other things as well and they're putting it up and the other one is looking at it in verifying it and in essence making sure that it's right and if it's not right it it will reject that image and that's where the adversarial comes from so these robots kind of fight it out uh until the images are correct into a state that's the desired outcome the idea of generative generative AI networks is are always becoming better at taking the information always learning how to do things better and that's an example of it so if you're designing um if you're designing a generative as of Serial Network and by the way you're normally not involved with that as an architect you're picking tools and Tech technology and the engineers are doing this on your behalf but the understanding the concept is that uh we're picking a particular approach to dealing with this information such as a gam and that the adversarial nature of that gets to a better more desired outcome again based on the information coming in that's in a state enough where it's able to you know generate the right the right systems so it's a it's kind of a fun little thing if you're picking this technology you don't see this happening but in the background you'll notice that initially when you come to these conclusions or basically creates these images and image could be a song could be you know text generation whatever it's the same sort of processes that are involved it takes a tremendous long time initially for it to train learn and generate the responses for the inference engines uh but over time it becomes more proficient that's where it gets like almost like humanlike characteristics as we train something we become smarter as people and the smarter it becomes the faster it's able to generate the information the better it's able to generate the information so the ideaas we're using this architecture is kind of a checkin balance um you know I like to talk to my clients and tell them it is about like two robots sitting across the table and they're having an argument and but they're really really smart robots and whatever they come up with is going to be an answer that's going to be better than the sum of the two decisions that they made and then iterate that again and again again end before they produce images that are uh that are amazing so it's the power of generative AI it's and this by the way applies to images and text and you know uh music Generation video Generation all these sorts of things are are available to us now the generative and generative AI is we can generate information and output in any number of ways that we want to do it so it's not limited just to images but it's interesting to explore kind of what gam is as it's related to images and that's kind of something it's easier for people to visualize yeah and that's well explained and most people don't really understand that aeral Network because they can't see it but it's so critical to see both sides of things so terrific David have you ever had to overcome a situation where your generative AI model suffered from node from mode collapse and if so how did you handle it yeah they all they all do at some point um so it's when the model repeats itself or you know fails to show a wide range of diversity in its outputs and in other words it's limited it's getting things wrong you know at the end of the day we can get into why it's getting things wrong but there's something wrong with the information that is consumed or how it's processing these things from an inference engine but it has to be fixed in other words we're just not getting the outputs that we need and so in mode collapse erroneous data you can call any number of things it's just a mechanism a way to define an error that is an error state that it's getting to and the cause of it's getting to so when you fix it you promote diversity within the data sets you may change the model structure there's some tuning capabilities you can put in there uh and your ability to go through and continuously improve the in uh information in the inference model so it's able to produce things uh better and in essence reduce the impact of mode collapse if not if not completely eliminated so diversity of information ultimately is where you get to the better outcomes in this in other words our ability to look at a massive amount of data where it's going to have a visualization across the data where it's able to discern patterns within the data so if you can't do that you can't provide the diversity of information and uh and the diversity of outputs of the system it's not going to get to a state that's going to be usable the good news with this is that there's tools and Technology out there that will fix this for you so Architects typically aren't dealing with this on a day-to-day basis this is kind of an engineering concerned they should know what it is and how to address it but as far as the mechanisms to fix it uh those aren't necessarily in the domain of the architect but it is in the domain of the architect to understand how these errors are going to occur and make sure that we're building an architecting a system that's not going to uh not going to go this way and if it does our ability to have an automated fixing or self-healing mechanism uh to reduce the impact of the mode collapse and and remove it directly from the uh on the system thank you for that yeah ethics you know so David in your experience what are some ethical considerations when working with generative AI particularly in the context of saying deep fakes or content generation that's a great question uh you have to remember that generative AI uh is a bit risky when it comes to ethical concerns just because of the power of the system so we're able to do things that are very concerning including generate video you know of uh you know people who have more than uh you know 10 15 hours of video out there on the open internet and it's very difficult to discern if not impossible to discern so we have a lot of ethical considerations because we know that we can do something we just need to ask if we should do something and that's really where ethics comes in so this information these knowledge models that we build is going to have um bias information in it they're all biased to the degree of bi how it's biases really kind of depend on the training that you use but it comes down to um looking at the harm and impact this is going to happen for human beings so where there's biased information deep fakes your ability to generate uh you know erroneous photos your ability to uh you know U produce things that are uh going to be um influence people more than they should all these things really kind of come to bear so it's working with people that understand what the best practices are in terms of ethics and one of the things I always do on my team when I run an architectural generative architecture program is I have an Ethics specialist you know someone who's has a deep understanding in terms of what the best practices are out there and what we should do versus what we can do we can do a lot of stuff we can look at uh uh the information and biases for example when I talked about that uh recommendation engine that we built there's lots of things that you can do with that data um that will help the business and provide it a better inside track and understanding the customer but there's also ethical concerns in doing that and you have to kind of make sure that we're not Crossing that line and this is a line we're going to come up to many times and so my advice is number one look at the reality of the in of how you're going to build these systems and impacts that can have on human beings and society and nature but also get somebody who understands how to deal with the ethical concerns on your team that you can uh chat with and we can put an Ethics plan together and bias elimination plan together and we can do bias audits so we can un people can understand if there's any bias in the system in a particular time and you have to remember that um it's not when you're going to go to it's not if you're going to go to court over this it's when you're going to go to court and so that being the case uh you always need to build these generative AI systems of if as if you're explaining it to a judge in a jury uh and so we're always thinking about how to practice defensive ethics doing the things we we we should put in there so we're not harming human beings but also the ability to um audit and track and log all the various things and decisions that we're making so if we do have to explain it to a court or a judge or a lawyer our binding arbitration or even the the board of directors things like that we're going to be able to do that so it's much different than traditional systems where you don't necessarily have the power to make a lot of do a lot of things are going to cause ethical concerns in generative AI systems you do and I think that a an Ethics program is going to be bound to all these projects in fact I wouldn't work on a generative AI Project without some sort of Ethics counseling and auditing that occurs within the system it just removes removes the risk really great response I also really like the way you said it just because you can something do something