Alation CSO Aaron Kalb on Generative AI with a ‘Human Touch’

foreign we're talking about generative Ai and how companies can protect themselves against the challenges of generative AI even as they explore the enormous upside of this emerging technology to discuss that I'm joined by Aaron Kalb co-founder and chief strategy officer at elation Aaron thank you so much for joining us today James thank you for having me man I'm glad to be here certainly certainly a Hot Topic this is not generative AI is anyone talking about anything else these days you know it seems like they're not and uh normally when this happens I get very frustrated I was I was so exhausted by all the talk about crypto for the longest time right um finally a topic that I think may actually be worthy of all the all the buzz I guess we'll see what what about cloud computing do you remember remember anyone people when you see people used to talk about this thing called cloud computing yeah I think um what's interesting is the cloud I think at the end of the day to the end user was not necessarily the most relevant it's kind of like oh you know my car has an amazing new engine I'm glad my ride is smoother my gas mileage is better or that I don't need gas at all because it's you know electric but at the end of the day I just want to get where I want to go I think that's interesting AI is it isn't just cios who are excited it's you know everyone and and their kids you know can kind of relate in some way to chat GT or dolly or some of the successful technology so so very true it is chechi BT and generative AI has really captured the interest of you know consumer public and businesses so the the question is of course for businesses the six are a lot higher so the the real question is how can Enterprises ensure accuracy and otherwise guard against some of the problems involved with with generative AI what do you see in this area yeah 100 right I think um a lot of the the applications of generative AI that have seized the public imagination or these kind of fun consumer use cases where you ask chat GPT to you know you know read the Steve Jobs speech as though it was written by Shakespeare or something and he said right right passovers and asking it for life advice which is like sometimes really good and when it's not good it's really funny on that same note I once asked chat GPT what the need of life is and it came up with a surprisingly cogent answer about how individual values and choice and it's like wow that was as good as some of those cheap self-help books I've read well totally and and probably you know uh the the GPT models have read those Chiefs cheap self-help books yes and all the blogs and all the you know everything in the public domain and so in some ways not as surprising as it sometimes seems that it can it can pair it back the best of human wisdom or human uh uh uh cliches appreciate it what have you yeah I think I think that consumer is different as you alluded to is um when it gets it right in the consumer case it's delightful when it gets it wrong in the consumer case it's sort of humorous and you can kind of have a bit of of a a a selection bias that sort of thing and it's okay when you get to an Enterprise context there are certain Enterprise contacts where maybe that's totally fine but many of their mistakes are much higher the cost and a mistake could cause really bad consequences or minimally could undermine the trust of your users or your customers and so I do think a lot of excitement in the Enterprise but a good deal more caution maybe than we see in in the consumer space and I think um I don't think that's a reason to dismiss the technology or to avoid it but I think it is a reason to kind of build it into processes that involve AI uh generative AI other forms of AI and mental Automation and you know manual human involvement by different types of humans by experts by different you know stakeholders to ensure that the eventual output is appropriate to the task at hand well I'm gonna guess it's easier said than done and that I think you know of the many problems one of them of course is that a data account financial information like there's one company who will remain nameless for the person enters some sensitive code into into the a generative AI program and like suddenly that was part of the larger database and that census material was quote unquote given away and like that was a problem so it's a it's a trade it's a slippery slope I mean if if there's a a series of managers out there who are looking to guard against it what what might you say like here's some ways to to push some guardrails in place 100 I think there's two aspects to it one is the particular implementation or deployment that's being used certainly if we're using one of these third-party cloud apis and we're sending data from within our firewall out to it and hoping for the best you know minimally you want to use one of the the services the sort of promises not to integrate the data you send into their training model um um and then you have to hope that if they say that that they're going to hold to it and they're not going to be hacked and we can all read the headlines about all the ways but I think um even if you have your own llm or a super trusted secure one even then there's this risk that yeah maybe the training data isn't leaked but maybe the output is is untrustworthy so I think one key thing to do is to consider the use case so an example that I like many people have thought hey writing isn't the most fun or is you know kind of kind of uh time consuming maybe I can save a bunch of time by having you know GPT 3gpt4 through all my writing for me sure I think the the desire for cost savings um is understandable and actually definitely possible and the Paradigm I'd recommend in that case is what I call sort of the GPT sandwich with human bread right so you might think I have five different ideas for articles and for each one I could write them in six different ways and you could just write 31 sentence prompts and have GPT write you know multiple thousand word articles for each of them right and then you hey rewrite this with punchier pros and we write that more formally or you know all this stuff is much easier to have the computer do a first pass on than they have to do it all yourself but the key is to them is follow it up with that human step where you read every sentence you can't just say hey gpg can you sunrise I'll be yeah I can make sure there aren't hallucinations there aren't inaccuracies you know it's in your voice I think if you do that you can again get that kind of creativity by proxy