Cognizant CTO Babak Hodjat on AI and Contextualizing Data

hi I'm James Maguire and on today's e Peaks we're talking about using artificial intelligence to contextualize your company's data the goal is to gain more insight more Competitive Edge from that data we'll also take a look at where the AI sector is headed in the years ahead pretty interesting stuff to discuss that I'm joined by a major industry expert with me is bobic hoat Chief technical officer of artificial intelligence at cognizant bobic I thrilled to have you with us today uh great to be here thank you for having me you know and and I do like your job title you're not just a CTO you're the CTO of artificial intelligence am I correct in thinking that yeah the only AIS work for me right now well that seems like it seems like a a very 2023 job title so I congratulate you on that oh well um you know I I normally ask companies you know Executives from companies to give sort of a brief you know elevator picture what they do and we can get to that but I think you yourself have had a really interesting career I don't want to make you the rehash the whole thing but is there is there kind of a nutshell thing of some of the stuff you've done before we get going uh I've been in AI since the late 80s uh I have a PhD in AI I've started three companies in AI in a past life I was the main inventor of the natural language technology behind Siri which isn't to say much today Siri isn't the top of the line as far as conversational systems are going um and I started sentient Technologies uh we spun off the first ever AI based hedge fund um it's awesome to be living in this uh time right now with AI being where it is and uh garnering the attention that it is right no there there's there's been an amazing surge in in interest obviously this year ever since cat gppt debuted in November 2022 um all right well then let's that's fascinating and that's that's grounds for a lot of conversation but I want to make sure we we touch on cognizant before we really begin um the I know a lot of people are very aware of cognizant very wellknown company but can you give for those who aren't as aware can you give the the short nutshell what what cognizant does cognizant is a Services Company a very large one we have more than 350,000 uh uh Associates around the world uh we're considered an extension of many of our clients uh and we help them on their mainly on their digitization journey and now that's actually uh our job is to help them on their AI enablement Journey okay great good so I think it's there's an idea I believe you may have written this I read this maybe in the official cognizant press release it's about contextualizing data effectively while aligning with specific business goals and and using AI as a tool in that I want to make sure before we even truly dive in what what what does it mean to contextualize data effectively while aligning with specific business goals let's just Define that idea yeah it's easy to think that a generative AI model like chat GPT can answer any questions and it can uh but it answers the questions without knowing anything about your business not knowing anything about the kpi usually plural kpis that you care about and so bringing that in is engineering and its work and making it uh act and behave responsibly is also a lot of work and so that is what we try to capture with those sentences that we just mentioned how do you orchestrate generative AI notes that are so powerful uh alongside the data that's running through your business uh and capturing the context of the decisions that you make in order to ultimately affect the kpi that you care about well that makes perfect sense well then that that enables me to to ask a follow-up question that that I I'll use one of the key terms of this year large language model every conversation at this point has to use generative Ai and the phrase large language model so here's my question with llm is is is part of that process having the right llm you got to tweak it it's got to be proprietary or or not necessarily the choice of llm is very very important even though all large language models uh are more or less trained on let's say a snapshot of the internet and they're trained uh as a foundational models to be able to answer any question um uh first of all they're fine-tuned to various different domains so that is an important choice also there's these this Duality uh where there's two Trends they're commercial larger currently stateof the-art uh uh models out there like gp4 or Claude and and uh you know others that you can pick but they're commercial and they're hosted by Third parties uh and then there are open- Source models that you can download and entirely run on premise and for example data security is a big consideration as to which way uh you want to go hm well you talking about training the L llms on a snapshot of the internet I I'm also thinking and correct me if I'm wrong there's a growth of or maybe there already exist of large language models that have been trained say for healthcare or for finance and they've really been trained with a lot of maybe Anonymous anonymized Healthcare data so it's a it's a large language model made for a healthcare application is that springing into use or is that already in use so basically what happens is that we do train base the foundational model uh is trained on um everything uh everything language because the larger the Corpus the more uh the more powerful these systems are and then they're fine-tuned and fine-tuning doesn't mean it's learning new things it's basically being biased to answer questions in certain ways so prefer certain answers over the others and that bias pushes them to be able to answer questions uh more from for example a healthcare perspective if we if we do that uh training in that manner um and so that there's there's two steps generally we're talking about picking a foundational model and then kind of skewing it a little bit towards a domain that we