doesn't necessarily mean you should do something that's something we all need to think about y so you know David you know you've been in this world for a little while what are your thoughts on the current state of generative AI research and where do you see it heading in the next few years yeah obviously it's uh uh the focus of most people are in the technology world right now doesn't matter if you're building operational systems or security systems or platforms uh everybody's looking at how they're going to work and play well with this model it's unlike anything I've ever seen in my career and I've been doing this stuff for 30 30 plus years so the focus right now in terms of the state of the generative AI system is probably too focused on the Tactical tools and Technologies um focused on the processors gpus tpus you know focused on some of the engines that are people being built and focused on you know basic who's turning out what whether a cloud provider is better at generative AI than the other cloud provider um which is fine but it's not something we can do longer term so longer term we need to focus on how this technology is going to evolve in a couple of years and where is it going to be and where does the business need to be in terms of consuming that technology so it's less thinking about the Cool Tools and the processors and the and the uh use cases for generative AI that are you know out of this world in terms of you know fun stuff to think about but the ability to look at how we're going to find the business applications so we're getting things planned we're getting things processed to make sure that we're going to build something and that's going to be a value to the business in one year or two years down the line so it's not focused on The Shining objects it's focused on uh the real world benefits of this stuff and I think utilization in the use cases of generative AI for most businesses is going to be very tactically oriented they're going to build smaller language models and rather than large large language models it's going to be dealing with strategic uses of the technology to differentiate the business and really kind of create the Innovative differentiators uh within the marketplace for example the the two that I mentioned the right the uh uh recommendation engine and the supply chain integration system intelligent Supply Chain management uh would be instances of that but the ability to automate manufacturing the ability to uh you know have a common knowledge mechanism across robotics within a uh a factory the ability to put intelligence at the edge so we're able to manage oil platforms and ways are going to be safer and more productive and so all these things should re be put on the radar screen right now in terms of how we're going to use the technology and the Tactical uses of the technology I think we have a tendency to be too focused on uh the large language models that are out there and what they're able to do and those are going to be important and we're going to use utilities out of those and you know they're accessible VI API are going to be part of the process but the bigger question which people aren't answering now is what this technology means to our particular business and what we're going to use it on so you go to uh meetings where they have a generative AI steering committee for a large company uh they're just reviewing the cool stuff they're talking about the uh releases of the next coolest llm they're talking about the different Cloud providers and how they're approaching generative AI uh and not necessarily the core Technologies that they're looking to build so the focus should be and my focus is on where that technology is going to be in two or three years and what is going to be the relative value and meaning to the business the thing is if we're going to be generative AI Architects or any kind of architect uh we have to have that in sight uh so it's probably too it's too much focused on the things that really aren't going to matter uh in a few years and not enough focus on the Tactical uses of the technology for the particular business domain and I think that's a harder and not as fun problem to uh problem to address and I think that that's where I focus and that's where I think businesses should focus on where this stuff is going so how we're going to use it how we're going to employ it and kind of watch that The Shining objects are not necessarily the distraction that they are right now and David you know excellent you know it's where the tech is going and what that Tech can do for the business I'm in complete agree with you not the speed the feed or whatever the coolest invoke Tech is thank you for that you're welcome David could you discuss a time when you optimize a generative AI model for better efficiency or performance what steps did you take yeah the best one would be the um recommendation engine that I built because if they don't operate at speed then they're not useful because people aren't going to hang around for a website if it keeps going away for five 10 seconds at a time so in tuning these models for performance and optimizing these models to provide better uh better Knowledge Management better predictive management better uh you know infant responses things like that it's going to the tools that are able to carry out a few core operations you know such as residual connections your ability to to make sure those are normalized the ability to look at Progressive growing and normalization techniques of the data so in other words if I'm building this knowledge model um making sure that I'm building it for uh a good reason in other words we're adding true knowledge we're not adding redundant information we're not overtraining the model we're not training it on things that it doesn't need to understand and you remember most of these things as we talked about are going to be purpose buil they're therefore a particular business case they don't need to be something that's going to learn from everything that exists on the internet they just need to learn about their particular domain and and the decisions that they need to make so it's looking at what the model is currently doing and kind of asking the question does it really need to do this and if it doesn't then that functionality that knowledge should be removed from the system you shouldn't worry about processing that stuff looking at how it's going to grow and how it's going to perform uh in terms of model performance over time but also how it's going to become better and more approved at the time so when you optim optimize a uh a generative AI model you're going to be uh constantly approving in terms of how it's going to come up with answers and how it's going to manage information so it's always asking the question good news is um were they these things used to be done by hand uh someone would have to write an application or a script uh to this stuff now there's tooling that's able to do this on your behalf so optimization of the generative AI model depending on what model instance you're running the tool that you're using the brand the type version all those sorts of things are going to have the optimization tools that are uh connected to it so a lot of this stuff isn't necessarily in the domain of the architect and a lot of the stuff probably isn't in the domain of the engineer we just need to understand that these things have to occur and ensuring that we have the right tools that are carrying these things out so and also this is something that's going to change a great deal in the next couple of years you have to keep that in mind so the way in which we optimize these models now which just uh you know got into it with the Advent of generative AI uh and learning as we're going with those is going to be very different in two years so um my recommendation as well as doing you know residual connection Progressive growing normalization all this stuff which is fundamental to doing this is that we're able to pick the right tools technology and approaches are going to assist us in making this happen that doesn't mean we're focused just on the tools that means we're focus on what needs to be done and the domain of tools out there that can help us get this done and our ability to bind it to those tools you know putting the complexity of the of the optimization of the generative a model in the domain of the tool versus it being something that a human being has to go after every few weeks that never scales because you can't get consistency there in repeatability there uh it's something that's not going to be normal nor optimized unto itself so if you're optimizing a generative AI model it's about understanding what you need to do to make that model better from a design perspective from an architecture perspective and beyond that it's leveraging the right technology so this optimization is done in an automatical way behind the scenes so it's really just kind of part of the infrastructure sounds great David David how comfortable are you deploying a models and production could you talk about any experience you have with say mlops practices yeah the the the great thing about deploying uh uh AI models specifically generative AI models into production systems