ChatGPT: What executives really need to know | Executive Webinar | Tech Horizons

thank you [Music] [Music] [Applause] [Music] [Music] all right [Music] foreign welcome everyone and thank you for joining us today for the first webinar in our technology Horizon series before we begin I'd like to acknowledge that at the University of Waterloo much of our work takes place on the traditional territory of the neutral anishanabe and Hoda nashoni's people our main campus is situated on the Haldeman tract the land promised to the Six Nations that includes 10 kilometers on each side of the Grand River our active work towards reconciliation takes place across our campuses through research learning teaching and Community Building before we get started a few logistical items that I'd like to cover you may have noticed we are recording today's session if you'd like to watch it again or share it with your colleagues the recording will be made available to all registrants in a few days also we have a lengthy q a scheduled at the end of today's session so if you have questions that come up during the presentation please add them to the question pane to those of you who submitted questions ahead of time thank you so much for sharing those we will do our very best to cover as many questions as we can in the time that we have good morning good afternoon and good evening thank you again for joining us we are thrilled with the overwhelming interest in this session we have over 1800 people registered from across Canada and around the world including Hong Kong Australia India and Germany just to name a few locations I'm Sanjeev Gill associate vice president Innovation and executive director of watt speed here at the University of Waterloo if you're not already familiar with watt speed we are a new unit at the University created specifically to help professionals organizations and Executives keep Pace with disruption through education that's designed and delivered by our faculty and trusted partners these days disruption is often rooted in emerging Technologies like Chachi PT we've created the theme of Technology Horizons as our commitment to Executives and Senior leaders by giving them the insights they need on emerging technology so they can be better informed as they lead their organizations we believe that technology has shifted from being an enabler of strategy to often where technology is the strategy I'd now like to introduce our moderator for today Dr Joelle blitt Joelle is a professor of Economics chair of the Council on Innovation policy at the University of Waterloo and a senior fellow at the center for international governance Innovation or CG his main research interests are the economics of innovation and Innovation policy he studies among other things the diffusion of disruptive Technologies and their impact on productivity and jobs Joelle over to you well thank you Sanji welcome everyone to what I think is going to be a very informative and very exciting webinar so let me quickly go over the agenda so Dr Jimmy Lynn is going to give us a talk for 20-25 minutes and then we're going to move on to a q a session now given the large number of attendees that we have for this webinar unfortunately we're not going to be able to get to all of your questions but what I will try and do is get to at least all of the main themes that you guys are raising and so I will try and post questions to Jimmy on all of these main ideas okay uh so if you have questions make sure you do submit them if you have comments there's also a chat function okay so let me get to uh introducing Dr Jimmy Lynn I can tell you that he's done lots of work both on on the academic side and uh on the applied side I can tell you that he's got main distinctions he also is the co-director of the ai ai Institute here at the University of Waterloo but instead of giving you a full introduction and doing all that work myself I thought I would ask Chachi BT to do it for me and so here's what Chachi BT had to say about Dr Jimmy Lynn so Dr Jamie Lynn is a renowned professor and researcher in the field of information Science and Technology he has made many significant contributions to the advancement of digital libraries data science and natural language processing with over a decade of experience in the field he is highly regarded as a thought leader and expert in these areas currently Dr Lin is a professor at the University of Waterloo in Canada where he continues his work on developing Innovative algorithms and methods for processing and analyzing large-scale data his contributions to the field of earnham recognition respect from his peers and his research continues to shape the direction of the field and push the boundaries of what is possible Jimmy that's really high praise from the most famous algorithm all right thanks Joel for the kind introduction but actually I should be thinking chat GPT so the intro is spot on in many places but inaccurate in others but it's behaving exactly as I expected which I'll get into my talk all right let's get started so in the same way the Joel introduced me I'm going to introduce chat GPT so I asked it what it was and this is what it told me so we all know many of these things already it's a language model it's based on the transform of Technology it's developed by open AI it's trained on a massive data set of text from the internet it responds to questions and generates text and I will be discussing all of these elements so I very much value your time I'll start at the end with a key takeaway messages okay so you came to hear about chat GPT but this talk is about a lot more it's about chat GPT it's about bar it's about large language models and more generally it's about AI right chat GPT is really just the tip of the spear and throughout this talk I'll be using chat gbt basically as a shorthand for all these things barred large language models and most importantly of all AI so this is the executive summary so I think tasks involving the routine and the predictable they're going to be automated by AI on the other hand tasks that demand creativity and require synthesis they're going to be enhanced by AI productivity will