prefer and of course if the uh use case is within Healthcare we would prefer to use a language Model A large language model that has been fine-tuned on Healthcare data all right so you've talked about contextualizing data effectively while aligning with specific business obje what what are the challenges to this process that businesses should be aware of like we we want to do that but why why is it going to be hard and how can we overcome that that that challenge um the challenges are usually uh very um the challenges that we typically face with uh with um uh you know bringing in AI models into into um uh into businesses uh it starts with the data itself um we have captured as different Industries have captured the flow of data that run through them uh this data needs prepping up cleaning up there's a mix of unstructured and structured data in there and we want to derive various different information from that so we actually spend a lot of time on that and of course generative AI helps us with that as well uh how do you take and augment data how do you take data that's coming like let's say you're a large company that that has acquired a whole bunch of other companies and the data that's being sourced is coming from different sources so the very same attribute might have you know seven categories here and five categories there and you want to harmonize it that's a big task a lot of work goes into that it's not sexy but it's work that has to be done and so that's part of it it's just getting the data in a form and flow that actually um helps us then derive decisions off of that so we need to be able to predict what's going to happen if we make a certain decision because then that that directs us as to what kind of decision we want to make um we also need to get a sense of how certain or uncertain we are on the recommendations of the system something that's often neglected uh we need to we need the AI system basically to tell us hey I'm in Uncharted Territory right now for this particular decision in this context so I'll give you a recommendation but I'm not as confident in it and so maybe that's where a human oversight is is important uh all the way to designing the ux and designing the interactions uh that allow us to effectively and productively and responsibly make this these decisions so I can't point at a single place where there's a challenge the entire thing is a a challenging process it's not TurnKey for sure right I and it really struck me as I was listening to you speak that um I it doesn't sound like companies can really do this inhouse I mean some I mean there's certain things like we want to you know improve our data and analytics practice so we can follow like data analytics you know five data analytics practices best practices and they can actually get more from their analytics application from listening to you speak it doesn't S like companies would really be able to handle this task inhouse yeah I mean the the company's running a business and right uh the data is flowing through and there is an infrastructure uh that allows it to um continue business and work and we have to Overlay this AI enablement on top of that as this is happening um and to bring in these AI principles so while it might sound uh like these AI models can do everything and when you interact with chat GPT it gives you that impression the fact is that to make them work and effective and productive and responsible takes a lot of engineering so you're very right um it they companies do need help uh and every company's workflow flow and decision flow is different so there's not a one-size fitall or template that we can just bring in and just plop in and and uh be done now there are use cases for generative AI that are TurnKey um and we do provide those as well in fact a lot of the work that we do for companies have to do with things like oh we want to do machine translation or we want to have our coding developers more productive or better quality so we want to bring in Te uh better test cases so provisioning that and have it be observable so the company knows where which uh kind of large language model is being used and how and it's all being logged all of that is also work but it's more TurnKey than you know the specific decision making and strategic kpi that a company has and they want to stand out against their competition I think at the end of the day there's a lot of off-the-shelf and routine stuff we can do with generative AI but you want to stand out from your competition by making use of generative AI effectively interesting well I want to I want to get to Cognizant I first want to cover one issue that you brought up I think it's really fascinating is that I may have heard you or misheard you but I thought at one point you were suggesting that generative Ai and a generative AI application can itself be used to train large language models which of course a large language model is is a foundation of a generative AI application so it's almost like this system can be used to train itself if I if that question makes any sense at all oh yeah I mean um AI designing ai ai training AI is a discipline it is a branch of AI and it does apply to uh generative AI as well uh there are pitfalls you want to be careful uh uh with how you do that and there are some very interesting um mainly research uh uh sort of tracks where we're looking into how you can actually make use of generative AI to distill data that uh is generated by um uh other uh generative AI models into a training set being trained uh for the next version of generative AI model so yes there is there is a track to do that and it's fascinating well actually I guess i' now that I think of it I've heard the idea that you know chat gpt7 will write chat gp8 Etc I mean 14 will we'll write 15 or so I I don't is that that will that I like the way we're skipping over five altogether um uh yeah I mean that is the track I mean the scale at which so first of all uh we're seeing that starting from