is they follow the same sort of processes and many instances the same tooling as traditional Dev SEC Ops devops so continuous integration continuous deployment your ability to version the uh the models out there your ability to uh uh lock the configuration in in terms of the hardware software Cloud configuration that you're leveraging you know all that all those toolings basically have followed the generative AI system and so they're able to provide the same automation so it's a good news bad news the good news is we don't have to hire lots of new people who are going to do the cicd you know integration and model monitoring and tuning and versioning out there and it's going to be very similar if not the same tooling but you need to bind it into the devops tool chain you need to bind it to the particular applications that use these models you need to version these models in much the same way and they're going to require a little bit different understanding and how we do versions versioning it's not just an instance of a database it's not just a instance of a of a of a Cod Tre um it's going to be a complete knowledge model which is this huge monster in many instances unto itself that we have to lock in in a particular configuration so we have to have the toolings that are able to make that happen by the way this is not mean that you're taking a copy of that knowledge model uh putting it off someplace because it take a huge amount of storage uh to maintain these things but we are taking a copy of the configuration uh the tuning parameters the model the model numbers you know some some of the U the knowledge data in the system even the training data so we understand how to return to that state um certainly as people start out with generative AI systems they're going to get to a point where they realize oh darn we made a mistake they probably won't say oh darn um and they're going to have to go back to a previously stable version of it and so that's the ability to hit the reset button go back to different state of training data different set of the model instance and if you've done this versioning and you've done this tracking along the way that's going to be fairly easy to do and you're going to protect yourself you're going to lower the risk of doing the development um I'm finding that people had have a tendency not to pay attention to that so they haven't put these into model version techniques uh uh uh configuration management all these sorts of things and they get to a point where they go oh darn we made a mistake and we built this model using erroneous data that was dirty uh for instance um and we're trying to get back to a previous state so we don't have to retrain the entire model and deal with the overhead and doing that and they can't do that so in instance what they did is made a mistake they have knowledge engines uh knowledge models that are built on something that's that's wrong and they have to figure out how to undo that stuff after the fact and those aren't databases those aren't uh Cod trees those are you know this state of a system where um it's more than just data but it's an array it's a network of knowledge relational systems that sits Within These uh llms and you're going to have to do a little bit more in order to get back to a previous stage so your Investments should be made in in cicd technology your ability to look at the model versioning technology and work with your particular uh AI integration vendor your AI tool provider so we're always going to get to a state where you use either one AI tools normally it's going to be three or four uh and how we're going to build these things and then use them as the sounding board for ways in which we should version their models you to remember they have the relationship with the with the tool vendors and the technology providers but at the end of the day this is a about you planning the processes and you putting the talent in place to make sure this stuff is being done right now and about 80% of this of the generative AI deployments it's not even if they have a very robust uh devops tool chain and they're leveraging uh you know the right processes things like that and they look great in that it's not including the generative AI systems yet and that's going to be hugely dangerous because we may find there's an issue with the model we have no way of diagnos having Diagnostics we can't get back to a previously stable State and figure out the differences between the two so it's a very small investment that should be made uh within your deployment system uh that will um save you millions of dollars I think at the end of the day so it's something people should do right now and look to do it with the right tooling the right technology but at the end of the day it's your planning that makes this happen sounds great so David what types of generative models have you worked with and in what context have you deployed them um well uh in the recommendation engine I worked with an image generation model that was also a Content generation model so it was generating text as well as images uh in some cases videos instructional videos for people who were leveraging the website so the retailers the generative AI model in that instance would generate a custom image very cool for a particular person based on who they were so it wasn't just presenting uh uh images of camping equipment as we mentioned before we you know we realize they're shopping for men's clothing so therefore they may uh there's you know 70% chance they they may want to look at camping stuff as well but your ability to generate an image that was specifically targeted at uh how that person perceives camping even the the area of the country where they are so if they live in Arizona it would generate a desert image if they live in the North Northeast it would generate a forest image all those sorts of things where it provides that extra bit of motivation to get them to look particular at the equipment Al also text of the system instead of generating static content the ability for the Genera tvi system to generate uh unique text for the particular individual so once we know who they are we know what level they read at we know um you know what they respond to you know key wordss that are in there even putting it in a different language if needed uh once we figure out that there their German speaker for example and they're they think it's very cool you figure out I'm a German speaker and you start showing the thing in German so it worked on those your ability to gen generate images and text um the ability to um uh in the supply chain scenario your ability to generate um logic around and Logistics around a certain supply chain so we're not generating text we're generating path through a map for example or and processes in place and the ability to kind of stepwise through this stuff and in terms of how you build and deploy uh a particular piece of equipment that you're selling and the raw materials good producers and what they should be doing all those sorts of things obviously we're generating text to you know generate the POS and the invoice and the way in which we control the information but the larger issue is that we're generating a map we're generating the content so we're gener generating in essence entities uh which may be represented as images and text but the larger thing is regenerating the plan and how these things are going to and work um ability to spot uh fraudulent systems in building Banks and obviously that was where we're inputting data and your ability to look at checks to spot for patterns that may lead to fraud uh in building those systems um looking at uh you know real-time images uh and the ability to identify them you know such as security systems they're able to look at uh um uh able to look at images of people who are approaching the building identify who they are and where they come from in fact they just had a uh issue in China where they're able to find a a serial killer uh based on this technology um and ultimately uh any sort of thing that's able to get to the state you remember the important thing here is that we're building the model the ability to generate uh different outputs of the model that are facilitate whatever kind of a business outcome you're looking for is a fairly easy step in the process so but your ability to build knowledge model in such a way where it's going to have this kind of flexibility the ability to generate this stuff in any way you want to consume it uh is is going to be where the work needs to occur so generative models to deploy all kinds of different outputs based on the needs of the particular business but the core models ultimately you're using the same steps the same processes to build the same stuff thank you for that so you know David quality matters how would you assess the quality of generated samples from a generative model well you can always look at the uh things like images and text uh to see if it's something that's acceptable you know one of the things if people have used chat GTP chat GPT before when they ask it to write something that's fairly simple you'll notice that in in many cases it's a mile wide and an inch deep so it just generates lots of content so your ability to