increase in both cases but I'm not so sure about this I'm not sure that employment will increase at least in the current roles at the end of the day Enterprises will be forced to climb the value chain but you know what that's nothing new everything will just happen at an accelerated pace right so yes there will be disruption but you know as I say the more things change the more they will stay the same and that's basically the talk so if you have a busy afternoon feel free to log off the rest of the talk is essentially a footnote oh oh wait you're still here okay I guess that means I gotta give the rest of my talk all right so here's the agenda what's chat GPT and how does it work I'll talk about the obvious issues I'll talk about how Enterprises can get the most out of it and then conclude with everybody what everybody came to hear how is it going to impact the Enterprise all right let's chat GPT well at the core chat gtt is a deep neural network specifically using a specific type of architecture known as Transformers you feed into Chachi PT a natural language prompt and you get out some helpful natural language response right what do you feed into the prompt well you can ask Chachi PT to do brainstorming to do generation of text to rewrite text to classify text to extract uh information from it all right so the first thing to understand is that although you just heard about chat GPT similar models have been around for a long time yes chat GPT was released by open AI in 2022 but before that it was based on something called instruct gbt which was based on something called gpg3 which was based on as you guessed it something called gpt2 and gpt1 all the way botting out bottoming out in something known as the Transformer architecture dating back to 2017. now Transformers are actually an innovation of Google right and so these are some of the models that you may have heard of or may not have heard of coming out of Google and not left not to be left by the wayside meta has its own series of models also Right Moving On how does it work well there are two tricks and I'll tell you about both right the first trick has a fancy name Auto regressive language modeling but it's just a fancy scientific term for saying guess the next word so here's how it works you take the Deep neural network the Transformer architecture and then you give it a bunch of text so in this case you say Ticketmaster recently angered fans of Taylor and you want the model to predict Swift and not Hicks right so you get a lot of text you do this trillions of times and lo and behold you train the model to generate some reasonable sounding text well now you might be asking well where's all this text coming from I'll get back to that in just a couple slides for now let's move on to trick number two the fancy term for this is reinforcement learning with human feedback but it's just a fancy way of saying you know ask a human if it's good so you feed chat GPT or any deep neural network some type of prompt in this case you ask you know what's going to happen if you fire a cannonball add a pumpkin you get back a couple responses the top one thinks that the the pumpkin is a magnet the bottom one is more accurate saying the you know the pumpkin's probably gonna be obliterated and so you get a human to say add you know the first one is bad but the second one is good and so like this you train the model to generate helpful text now a natural question is well where are the humans coming from hold that thought I'll get right back to that okay so that's uh summarize these two tricks and I want to answer the question of what do you really need to get these tricks to work right the important feature of the first trick is that it's completely unsupervised it happens automatically you just need to spend a lot of money and a lot of compute resources mostly gpus in the cloud and you'll train the model using trick number one now the question where's the data coming from well chatgpt already gave you the answer it's information that's harvested and crawled from the web and so all these organizations are essentially crawling the web and basically shoving it into the models right let's move on to trick number two it's slightly different here because it requires human feedback so not only do you need some resources devoted to compute but you also need human labor the narrowed very natural question you have is well where are the humans coming from as it turns out this was recently revealed in a new recent time article and uh according to this article these opinions are coming from humans in Kenya that are being paid roughly two dollars an hour now this raises lots of issues about exploitation and the equivalent of digital sweatshops that I'm not going to be able to get into today but it's a very interesting discussion that we need to all have right that mean move on now that I've told you about how chat GPT works and the two tricks I'm going to talk about the issues and actually the issues will become obvious right so the first issue is hallucination which is just a way of saying it makes stuff up all right so if you take chat GPT and feed it a prompt and look at the output it kind of has this like Ransom note quality right well well duh that's just an artifact of trick number one if you play this game of guess the next word this is the type of output that you're going to get right so hallucinations is uh making stuff up but really it's just another way of saying writing fiction right so um it's also the Hallmark of creativity right so if we want chat GPT to help us brainstorm for example we'll need to preserve some of this so here's an example of chat gbt being very creative in my opinion all right so the problem is you got peanut butter stuck in a vcr and you want chat GPT to help you solve it but in the style of a Biblical verse and reading this every time I can't help but smile so this is good let's move on to example of the bad right you've no doubt seen lots of examples of this this is an example of the new Bing interface for search recommending you a phone and you can see here uh based on this public screenshot on the web that it's just riddled full of