these simple first principles of you know predicting what's the next token and a string of tokens and training it on a lot of data and scaling it making it larger and larger makes it more powerful um and so if we're on that trajectory of scaling it we have to scale it automat atically some way somehow and so you know utilizing AI to help us with AI does make a lot of sense when you think about it we've been doing this forever with technology like we use technology to create new technology all the time and so it's no different than that in that respect well let's drill down to Cognizant I mean what how how is cognizant and you've talked about it a bit but let's make sure we really cover it how is cognizant addressing the AI and data needs of its clients where it might be some sample use cases of of cognizant right um so I I mentioned a few where uh when we engage with a client um there is this expectation given our interactions with chat GPT for example that it that there are certain things that we can do write out of the box their um lwh hanging fruit uh as far as productivity is concerned especially developer productivity quality of code um uh and and uh you know translation and so forth so there's a bunch of U use cases there uh but where we actually strive in engaging with our clients is to identify the most critical decision making that they have and the kpi that they uh are targeting and work back from there and so what we've done is we've created a number of templates for these types of uh use cases and we have a platform called neuro AI where uh we can make use of these pre existing templates or actually build new systems end to endend that start from data all the way to a decision-making UI that might be augmented with a sort of a chat GPT like interface for different use cases and you name it uh there are many many use cases there uh from you know augmenting a call center operator all the way to coming up with uh you know loan uh criteria and mortgage pricing to actual pricing and uh optimization of pricing um all the way to uh things like um manufacturing supply chain uh you name it there are many many use cases here um some of them have kpi that have to do with for example ESG and uh and climate so you you you bring in uh that uh those aspects as well um it's very fast for us to actually creates these use cases it's amazing my world record is 45 minutes believe it or not where wow we had a lunch meeting with a client and then I had 45 minutes I actually used uh a generative AI model to generate synthetic data uh based on what my understanding was and I also consulted generative AI on what the outline of the data that it expected to be running through the model and then I built an endend system where they could actually query it and make decisions so we often get in front of clients and show them an endtoend use case that uh they typically haven't really thought of uh that that targets their kpi um even in the first meeting and so that's the power of neuro aai the the platform that we've created here it's pretty amazing then the 45 minute figure is pretty stunning it seems like one of the challenges facing cognizant and or the business that might work for is that you're you're helping them with their artificial intelligence Journey but of course that's not a separate Journey from the rest of the business it's so intermingled with the business it's not like a company is hiring someone to do their Trucking and the trucking is handled but it doesn't really address the rest of the business when you're when you're doing what you're doing it sounds like it's it's it's really M making a major decision how they're going to run their business itself the core of the business that's exactly right I mean we look at it holistically um under Ravi our CEO's guidance we Advocate pervasive and responsible use of generative AI so we're coming in saying don't be scared of this technology let's endorse it let's bring it in and um make us your partner in bringing it in and absolutely we want to Target the kpi I mean if we're this is analogous to the transistor Revolution and the PC Revolution which is hugely disruptive the main difference is it's unfolding much faster and the barrier to entry is much lower you know to set up these generative AI based agents and orchestrate them all you need to know is your native tongue just to explain to the generative AI what its personal is and what you expect it to do and so if large businesses do not think about how to disrupt their own business flow others will do that for them this is that l i mean I think that I I I think that's what we're this tsunami of demand that we're seeing on adopting and bringing in generative AI um is is hard to stay in front of I can tell you in indeed and I guess that that brings us to the the trillion dollar question which is the future of artificial intelligence and Enterprise settings you know something you said about the transistor and the PC Revolution you compared AI to that and I makes perfect sense I guess the one thing I would say is that artificial intelligence is the first technology that can grow by itself it can develop by itself without human intervention if you the algorithm is set it can iterate and learn unlike the PC of course only actually goes downhill once it's booted up on that first day um so I I don't know with the future of AI and it its ability to self- evolve what do you see in terms of Enterprise use in the years ahead yeah I I I would caution to say that that that ability that potential is there uh as you mentioned but that's still in research and we're we're still a bit uh far from that completely self-contained and autonomous uh improving system um one one of the main issues is these models are so large and and it takes so long and so much data to actually train them um that unlike I mean we use a lot of terminology here on machine learning and you know zero shot few shot blah blah blah that implies that these things learn as they go but