have a logical assessment of something as to whether or not it's going to value or not it's something you need to do so it's visual inspection you know quantitative metrics domain specific evaluation criteria that these businesses put in there in terms of what's acceptable in terms of what the outcome should be so everything in terms of measuring uh the accuracy of a generative AI system is going to be determined in the outcome in other words the output excuse me so one of the things you need to do if you're going to test your quality is create some scenarios where you know what the out output should look like you know what the picture should look like the map should look like you're getting to the right answers and you're able to ask it a number of these questions and the inference engine puts out the particular data sets that you're looking for and do so with multiple random ways um other good news you can leverage generative AI technology testing tools to test these models to figure out what kind of quality you're getting out out of them they'll provide you with a ranking and they'll provide you with ways in which you can fix and adapt to the model all those things are available but when you look at how these things are tested it kind of comes down to the way we tested applications in the past we put stuff in them and stuff comes out of them we're trying to make sure the stuff that comes out of them is in a right is in the right State and if it's not something's wrong with the application whether we're white box testing or Black Box testing generative AI a bit more complex but we're doing the same thing in other words we're asking it the questions we're invoking the apis we're providing with the business problem and in these situations we have maybe a thousand different scenarios where we know what the what the output should be and what kind of answers we should get out of the system and you look at the the generative ability to get to those answers by the way it's never going to be 100% these things are going to make mistakes are going to be off a bit things like that but that's how you tune the system also use testing tools to make this happen those are starting to emerge right now uh we're starting to see huge amount of interest in there and how you test these generative AI models so you can look at the quality of output also the ability to leverage these testing tools where the outcome or the results of the test in this case quality assurance goes back to the model so can iterate itself and improve itself over time just like we do with human beings in other words we're on a factory floor and the someone's putting the door knob on the uh on the Cadillac wrong um and it's a repeating pattern then we can go back to them saying you did it wrong let's put it the right way and it's fixed generative AI models work in the same way they can be instructed they can be trained um and they don't seem to take it personally so I like working with them yeah they're not going to take it too personally great good point there David David could you describe a challenging project involving generative models that you've tackled yeah ultimately the ability to build a uh fraud detection system uh for a big government government agency was probably one of the most challenging things that uh I worked on number one the technology was very new it was just more of an Innovative uh research projects at the time they obviously wanted to invest in it so there wasn't a lot of technology and talent out there I had to in essence build the approaches to make the thing happen and kind of understand my different architectural parameters going forward using what I knew from the world of cloud computing and Enterprise architecture and application architecture things like that so it was what I didn't have a lot of things to draw upon other words there weren't a lot of people I could learn from there weren't a lot of books I could read about this there was some PhD dissertations things like that but it was something that was uh was definitely going to be me me figuring this out as we go along well that became the challenge so we you know looking to develop this deep learning generative AI models your ability to leverage the technology in such a way where it's going to have the right outcome for the client and your ability to an Essence build lots of these things from scratch doing full well by the time you deploy this thing there's going to be other products out on the market they're going to solve this for you so that was a difficult thing to do because I felt like I was going to be repeating the wheel or so Reinventing the wheel for something else and someone else was going to step in and make it happen so what I did was to make sure that everything was very modular in nature and so in other words we could plug in the generative AI tool set again if we have something better coming along we're using ETL tools the ability to uh deal with data hygiene and batch processing you load the training models your ability to deal with model data management in different ways your ability to bind this you know into a devops tool chain and that was was difficult because it was probably too early and that was my advice to the client other words we're building something before I think it's going to take off and if we wait a little longer then it's something that we can build with better tools and have an under better understanding of the market the more we let it mature and maybe we take a baby step now build the training data system and the the business analytics system and then we put this on later uh that wasn't the case uh they they wanted it built and uh it was built uh 6 months and that was a very aggressive timeline didn't know exactly uh how it was going to be built when I started out didn't have a good understanding of how it was going to be uh done because the tools didn't exist um but we're able to pull it off and uh pull it off with lots of very talented people around to make it happen and by the way that's key to all this your ability to lead a team and attract a team that's able to take on various components of the system and take on their role and responsibility so they're doing the data science they're doing the AI engineering they're doing the this case it was a deep Learning System was bound to a generative AI system uh they're doing the model development and management uh they're doing uh the selection of the infrastructure and processors uh whether it's going to be on premise or in the cloud and you're bringing the whole thing together uh the what makes me happy is that system's still running it's uh still viable it's still working well um um probably needs a refresh in terms of its technology but I was asked to do something at a certain amount of time um I advised the client as to what they should be doing uh so some of that was taken under heart some of it wasn't but we forward anyway we're able to pull it out and that's probably the most uh fun I had and also the most challenge that I had in building an AI system and I've been building them since the 80s now David it's a great response and I know exactly what you mean when you're doing something that's too early to be done but at least you made the point you talked about making it modular so you could change it and adapt it I love that I also love that you made something very clear you told me that you know it was a team because so many people when they come and interview for AR an architect job they try to act like they can do it themselves and they can't it's going to take a team so I'm glad you're mature enough and have the experience enough to to be able to communicate that to me thank you you don't do this stuff on your own it's a team effort absolutely you know David I know we touched a little bit on ethical considerations before but I want to go there a little more because it's so important with this topic what ethical considerations are crucial when deploying generative models and how do you address them yeah it's a great question I I think the the ethical consideration that everybody should be concerned with would be uh the bias and your ability to mitigate it so all of these models are going to have bias people always tell me they don't they don't want a bias in their generative AI model in their llm and there's no way to there's no way to get away from it in other words if you're using real data data is going to have a bias and sometimes it's going to be a bias that you have to mitigate as it comes into the system and in mitigating the bias we need to identify it we need to figure out what part of the model that it exists in and move it out of the model or basically even put counter you know counter uh data in there so we're mitigating the bias so that's the big one because that's the one that's going to get you in trouble that's going to have you um uh turning down loans uh for a certain demographic of people and perhaps uh not selling in certain parts of the United States because we're not addressing that part of the United States for whatever reason uh your inability to uh identify and use information in a way that's going to be fair to everybody who uses it and so the ethical considerations really kind of lie