inaccurate information it's wrong about the price it's wrong about the quality of the um the the camera it's wrong about the capacity of the batteries it's just wrong this is bad so we've covered the good the bad now let's move on to the ugly so that's the second major problem toxicity right so I'm using this basically as a coverall term for generating bias content generating toxic output generating dangerous falsehoods misinformation Etc okay so why do these models have this problem well you already know the answer right ask yourselves where is this model coming from where is all the data coming from it's coming from the web and we all know that the web can be a toxic Cesspool right so if you play this guess the next word game with toxic information it's easy to see how chat GPT can basically just regurgitate this toxic content right so here's an example chat GPT will happily write you an essay about the health benefits of eating crushed glass expanding on the the benefits of the mineral content of silica here's an even uglier example right it's so bad that I'm not going to read it to you and we had to cover up some of the words but you can imagine what's under those black boxes all right this is just really really ugly all right where do we go from here in front of the audience I am going to put my professional reputation on the line and make some predictions so with respect to the first problem I think it's a technical problem and I think it will be solved relatively soon with respect to the second problem I think fundamentally it's a social problem and it won't be solved ever but in the near future I think under most normal use cases neither will be a much of an issue okay let me expound on this and go into a little bit more detail right the first is a technical problem of uh hallucinations and making stuff up all right so there's a key promising technique known as a retrieval augmentation and I'll tell you about how it works right so the key is not to just ask GPT the question like the efficacy of booster it's going to make stuff up and it's probably gonna make stuff up that's full of misinformation instead what we want to do is something along these lines what we want to do is feed chat GPT a couple reports report number one report number two report number three and ask the model to synthesize and summarize the uh the efficacy of boosters with respect to these models right and so Chachi PT if you do it today will still make stuff up and still give you the wrong information but I'm talking about preserving the Fidelity of the underlying information that's fed into the prompt and I think this is a solvable technical problem now once we solve this problem the answer quality will come down to the input that you feed into chat GPT right if you feed a journal articles from the New England Journal medicine it's probably going to give you something reasonable if you feed it misinformation well you know your mileage may vary right but the key point I want to make here is misinformation is already a problem we need to deal with and so in this respect chat GPT is not going to make it any worse let's move on to the second problem the toxicity problem so in my opinion I don't think this will ever be solved and this comes back to trick number two remember trick number two is ask a human if it's good well who determines what's good I mean philosophers have been debating this problem for literally thousands of years we're not going to solve it in the near future the fundamental problem here is that humans hold opinions that other humans find toxic right but within some bounds the susceptible bounds of discourse we allow this right this is called free speech right so so in this context how can chat GPT do any better I don't think this problem will be solved but I think things will get better I think the scenario is that the output that's going to be generated will converge to social norms you know things like we can all agree on like racism is bad and uh violence against children is bad but I don't think we're fundamentally going to do better than that right so where we go from here I think problem number one will be solved soon and problem number two will never be solved but things will get a lot better now at the end of the day it is always going to be possible to get toxic content out of these models right but here's my perspective on that you know that's a little bit like recording your Racist uncle ranting after a few drinks right but for the most part you know your Racist uncle is probably a normal normal functioning member of society and observer observe social norms right and so I'll think it'll be like that for large language models and so I think in normal use neither of the issues I talked about will be much of a major problem okay let's move on to the next section how do you take advantage of it so the appeal of these large language models is that you can just say what you want in plain English and it will respond to you all right but my caution is don't really believe it because to some extent it's a mirage what do I mean by that so beyond casual use it can be really difficult to get chat GPT to do exactly what you want you almost need to become like a llm Whisperer all right today this is called prompt engineering and there are people hired with this actual job title and here's an example of a job ad soliciting a prompt engineer at very good salary as I might add all right so where we go from here once again I'll stake my professional reputation and make a prediction so I think eventually prompt engineering will evolve to become just another programming language a highly technical skill that's uh non-trivial to learn so in other words prompt Engineers will just become software Developers all right the prompts may be in English but I think that's tremendously deceptive the example I like to give is legalese right it is in English it is a natural language right but it has all these structures and the words have all these different meanings it's really incomprehensible to many non-experts and plus you have to get go to law school to uh to become proficient in this so I think that's going to be the