save a lot of time but if you skip that last step then and get you know a little bit uh you know lazy as it were uh then I think there's a huge risk of something really embarrassing happening or Worse depending on where that text is going as one example really interesting so I mean you I think you may have talked about this though you you've said at one point that generative AI models still need to balance Innovation with human touch so you mean that sort of the chachy peach sandwich idea kind of idea how else can they balance to get my touch because it seems like it's clearly needed yeah I think absence of one case is just making sure that whatever the gender model produces a human um reviews before it goes out and again easier said than done you can see the temptation to say oh I want to have GPT write an email to all of my 100 million users right telling them you know a personalized message about how much we love the way they've been using the product and we want them to come back um you know that might be tempting but I would argue that in that case you're probably better off having your copywriting team make a template and filling in the user's name and an example of a thing they bought recently or did recently and like one of the little thing you know in more of a Mad Lib but it's much more constrained because the creativity of gbt without a human in the loop you know maybe more trouble than it's a work and I want to have more control and have automation is a little bit less random um I think there's other ways humans can be another loop too one of them is just having an intelligent thoughtful human design The Prompt design the interface design the process that's in which the AI is embedded I I see unfortunately some human beings acting like they're robots and just sticking you know gender AI into a place that doesn't belong without thinking about it so in some ways part of being a human is to like Leverage your humanity and and the context you have about the world and how it works and what your users want and having empathy um instead of just blindly uh using it there's a lot of manifestations that were sort of that human balance can can come in um I think the other key point is you know a model is only as good as the data that goes into it right and yes it's true that if you know the internet most things on the internet today were written by humans one day that's not going to be true anymore one day soon probably but for now um but I think it's important to say okay what training data do we put in if we train these models on the most you know you know racist xenophobic illiterate parts of the internet we're going to get output we really don't like and so part of it is that human judgment and putting the right data in and in an Enterprise text we were not talking about the public internet I think finding the right data as input becomes a huge challenge I think very often if you put in stale data or inaccurate data or just the wrong data in any way you're going to get garbage and garbage out and it takes so much higher when the AI is more economy well on that note I mean you referenced earlier about you know what if a company has its own large language model do you think most companies have their own large language model I mean I'm going to guess the answer is no to that but what is your take on that I think it's interesting to see how this develops um one hypothesis that I have is in the same way that AWS and then later Azure and Google Cloud you know platform um you know made sort of you know elastic you know internet hosting and alternative cloud services available to the masses um I think we will see a similar thing emerge where there'll be a not one uh the way they might seem to be today I already that's changing but there'll be a there'll be a handful of large and a long tail of small players that offer you know different generic but customizable fine-tunable you know parametric you know uh uh uh uh settings um for large language models yeah yeah exactly so I think in the same way that everyone has their AWS cluster or their GPC cluster or their Azure cluster everyone's going to have their large language model but only a few players that are very technical and have very particular needs and have the means to do so will actually roll their own and the rest will use a service but I think we will see some uh people doing their own in places like defense and certain places where it's like highly regulatory I think in certain parts of financial services where you need to have a certain level of speed that you can't afford to have the round trip to a third party and I think there will be vendors who will offer um you know on-premise uh models where you know it is more uh an Enterprise's own model than than the cloud case but isn't built from scratch we'll see that a lot more as well I think well it seems like it's going to be a major competitive advantage to to truly roll your own customized bespoke large language model as opposed to those folks that are using kind of a generic one that's been dressed up and tweaked for them it's behind the scenes it's going to give a big advantage to those those custom folks wouldn't it yeah I think there's a mix of learning and unlearning that has to happen so I think if your use case really relies on that apparent um uh linguistic intelligence of like seeming to really understand what a human is saying and produce you know human thoughtful natural sounding response and you know we're focused very much now on the text case there's obviously other applications of gen AI but um in that case I think the bigger the the core language model is the better and so you might want to fine tune it with some proprietary data in your system but something that is trained on as big a swath of of human knowledge and human text as possible it can be an advantage there are other cases though where the breadth of these models is actually a bug not a feature right where you don't want it to have read all the travel blogs and recipes and and you know uh social media accounts you know you want something that um only knows about investing or only knows about retail or what have you and in those cases is actually a much smaller model could not only be cheaper easier to maintain you know what not but actually also more reliable and dependable because it doesn't have as much to hallucinate about so I think it depends on the use case whether folks will want to take a big model and fine tune it the way they'll actually want to build their own you know from scratch as it were or using a one of these toolkits that I think will increasingly be