there if you look at all of the the um legal cases that have occurred uh in the last 10 years around utilization of of AI it's around biases people are impact Ed and they're damaged because they can't get a loan they they can't buy something they can't uh uh get a procedure done uh in healthcare I heard something the insurance companies were leveraging these systems to pick and choose what uh uh what um treatments people could have things like that and when you do that um you need to have some sort of a process that you can document that you're getting through to audit and remove all of the ethical issues that these systems are going to uh that these systems are going to have and so that's core how you deal with bias but you know at the end of the day um your ability to look at what this thing is doing as to whether it's harming people or group of people is what you need to ask things like making sure you have consent um with uh with people who are going to be producing in the network your ability to have um uh copyright uh issues with people I noticed that in llms that I'm using my work comes out a lot of times when I ask questions um so making sure you have the legal uh roles and legal um documentation and Concepts and consent for all the information you're going to leverage as well so lots of Dimensions kind of come into it but at the end of the day are you harming people are you uh going to get yourself in trouble with the law uh and are you doing the right things which you can explain to your stakeholders and I think that's what we need to kind of move through so the considerations are really that so people get into these bias Frameworks and things and bias Checkers and things like that that's really not those things are easy to deploy uh you need to have an ethical framework an ethical set of guidelines that people can follow so we live up to some sort of an ethical standard that's why I love having an ethic specialist on the team and the reason why he or she is going to be responsible for ensuring and auditing and making sure that things we do are things that we should do it's not that we can do it we can do lots of stuff we can do deep fakes and we can put out uh you know produce things that are approaching fraudulent marketing uh marketing campaigns things like that but your ability to say should we do it in other words we all can should we and I think that should answer is really where the biased and ethical standards lie great answer with that David thank you for that David could you explain the concept of latent space in generative models yeah latent Space is really kind of a work area so um you know it's low dimensional space that uh you know we're able to put features and variations in it's able to leverage something that's able to put it into a a a space for a certain amount of time as we process the system so it's a it's probably best people Define it differently but it's best defined as as that so the generative a model is going to put stuff out there is a temporary storage area and it's going to process information and handle the information within the space and it's it's uh something I think as an architect you don't have to deal with because it's something that the uh tool set into itself is going to do do on your behalf but it's good to know that it's there and it's good if you're talking to a vendor to ask them how they're managing Laten space and whether it's persisted or not and you know how they use it how it's effect on the processor and things like that thank you for that David have you implemented conditional generative models and if so what techniques did you use for conditioning yeah they're models that generate specific data um based on given condition so um you know conditional vae which is variational auto encode encoder you know that generate image from handwritten systems would be an example of that so in other words we're able to look at something that a request that's coming into a knowledge model and it's able to generate the image based on what it knows and how to generate the image in other words I'm telling it to generat um handwriting that spells out the name mic so it has to know a few things in other words it has to know how to write mic uh from a veral cue to a to an uh to a written queue but also knows needs to know how to write it out in handwriting so it's able to generate images uh you know based on desired digital uh label and it's able to learn to produce images that match the provided conditions allow the targeted data generation so what we're saying is um write mic and it knows how to spell mic and it also knows how to write it out in handwriting and so in looking at that the um where it's you know looking at the conditional you know classifier attention mechanisms to make sure that's done in a proper way um again good news is nor normally Architects don't have to deal with that it's good to know that it's there and the implementations that are there all of the models and all the Technologies out there are going to do this in a different way so um it pays to you it pays for you to understand the conceptual understanding of what these things are but not necessarily how it's carried out in detail because that's going to be the domain of the particular tool that you leverage absolutely now I know you partially answered something like this earlier but I'm going to dive a little deeper how do you train a a generative model effectively that's going to have either limited data or noisy data yeah the the best thing is to you know do things like transfer learning um you the ability to augment the data and even providing some semi-supervised systems in there as well so in other words the great thing about generative AI systems is that they they deal with information learning through unsupervised techniques and we don't have to you know tell it what to do and so we can that's why we can take take these llms and point it at the open internet and just go out there and read everything out there and discern it through patterns and come up with conclusions and become very smart at the information out there um however if we have noisy data or some of the data may not be there it need to be backfilled we have to have some sort of assim supervised capability where we're backfilling the data we know if the ZIP code is missing that we're going to have to take the city and state look it up figure out what the ZIP code is and put it in the data before it's consumed in the in the generative AI system uh the ability to make sure that it's spelling things correctly the ability to make sure that uh you know we don't have a transaction that's wildly out of uh out of gauge you know we're not charging someone $10,000 for a safety pen things like that so your ability to put these kinds of conditions and these kinds of monitoring systems around the consumption of the information is is really how you beat these this noisy data uh that's one way to do it another way to do it would be to go back and fix the data um which which is my preferred way of doing it but many businesses can't do that or many uh instances they tell me that that's somebody else's data we can't fix it so you need to figure out how to do it on the Fly well that's a way to do it on the fly but you should go fix the data that's your best CH that's your best bet sounds great David and thank you yeah the better the source of the information the better the results will be right so what strategies do you use to ensure stability and convergence and training generative models yeah things like normalization you know understand the progressive growth in the model how that's going to be managed and your ability to deal with adapted learning rate schedules in other words when those things are going to be pushed to the model and how often that's going to happen so you have to remember that these things um basically function on external stimulus that occurs on them so we have to look at the stability of the models based on what needs to be changed the model that's why we put like things like adaptive learning diversity promoting loss function preventing out you know overfitting stabilizing training processes getting into the models Rel liability to learning uh so there's any number of ways to do that but at the end of the day what you're doing is you're trying to make sure that the model can learn in a repeating rate we have a way in which the model can generate outputs in repeating rates in other words if we ask it a question it's going to generate inferences based on that model so this is going to be that this particular model is able to learn in any number of ways and we have stability and the ability to train the model any number of ways and we've assured that the stability and reliability is there and how it's going to be trained uh again good news bad news um good news is you need to understand what that is and how that works you uring stability convergence training of these models the um I'm sorry that's the bad news the good news is that we have tools and Technologies to make this happen so uh most of the technology that I deal with in generative AI this is an automated process that happens behind the scenes it's good to know what that process is very much like you're