analogy here moving forward all right Next Big Challenge the limitations of training data let's think back to this trick number one remember this right so the natural implication of trick number one is that training data will always be out of date and the second is that the models will never see private Enterprise data so that's a problem well there's an obvious solution the obvious one obvious solution is to further train on private data that's immensely expensive but it will be worthwhile in some domains but I'm going to present uh another category of solutions that I think is more practical and it goes something like this right so you're going to take private data you're going to inject it into the prompt and then you're going to take the entire prompt with its injected private data or more recent data and ship it off to chat GPT and you're going to get the response now I know this is being a little bit vague but um how do you exactly do this well that's probably why you need a prompt engineer okay the next implication and the next potential barrier I want to talk about is usage of chat CPT will require calling an external API so it's basically using AI as a service okay I can just imagine it now there's some CIO out there that's saying oh we'll never use chat gbt as a service because our data is way too sensitive all right or insert any other adjective here you know to this I respond yeah I think I've heard this objection before I think this was the same CIO that said about 10 years ago you know we'll never use the cloud because our data is way too sensitive they're security risks their privacy risks there's all these risks and they're absolutely right yet Banks today at hospitals and government agencies routinely use the cloud so I think there will be lots of possible solutions and this problem will be solved the simple is the notion of uh just like you have a HIPAA compliant Club you'll have a HIPAA compliant API I think this problem will be solved soon okay let's move on to basically the the main event of how is this gonna affect the Enterprise so in fact I've already shared my thoughts right the routine and predictable they're going to be automated by AI things that demand creativity require synthesis they're going to be enhanced productivity will be increased in both cases but not necessarily employment in the current roles so no here I said tasks not jobs okay so let's drill into this a little bit more right the routine and predictable so what will happen to jobs that are dominated by such tasks well the classic example that everybody brings up is call center representative and so when you throw chat GPT into the mix productivity definitely will go up and the conventional wisdom is there are going to be fewer of these jobs and I think that's right that's take another type of task the other tasks that involve creativity and synthesis right so what will happen to jobs dominated by these tasks journalists risk analysts data scientists so that's throw chat GPT into the mix their productivity will go up so the conventional wisdom is that these jobs are safe for now but you know what I'm not quite so sure right it's already begun so here's an example of some headlines saying that BuzzFeed laid off a bunch of journalists and are going to replace them by chat GPT so what do we make of that well at the end of the day I think it boils down to this question is the pie mostly static or growing so what I mean by that okay that's considered the scenario Professional Services the Consulting business let's say right so today you have Consultants that in the near future are going to be augmented by chat GPT okay what's gonna happen then well there's two scenarios under the first scenario you need fewer Consultants to deliver the same work quality and so what's going to happen you're going to need fewer Consultants so lower employment but you're going to increase corporate profits in scenario B you get higher quality generate greater scope generate greater demand and you're going to lead to both higher employment and higher corporate profits now it's not clear to me which scenario is going to actually pan out right that goes back to this question is the pie mostly static or is it actually growing okay now I want to be a little bit more optimistic now and you know I was careful to say in current roles I was talking about employment in current roles right but AI will create new roles that today are unimaginable you know by definition I couldn't tell you what they are right and that's what makes AI so exciting right so yes this is an important question is the current pie our current jobs current sectors are they growing but I am confident that there will be new pies and new opportunities that AI will create so yes there will be disruption but we have a bright future ahead of us however I do want to end with a word of caution so during the Industrial Revolution people left Farms to go work in factories and for the most part the new factory jobs did not require new skills unfortunately this is not going to be the case for the AI revolution in the 21st century as an institution of higher education this is the question that we struggle with every day but this is a problem that requires all of us Academia industry and government to tackle together all right that's all I have and well thank you Jimmy I think that was an absolutely excellent talk very informative uh and also very entertaining I would say uh and and I'm sure that uh you uh on on the computers whether you're an executive whether you are a policy leader or an Engaged citizen I'm sure that you've also found that it was very thought-provoking so we've got lots of questions uh and what I'm going to do is try and sort of uh put them into three main buckets so we'll talk about the technology then we're talking about the impact on business and finally we'll talk about the impact on society and so let's let's get started let's talk about the technology uh so one of the things that you said is that uh Chachi PT has a tremendous ability to retrieve Fox to synthesize facts but also to fabricate them and so Taryn Joseph of Scotiabank asks how can you ensure