available makes perfect sense all right so let's let's drill down to the role that Elation place as to where this this new world of generative AI it is it is a data hungry world to be sure um how does the Elation move this world forward 100 so that was two-way resolution thinks about about uh gen Ai and AI in general um excuse me so the first thing is um almost every use of data I really ever use of data um basically every use of data is is garbage and garbage out and I think you know for the last several years we've mostly seen people using data for the purposes of analytics and business intelligence and so that's garbage and garbage out but the nice news is there's sort of a human in the loop so if I'm a a a person that I use for example Elation to discover a data set to do some analysis about my customer base or my supply chain or whatever it is and I um well if I'm using Elation I'm going to get the right data because Elation will help you know me understand okay this is the data that's most used here's the data that it's trust if it's use case Etc but in the app specific data catalog your data intelligence tool I might you know ask somebody what data to use I might pick a table that appears to have the right name for my use case and I might be wrong about that data's applicability you know uh quality relevance Etc and if I use bad data I'll get this bad result of my dashboard or this bad Excel file or whatever comes out of it and the hope is that the because I know my business I'll say you know what that doesn't look right you know we don't have that many customers or a supply chain isn't that distributed whatever the the clue is right and so AI is just like bi and if you have bad data going in you have bad decisions bad bad indications coming out but the problem is at its best the human is less in the loop you get that lift when the automation is able to have some autonomy and go quickly and so I think it's the same universe of needing to get trusted data but the stakes are even higher and it's even more scary to imagine it going wrong and so software like relation that again can help you discover relevant data distinguish the good data from the bad data is all the more vital I think the other thing to consider is software like Elation also doesn't just enable AI but also really benefits from AI from the very beginning even AI built into our product using ML on different kinds of AI to you know figure out what data is useful for this case you know automatically figure out what data can be trusted all that sort of thing and I think as we um fold more and more generative AI into our product we're going to see even more of those uh advantages of Automation and intelligence uh be bestowed upon our customers and users so really it's a system that is at least partially powered by AI helping to create artificial intelligence itself I guess that's not a totally new idea I never that had never fully you know I never fully realized that idea of course that's that's the way it has to be there's no way around that I think that's true and I think the one key point is it's a loop a virtuous cycle of AI to AI but with it Junctions you know the right humans in place what we want to do is do you want to eliminate human toil all the copy paste you know transcribe from paper onto the computer like all of those tasks that you know are just you know boring and mundane and don't require judgment and actually in certain of these tasks a human is more error-prone than a computer that's a great place to bring an AI and free humans up but if you jump too far ahead and pay humans out entirely this is fear of a snake eating its own tail if it just says hey hey I grab whatever data looks right to you and then whatever conclusions you want to come to and take whatever action you want to take that's when we get these like Terminator sci-fi terrifying scenarios but manipulation the idea is you know the uh the computer makes the job of data documentation data curation data Discovery easier it doesn't do it all foreign augmented but what what is the human role then in that process is making judgments about what what chunk of data is going to go where what what does the human do that's exactly right so let's suppose I'm an AI engineer or you know I I'm fine-tuning of building my own model um for a company I have to say Okay what data are we going to train on right if this is let's say this model is going to be a chat bot that talks to customers dynamically about their um you know about their customer service issues and then routes to a human only when it's the hardest problems but it's kind of a first line of defense themselves the early you know things quickly so I want this I want this um this bot to kind of have at its metaphorical fingertips what a good customer service rep would have like okay how long has this uh customer been with the company any what were their recent purchases you know it was there a delay is that why they're angry like all that sort of content you might want to feed all that into the model well your database might have tens of thousands of columns with different bits of customer user data and you need to create the right training data set or even just like on a per you know prompt basis if you're prompt engineering or many different applications in an AI engineering context but you want to get the right data to get the right outcome and each one of those decisions um if you don't have a data intelligence tool like Elation it's it's quite probable to put in the wrong data and then produce a bot that causes more harm than good all right it makes perfect sense so the big question is what's going to happen in the future if you look a few years down the road the the future of dinner to Ai and I think that sometimes people forget it is really it's not new totally an Enterprise but it's it's new in the public Consciousness since the end of 2022. so of course it's still hallucinate still with some problems but of course it's it's so new a lot of these problems are going to be solved uh and and then what happens what do you see happening oh a few years in the future of generative Ai and the Enterprise you know but no one can predict this if if you could really predict that you'd be in a beach in Tahiti right now but still I'm asking you anyways that's right exactly I'm hesitant to say it because it is moving so quickly um I'll say one thing that I do know however is um the models will absolutely improve but there's certain barriers I'll say are challenges we're more intelligence alone can't solve it so one reason I'm very bullish about the space elationism is no matter