dealing with a database you know how to add and delete data you know how to deal with data normalization any relationship diagramming and and and uh and and foreign keys and primary Keys things like that and here it's looking at the particular way in which the generative model needs to consume information and making sure we're tuning to make sure that's going to be something that's going that's going to work every time and it's going to provide them with uh some core capabilities to make sure that it's able to learn at scale I guess is the best way to put it thank you well explained David if you could discuss the tradeoffs between different generator models such as Gans or versus vaes yeah in other words um the adversarial network uh generative adversarial Network it's able to prioritize sample quality and so in other words it's it's again the robots that are working across each other in the conference room and their ability to throw something up like we're going to evaluate this piece of data and they say I have an opinion of this piece of data here and I have an opinion of this piece of data they match it moves on I have this opinion I have another opinion and it doesn't match then we have to come up with some sort of an arrangement in terms of what what's the best outcome that's why these things are very good at producing information that is um is right most of the time based on the training data that it has is because we're going through this kind of scrutiny the data is in essence always questioning itself and improving itself moving forward so um and of course there's VA VES they emphasize encoding and decoding variations uh and more most of the stuff in generative AI is going to be the adversarial networking stuff which is able to deal with sample quality because we want the quality of information uh to come up so and again these things are going to be handled by whatever tool set that you're leveraging so whe whether you're uh open- Source uh AI tool sets um something that's in a cloud provider things like that uh all of these things are going to be built in the particular tool that you're leveraging so you're not even going to know what the mechanism behind the scenes to make it happen but these are good questions to ask the vendors of the generative AI Solutions in terms of how they handle that they should be able to explain it in other words how does your gan gan work relative to the other Gans in the marketplace um and you know what does your vae do uh in terms of encoding and decoding systems how does it run through the process all those sorts of things so when you look at something like this in terms of architecture it's good to understand that they're there I kind of look at this as textbook learning in other words you have to understand it to have a discussion but you don't want to lay too much in this sort of a de detailed system because it can get you down in the weeds and lots of the stuff so what you need to know that this is handled by the technology You' be able to ask the questions in terms of how it's handled and that gets you to the productive instate or able to look at the different generative AI models and use whatever one are going to be most applicable for for our problem space that sounds good good thank you we discussed bias a little bit before but I'm going to ask a little more how would you mitigate biases and generative models especially in say sensitive domains like healthcare or Finance the big thing is diversity of training data in other words they're getting the right data into the uh into the system and even understanding if there's biases within that data as as it's uh as it's being trained so what I would do is understand any bias in the data before it's consumed in the model and if there is bias in the data maybe this is something that we can uh uh change within the data that's being consumed to the system without changing the facts within the system maybe there's an error in there things like that um and if there is an innate bias it's your ability to come up and explain it by the way all biases aren't bad it may have a bias against uh uh you know people who are you know have felony convictions um if you know to to uh you know get background checks for firearms and things like that that's a bias in other words we have a FBI database that's going to tell people you can't own a weapon because you're you're in a um you're a you're a convicted felon however there's biases that are not unintended and erroneous in nature which need to be spotted so the core thing here in mitigating bias is really to externalize any biases that may exist within the data and may exist within the knowledge mod model so we know they exist we know how to manage them and also we can identify them either to remove them or manage their impact and in some cases this is going to be a discussion with the board of directors that we've gone through the knowledge model and we have this particular bias which is innate to the model is this something that's acceptable talk to the lawyers when the parameter something is it going to get us in trouble and then we have to make decisions on how to do it but the big thing is knowing what it is uh in other words not having unexpected consequence where suddenly we realize that it's uh it's doing something it's discriminatory and that is uh not necessarily something that's legal and we end up getting fine so your ability to externalize that information understand that those biases exist and manage those on the behalf of the system is kind of how you mitigate this whole stuff so understanding learning figuring out how it works and then making decisions around how they should be managed sounds great do you have any experience working with text generation using gen of models yeah and they're it's using an RNN or reoccurring neural network um Transformer models uh we already talked about adversarial Network and the ability to do dialogue generation within the system so ultimately when using these models and generating text whether you're generating a song or generating a thank you note or generating a paper um you're looking to train the model to create net new texts based on patterns that it learned from an existing text uh this is why people who ask uh chat GTP to write an article for them aren't necessarily committing plagiarism because it doesn't exist anywhere else it's Unique to them but it's learned that information from past text models that it's consumed so great thing about generative Ai and probably the most powerful thing about the concept in general so it's able to learn from the patterns in lots of different ways and learn from V uh videos and audio and written text and books and stuff like that I know that I know that for sure and its ability to generate text based on the system and it looks at the patterns and basically how it's going to write text what language it's writing in and generate the text in any format so all of those things are you know really kind of foundational to what generative AI systems do again we don't have to know what an RNN is or Gan is for that matter just understand that the text is coming from somewhere we need to understand that as as the architects in other words it could have some IP in there we have to be concerned about but it's understanding the patterns of information and it's generating the output that it thinks we need to see from the particular pattern which is having a dialogue with the chatbot with the generative AI system and it's generating whatever text we need and the format we need and getting to the response that we need at the level of detail that we need wonderful have you explored creative applications Beyond traditional data generation using generative models I think everybody's use an llm to ask it to write a song about themselves which is probably the most narcissistic thing you can do but um it's really that's kind of a foundational greatness in the dinner we're able to uh from the technology we're able to create artistic outcomes based on ask asking a question and that we can personalize and take to uh whatever extreme you want you're you know draw a golden retriever driving a truck you know down a country road okay we'll make that trucks out of the 50s instead of the 70s okay we'll make that truck out of the 20s instead of the 70s and it's able to draw this and create artistic content based on what it thinks that you need well in looking at the applications there your ability to in essence create stimulus on demand for example the recommendation I talked about your ability to write and generate photos the ability to which which are in essence deep fake like and your ability to generate art based on whatever input you have that you're dealing with the human being is incredibly powerful because now we have not only the ability to generate generate custom information to a human being in a way they want to consume it but we're able to generate visual stimulus as well and way they want to see it whether it's a video or a picture things like that so it's able to create the artistic benefits of leveraging generative Ai and put it to use for a particular ular business or the ability to even draw a um uh a maintenance diagram on the Fly for somebody