that responses from chat gbt are factually correct you suggested that one of the things that we can do is we can make sure that into the prompt we also are putting in relevant information that that is truthful but you also said there will be a potential technological fix and so could you expand on that okay great so Taryn thank you very much for the question so what I describe in my talk is a very precise and relatively circumscribed technical problem it's if you put information into the prompt will chat GPT essentially mess it up and generate false information so um I think that's a since it has a precise technical solution we have lots of smart people thinking about it I think it's solvable right so beyond that you can use this as a building block and you can put different things to into it and this is where I sort of go back into the notion of garbage and garbage out it all depends on what you put into these problems now with respect to the other issue you mentioned a misinformation well you know I think this is an underlying societal problem right we live in the age now of alternative facts filter bubble and I think these issues are far beyond chat GPT right these things call for the importance of information literacy in education um teaching everybody to become critical consumers of information you know don't trust what you read online you know there's a lot of uh wrong stuff on Wikipedia right so I I think there's a tendency that when a new technology gets introduced we sort of expect it to fix problems that are Beyond its scope all the problems that are sort of unrelated and have already been there and I I think I think about technology adoption slightly differently I think like to think in terms of well you know is this new technology making things fundamentally worse I I think it's a fair point you're you know some of these falsehoods that it's uh creating is a reflection of society but I'm not going to let you off the hook that easily okay uh one way to think about chat GPT is that it's it's a technology that's fundamentally lowering the the cost of content creation yeah uh and so what we might expect to see is a flood of content uh much of which might end up on the Internet uh and much of which of course if you're just saying might be wrong and so is there risk that we're going to see a tsunami of falsehoods uh and moreover is there a risk that uh you know chatbots by creating the tsunami of falsehoods are going to be sullying their own future training data therefore creating this positive feedback loop of falsehoods yeah so yes I think that's a very very compelling counter argument right the summary is basically uh misinformation is not a new problem but chat gbt essentially makes the generation of misinformation free right however I think the important thing to realize is that it's an arms race right there are already models today that can detect if a piece of text was generated by chat GPT um the uh there's a watermarking technology that's going to be developed to be able to easily tell if a Content was produced by a large language model right and so I think it's going to be uh they're gonna it's an arms race they're going to be moves there's going to be counter moves no doubt chat gbt War the future duration of large language models will evolve to evade these detection measures and around and around we'll go right so the the great example I like to give that's sort of along these lines is the industry of SEO which stands for search engine optimizations right today you can go and pay these companies to place your website higher in a Google results list right they've claim to re-reverse engineer Google's algorithm and will get you a higher placement for a charge all right now Google obviously since they want to preserve the Fidelity of their rankings and wants to defeat that so it's an arms race and around around we go right so yes right now uh I would say that the generation of misinformation these models have the upper hand right but it's a dynamic game it's disruptive all right now but I think there's going to be reactions and counter reactions and pretty soon we'll get into some type of new equilibrium and I don't think the new equilibrium will be very much different from the present state okay very good as an economist I always like to think about equilibria and changes in those real that's great so we also had lots of questions around and you know something that you talked about which is uh that these Technologies struggle with very recent data because of course they're trained on older data uh and uh and this is of course related to Search Because unless you have recent data it's very hard to search so Rahul shagritaya of connectrix asks uh we'll chat GPT in the future be able to gain live information and also do web search yeah okay well thanks rohol for your question I I think you're uh putting me in a tough spot right because everything I can I I say about the future of search you know the product will literally be released tomorrow that might uh that might uh contradict everything I'm gonna say however I can sort of speculate on the future of search at a at a higher level right so traditionally historically users interactions with search engines have roughly been divided into first two phases the first phase is you gather the information and the second phase you kind of make sense of it you synthesize information from it right traditionally um the the first phase has always been the purview of the search engines right you issue your query to Google you get back your two 10 Blue Links and then you have to click through them and then you have to put together your answer the second synthesis phase is what's going on in your head but as we've noticed as search has evolved over the over the decades the the computers are becoming more and more proficient in the second part we see it in question answering sometimes uh the chat the search interface already gives you the answer and form of a snippet we see it in the form of info boxes and other sort of search techniques also and I think with this respect Chachi PT is just a natural evolution right with respect to the live information