they don't so let's let's not get carried away on that aspect these are fixed pre-trained like the PT and GPT stands for pre-trained um and they remain static so we do need to solve a whole bunch of problem before that aspect um is is completely uh autonomous and maybe maybe some major breakthroughs in in the technology are in order having said this um coming back to your question of what is the future um I I do think that as decision- making become is disrupted by bringing in generative Ai and other forms of AI and decision making into the flow I think we will see better quality decisions and we will see um a a disruption in the way that enterprises are even set up because we set up and we organize around let's face it our human limitations in decision- making uh we're for example not very good at multi-outcome multi-objective decision making and so we set up our organizations so that one unit is only looking at Cost let's say like the finance department and then we have one that's only a cost center uh or is looking at Revenue we have this other like legal department and so forth but if you can actually manage all of those at the same time and find the right balance between them uh in that decision workflow that might call for us rethinking the way we set up our organizations as a whole don't ask me what that would look like I don't know but I I I just think that there is that potential for sure you know I'm glad you clarify the idea of the uh the AI applications not being self-evolving exactly that's that's a it's a future research there it presumably will happen but it's not at the moment they're static I guess I for my own education that's I I I thought there was more self Evolution going on already but not necessarily like I think about an algorithm of course you know the algorithm knows that I shop for a lawnmower so it's going to you know suddenly start feeding me lawnmower ads but of course that's that's part of a static nature I mean so really as of this point the AI that we're using the Enterprise is actually static it's not really this this evolving autonomous thing the the way it keeps up with the data that's coming in in a changing world is by virtue of engineering it alongside models and data that uh uh you know can make it keep up to dat so I'll give you an example if we treat these static models as knowledge workers then we can put them in a framing where we say look here's a search engine go check the search engine for the latest data and then distill an answer for me and that way it's kept up to date if I don't do that and just say answer this question it's not going to know what happened uh since its training so I think it's very important for us to um remember that these are statically trained nodes but in the larger picture we can engineer them to keep up to date as as as they go um and and so we might actually still relying on what as of nine months ago is traditional ml models that can be trained very quickly as the data comes in and optimize very quickly as as the data comes in and then uh allow our generative models these larger language models to make calls to these uh AI models and um get up-to-date information and and respond so that it's very much engineered right now um and these models are not the the model itself is not learning as it goes I I'm kind of relieved to hear that it sounds like we all might have a job for a little while since since they're not learning without us I'm kind of relieved to hear that uh well then the last question then is when is is there a time frame 18 months you know four years when when the actual there is a greater sense of self-evolution within an AI deployment um I I don't know the answer to that uh I think that we the big breakthrough here is that we've a proof of existence we did not think that we could develop an entity that abstracts the world the way humans abstract the world and understands language the way we understand language that's the big breakthrough um to make that now be able to learn as it goes might require a completely different approach so we know that it is possible to get to this end point but can we get to it in a manner that's more incremental and it's training as it goes like us humans seem to do um that might actually take some breakthroughs in the technology so that might happen in a few months it might happen later I can tell you the there there is some uh bad news here which is that AI is now very myopically focused on a very certain architecture trained in a very certain way and scaled in a very certain way which which means that we're kind of not looking elsewhere as much as we should I think and I think there's a disruptive change that is in order for us to be able to um fix this this shortcoming I I don't want to take too much of your time could you do elaborate a little bit on the last point I'm not totally sure I understand that so um AI uh used like machine learning reinforcement learning uh uh was all about you know keep keeping up with a changing environment and changing your world view and being able to operate in that uh currently though AI is a very specific architecture it's a very specific way of training things when we say AI it's all now synonymous to generative Ai and not any generative AI but GPT style generative Ai and so most of the research most of the F you know investment and R&D is going to just that and making that better incrementally whereas I think we need this like we need to bring in uh from the older AI disciplines outside of that uh in order to solve some of these fundamental issues such as um online learning that makes perfect sense bobc I think you said it uh fascinating fascinating stuff it's going to be uh really really interesting to follow the sector in the years ahead thank you so much for sharing your Insight and I hope you come back and talk with us again sometime it's been a pleasure thank you so much

This transcript was generated automatically from the video's captions and may contain errors.