who's maintaining a robot in the system based on the changes that were just made to the software so instead of having to reprint a new manual based on the changes in the system we're able to regenerate not only the text but we're able to generate the images that they need to see how that the diagram and get to the engine so it's amazing the amount of applications are in that space and you're also our ability to generate things that are complet custom and useful to us which is uh the power of generative Ai and wipe everybody's into it in the in the begin with generates what we need absolutely you know you got me a thought maybe I'll have a generate a photo of my cat flying an airplane just you can upload a picture of your cat and it'll do it have to do that so David how would you manage scalability and computational efficiency when Ling when working with large scale generative models great question because you have to remember these things are computationally uh expensive uh both from a they use a lot of processor power they use specialized processors gpus tpus um any number of things and I think that the scalability of these systems the way most people react is to toss money and processors at it so in other words if we're especially certainly if we're leveraging a cloud-based system we're just going to allocate faster processors more processors we're going to leverage them at scale we're going to put more memory in the place more storage we're going to increase the io rate all these sorts of things that's not the way it works and you deal with scalability we're looking to get to the minimum viable systems in other words we're looking to get to the most optimized way to provide the uh performance that we need for the particular use case that we have and provide a technology configuration to provide the performance and so that's a little bit of an architectural dilemma that people normally don't deal with they throw money at it they overr provision they leverage very expensive processors I understand why they're doing that because you're always going to be right in other words you're always going to have something that's going to work and it's going to perform well but it may cost you 10 times what it should cost and that's where we get into trouble as Architects we have to pick the right uh tool for the job the right uh system to uh to set it up so it's looking at performance metrics it's looking at performance profiles it's looking at IO rates it's looking at uh Network saturation rates it's looking at storage utilization it's looking at memory utilization and you know even using things like distri training Frameworks where we're moving things across multiple processors and doing so in some in in a more efficient way and there's ways in which you can use CPUs instead of gpus in doing that which is awesome so therefore gpus cost a lot of money to run they use a lot of power order if we're able to use uh a cluster of CPUs to do the same thing um in some instances on some use cases you can do that it's your ability to have a computational scaling model based on the fact we're leveraging a unique creative use of the framework uh model distillation and Hardware accelerators you know all the tricks in the book and how we're able to make this stuff happen are increasing the cash rate um and again uh moving it to a different platform moving it closer to where the applications that are leveraging it because it's uh has some n network latency on all the scalability tricks that we played with in traditional systems are really kind of applicable here the thing that we need to remember though is that generative AI systems uh are operate extremely efficient and better in some parallel processing situations that's why people love using gpus because that's what gpus do however that comes with a cost and so we got to get out of the game of just throwing money and processors and hardware at this stuff we have to think logically in terms of what is the configuration that we need that's going to provide us with the most optimal way of doing it and that's a harder problem to solve uh and that's something that very few people do when they're building generative AI system but it's something I build into the architecture because I think it's probably the single most important thing for you to do because you're limiting the amount of resources in this case money that you're going to burn to solve this particular problem absolutely and that money can be invested someplace else in the business I often wonder if it comes from those of us that worked with 10 megahertz CPUs that we just throw so much compute power because we almost remembered what it was like when nothing worked because we never had power it's always easy to throw larger processors at it because we have them available and they're not that expensive um but everything adds up fairly quickly very quickly could you uh cite a real world application where generative models have significantly impacted an industry yeah that's a great question uh I think the um ability to generate information processing uh within the marketing uh departments that's kind of an unsung here generative AI so in other words uh we talked about the recommendation engine at the end of the day that's a very sophisticated marketing engine that's able to sell something if you remember markets marketing is your ability to generate demand which is able to generate um uh more sales so quietly uh all the marketing groups are making well use of generative AI systems and I say quietly because it's something I don't think people want to advertise you're not doing a press release that you're building generative AI systems and chat Bots into your core engine but that's something where they're able to determine the viability of the market your ability to determine what people are thinking based on your scanning social media and using that as training data your ability to understand Logistics all the things that generative AI can do but then customized stimulus it's able to produce the most um the most benefit uh to the business so they can generate custom images images they can generate custom text they can generate at uh audio uh to communicate with people in ways in which they're not going to uh they're not going to get uh uh turned off because it's AI um so that Nuance was always missing from marketing and people didn't like don't like spam they don't like uh coold calls they don't like all the thing a lot of the things that the marketing folks move into now it's something that can be strategically directed to provide the best benefit for making it happen and that's changing the game quietly obviously we have supply chain integration uh we have Uber that uses it uses its AI system so we're able to you know get the best car and the best uh the best driver for our particular situation depending on where we are all those things are being tactically deployed right now but as far as something that's systemically changing the industry I think it's quietly the marketing industry okay that's great David what role do attention mechanisms play in generative models and how do they enhance model performance well they facilitate um focus and uh and ultimately put an eye on something that it should be paying attention to like the name name applies in other words if I'm creating my uh uh system that I created to spot fraud in this case I'm focusing on checks people are writing and credit cards and even videos of them at an ATM then I'm focusing on particular things are going to be more meaningful to me than others your ability to look at the signature the check back your ability to look at the uh quality of the paper all these sorts of things are going to be more meaningful to me than other attributes whenin the system so I'm going to have an intention system that's going to put a spotlight on a particular attribute that I think is going to be important and it's important to do that got remember generative AI systems if they're not taught different they they treat all data as equal other words it's equal stimulus that come into the system but just like we do as humans um we want the bank teller to look at the signature to see if there's fraud going on if they're getting a check we can ask the generative AI system to focus on that as well and therefore the attention is paid in all the parameters but specifically that and so that's a good feature to understand under know in terms of how you deal with architecture great Point David how would you ensure the robustness and generalization capability of an generative model across diverse databases well the the main thing is doing cross validation within the system so making sure the output that we're getting is expected and no and if it's not that it's able to train the problem you know out of the system so we're able to do regulations we able to put regulations around the data making sure we're doing adversarial training so asking the questions remember the two robots they're looking at a piece of information and seeing whether it's valid or not but it's any number of things and it's really kind of a data testing methodology at the end of the day we're taking information that's being that's flowing out of these particular systems