how intelligent you are whether you're artificial intelligent or human intelligent um intelligence alone cannot substitute for knowledge right right I might be able to write the most brilliant algorithm or do the most brilliant analysis in the world but if I'm wrong about the facts I'm going to draw the wrong conclusions you know even even if I'm infinitely intelligent so I think they're always going to be a role for data and thinking about data quality and data governance and data Discovery because no matter how smart the systems get in fact the smarter they get the more vital it is that they nipidate are correct about what's true and what's false because they're going to run wild and be increasingly empowered with whatever they have that being said um I'm happy to speculate a little bit oh good um I think I think we're gonna see um AI what they're going to see is like designers and developers get better and better we're gonna get more like uh AI Legos if you will and aisdks and AI Frameworks there'll be books written about designing with AI and about you know policy and process with AI and so we're gonna see I think instead of this kind of wild west of AI That's dangerous people trust it too much people don't trust it enough I think we're going to see some um you know kind of like with mobile devices with the Internet or whatever we're going to get to see some paradigms that enable us to integrate AI intelligently into all aspects of life and it's going to be really great in some domains I think we're also going to see some weird two steps forward one step back phenomena one thing and the experts who disagree on this but I am actually quite concerned that as the internet um gets filled up with content written by these Bots many much of which looks great some of which is great some of which is actually you know terrible and wrong and working um that we're going to have this sort of snake eating its own tail phenomenon where either you stop training on the new data and then everything you know is stale by one year two years three years or you keep learning new things but a lot of it is is you're kind of reading your own garbage right right so it's actually training so its own bias is being reinforced because it keeps re-reading it that's right and so in the same way something like personal Computing or the internet you know creates all this incredible opportunity but then creates these problems we didn't expect I I think and obviously I don't know what I but I don't know what to expect one thing I do expect is a whole industry about this problem we have today about fake news and trust and whatnot is about to get even worse right industry dedicated to helping you as a consumer decide what parts of the information out there were actually if not written by a human at least verified or validated by a human sure um and by the way not just so that you can trust it as a human but not only our AI is going to be doing so much writing they're going to be doing so much reading I mean be a few years from now no one's going to read their email they're just going to have their intelligent assistant say good morning sir you know good morning Madam you're right here's most important out of the millions of emails that I read most of which were written by other AIS I think in certain areas it is dystopian yeah yeah it used to be like my people would call your people now it's like my bot will call your Bot I think I think so I mean I think of these things that are are you know seem funny or science fiction now I think are right around the corner I mean normally you know you have to say oh what's gonna happen in 50 years to get the really crazy stuff but I think now your question of two three five years and that's enough time for the world to look very different well then there's the other question of um the question about jobs or whether it would take jobs you do a good job apocalypse and you know in earlier is changes like you know from the farm to the fact that they said well they created more jobs in the factory the farm there's no problem but the problem is uh the AI jobs really require training and education and some people that are losing their job now don't have the trading education to move into the new areas I'm worried about the job piece in this picture yeah I have a few thoughts on that I think um I think certainly there's the individual workers human beings perspective they can be quite different from the economists you know aggregate perspective right so maybe certain jobs are are eliminated and certain new jobs are created but an individual person might not be able to get that new job so obviously that disruption that every other human cost that we shouldn't Overlook um I think what's a little bit interesting though is this revolution unlike fire revolutions might actually cut the other way a little bit so I think about like the sewing machine if my job is to stitch and I can do 100 stitches in a day and then a sewing machine you know can do it in in you know a minute suddenly I can change my job just from doing stitches but to actually making clothes that's the good news probably more rewarding more fun job um that said I have to now buy a sewing machine learn how to use it and you know um um I think previously the jobs got more technical you know first it was like I know how to use Microsoft Word I'm cool now it's like you don't know how to use Microsoft Word no yeah right so right I think that R has gotten higher and higher for chemical knowledge what's interesting is that AI even though under the hood it's the highest Tech possible it actually enables you to get computers to do things for you just by like typing in right right so it's low code no code and then that void The Voice interface I see what you're saying right yeah so it's a weird thing where like to be able to make a new AI algorithm might require a PhD in computer science but to get the new AI algorithm to help you be better as a writer as a lawyer as an accountant whatever the job is right actually it may be extremely easy if you're willing to to you know just retrain your brain to not do everything yourself and to instead invoke the assistance of this new power right that is fascinating Aaron I think you said it it's going to be a super interesting sector to follow and who knows what's really going to happen I I like your predictions though uh I hope you come back and talk about this again sometime it's a really really good time thank you thank you so much I certainly don't know what's going to happen but I'm eager to watch with you great

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