and we're looking at the validation the ability to validate the information is being correct and we've been doing this for years built large databases and they obviously produce information and business information when we first use these systems we put it through a testing cycle we're going to validate that the information coming out of it is correct and of good State and same things work here we just have uh more tools and Technology able to throw at this that are generative AI specific even that leverage generative AI themselves the good news with generative AI systems if they do make these mistakes they're able to spot the mistakes and they're able to uh in essence feed that back into the reiteration of the model and the Improvement of the model um so we're leveraging tools and Technology to make this happen it's good to know that we have these processes in place and how we're doing the testing mechanism from an architectural standpoint but again the good news is normally these tools have mechanisms that are bound to them that do them on your behalf thanks so much for that so couple more questions for you could you explain the concept of style transfer in generative AI models and its applications yeah it's like using Artistic Styles of one picture to change the look of another picture um and you know I have a you know example of this where you know I would s uh something to write a paragraph but please use the writing style of David lyam and so I have enough things out there it's able to figure out what my style is um and rewrite the system so it sounds like something that's coming from me or sounds or sounds like it's coming a simulation of what they think I'm going to sound like sometimes it's a bit insulting but the idea is that we're able to fine-tune these models now not only generate something we're looking for but look at a particular um style and iterate that style in a new instance of it so we can say like you know draw a picture of me based on this picture I just uploaded but uh make me look as much like the Mona lease as I can it'll generate that I've done that before um I don't know why people would want to do that but you can but there's obviously you know different ways of doing it if people have a uh uh a style that they use for their company where they have a format that they want to use so everything looks very similar they can reiterate that style um they can do any number of things which are going to carry certain patterns from one the next so the the great thing there is we're not just having it do something static but we're having it look at an existing piece of work and denote the patterns within that piece of work which are going to be more fine grained and then restating those fine grained patterns in some different use case and how we're using it like David monisa there you go so what methods or techniques would you use for hyperparameter tuning in generative models well ultimately it's just finding the best setting for the model it's going to be tuned for the best performance and the best outcome and so we look things like learning rates and batch sizes um and make sure that the output is occurring in a way at a speed that we're that's going to be useful to us things like that and so generative AI systems and models at the end of the day are going to have tunables within them another words thing were able to configure to make them operate better focus on particular things that we need to do faster and better uh focus on um uh uh running things uh at a rate that's not going to necessarily cause errors but it's able to produce and generate the performance that we have and so this is basically dealing with tunables uh very much like a database and so they're all going to be different by the way there's no one approach to tunables with all these various tool sets that are out there but learn what those configurations are learn the best practices for setting these configurations and you're able to tune the best performance out of these systems and that's the idea behind them so we're tuning it for its particular use in other words they're all going to do things differently they're all going to set the performance pars differently they're going to deal with iio differently we're setting up that system very much like we tune our automobile to be optimized for the particular configuration and use case that it's doing wonderful David I just have one final question for you and then after that a few things we can talk about how do you stay updated with the latest advancements and Trends in generative AI that's a great question um I just re uh read as much as I can uh so I have a uh uh Google alert that's set up uh for generative Ai and generative AI architecture and generative Ai and cloud and I get Oodles of information obviously coming back these days but I'm able to sus through it and see which things are um you know probably more advantageous to me understanding and so that's how I stay stay up on it watching YouTube videos sometimes and certainly taking training courses uh and reading books would be the way that I do it but obviously this stuff is so new now that you're going to find the best uh opportunities for learning based on content that other people are generating so I'm always willing to get ideas from other people and how they're thinking about this technology and new ways of doing it I'm looking at what the vendors are doing and how they're taking the technology normally people's thought processes go beyond it and may attend conferences you know pretend uh participate in online communities on LinkedIn or Reddit uh I have a number of ways that I do that where I just get a diverse array of information that's coming from different perspectives and I'm able to take this array of information and boil it down into what are substantial in terms of where the industry is going and that's the game how it's played right now now unfortunately in a couple of years we're going to have more books to read They're going to be more thought out we're going to have more videos and more training on how this stuff works that we should certainly participate in but right now it's a bit of uh the wild west and US understanding uh ocean of information some of which is not going to be helpful at all sussing through it and figuring out what are the important bits and know David I agree with you it reminds me of when we built out the internet or the main components of the internet for the Enterprise and there was no documentation anywhere and nobody really knew how to get there so David I really want to thank you for sharing your expertise and your knowledge on how to interview for a generative a architect role we went through some technical questions very Tech technical questions honestly for a generative AI architect you got into some engineering questions there a little bit as well which I think is good because no matter what it gives our Architects for generative Ai and our generative AI Engineers even think people things to think about not just to pass this in in an interview but also to think as they're doing getting their generative AI training or their AI architect training do I really understand this can I put it into perspective can I explain it to an employer simply that the employer can understand that I'm competent because at the end of the day that's what really matters in getting hired for all of you that are here we have a generative AI architect program with David lithicum is the primary author he's come up with all the wonderful generative AI architect Concepts and it's terrific L if you and our component to this is and is go Cloud careers is all those other architect things those business acents those leadership those sales skills that executive presence that emotional intelligence those presentations skills those negotiation skills stakeholder management skills all the things that are critical to the architect career we've got some free resources to help you in your architect career in the description of this video we have a how to become a generative B architect we also have some free webinars to learn how to become an AI architect or a cloud architect or other Architects like Enterprise Architects these webinars are free we'll go over the role we'll talk about every Skilling you need to get hired and how to get hired and of course it's on Zoom so you can talk to us free we'll give you lots of information we just want to help you get to your goal if you Eno enjoyed this video please give it a like subscribe to our Channel and hit the Bell to be notified of new videos to assist you in your architect career whether that be a generative AI architect a cloud architect an Enterprise architect or who knows different kinds of Architects coming in the future thank you so much for watching and I look forward to seeing you in another video [Music]

Channel: Go Cloud Architects

Views: 1,332

Keywords: solution architect interview questions, genai interview questions, genai, machine learning interview questions, artificial intelligence questions and answers, ai solutions architect, ai architect career, david linthicum, go cloud architects, go cloud careers

Id: LEi6unJkYKk

Length: 80min 41sec (4841 seconds)

Published: Mon Jun 17 2024