based on my understanding what uh bing and other search engines are are going to do or if they're not already doing it is going to take live information put it into the prompt and ask chat GPT to generate some type of answers around it right so uh if you ask it about yesterday leaves games for example example it's going to do the web search pull up the article and then synthesize an answer based on that right this is not unlike the prompt engineering and the prompt uh injection that I I talked about in my presentation that's a really creative way to solve that problem that's that's really good uh okay so just to wrap up on the technology front uh we had lots of questions as well on where's this technology going uh and so Dan Fernandez asked are we going to be seeing uh 100 trillion parameter models uh cure kovalenko asked uh as there's more and more power are we also going to see a lot more features coming out of this and so what's what's next like can we expect exponential growth in the part of these models can we expect lots of new features what do you think yeah that's a that's a very interesting question so um I think the number of parameters is a great marketing Ploy and every everything every time I hear about these numbers I I imagine uh Austin power memes with Dr Evil so that's that set that aside but I think in seriousness um the the future lie lies in multimodal models right so currently chat GPT is mostly a text-based model right so it it it it manipulates textual information but today we already have models for different types of modalities they're models for speech they're models for uh for image generation in terms of a stable diffusion their models for Speed generation and speech understanding so imagine what if we could combine all of these modalities together into a single multimodal model so with that the future of these models you can actually talk to it it'll actually talk back to you you can use it to search for images uh you can follow up follow up and ask questions and engage in conversation about things that are in the image so I think that's a very exciting future well I I think so too that's a very Vivid uh picture of what that future can be I I really like that it's exciting um okay well so so I I think that's those are sort of the big topics I want to tackle on the technology front so maybe that shift gears a little bit and talk about impact on business sure uh and so as we know uh in Canada we have a bit of an innovation productivity Gap our our businesses are not always known for being the fastest adopters of new technology uh and so the question is how could we make it different this time when it comes to chat GPT so how can our how can our Executives make sure that their companies are fully leveraging the powering capabilities of this newest technology and that they're doing so as quickly as possible and so Michael Dawson who is a the chief technology officer at cat Pension Plan he asked a question very much along these lines he says what are the most practical applications of chat ubt that an organization could leverage today hmm well Michael thank you for your question and that's a great question so I think the lowest hanging fruit is using chat GPT to enhance written communication in the Enterprise today all right so what do I mean by this written communication is integral to part two all sorts of processes in the Enterprise right so memos reports emails you know look at uh look at all the all the information that's piling up on SharePoint sites everywhere right and so I I think a immediate use is to streamline the process both the generation and the consumption of written communication okay let me talk about both sides in a little bit more detail right on the generation side this is something that we can all use today you can use chat GPT to take feed it a bit of prose that you've written and ask it to refine it to improve it to increase the fluency to change the tone and it will do a reasonable job right this is actually a perfect use that can be deployed today because you're not just going to blindly use the output of chat GPT right you're going to click the button a couple times and you're going to be like yeah that's right yeah I'm going to pull that together so in this context it really is augmenting your writing capabilities right but sometimes um you just have problems writing like how many of you I'm sure all of you have faced writer's block and this is where another use of chat gpg I think is very very promising as a brainstorming tool right sometimes you just don't uh don't know where to start hey this is where the hallucination powers of chat GPT might actually be useful all right so that's the generation of content side all right on the consumption of content side well um Executives today need to pour through lots of written material imagine memos reports that are sort of trickling up the hierarchy that an executive ultimately needs to digest and from it make decisions right so can we use chat GPT or some type of technology to synthesize all this raw material into high-level executive summaries and talking points you know the answer is actually no for the hallucination problems we talked about today but as I as I've sort of put my reputation on the line I think that's a problem that will be solved relatively soon and and going back I recall we were having a conversation earlier uh just before the beginning of the session you have some experiences using chat GPT for brainstorming right you want to share that yeah yeah I mean absolutely I in case I hadn't yet come out I'm a big fan of chat GPT I've I've been using it maybe five ten times a day and I I really have been using it a lot and and often I use it as you're saying for you know if I have writer's block uh I I'm I'm struggling to write something I I give it some prompts it gives me a couple of paragraphs and I'm able to iterate on it and and come up with something good that truly is my own uh but as you were saying it I find it extremely useful as well for brainstorming and so uh you know recently with my team we're trying to come up with a name for a new center and uh you know we're struggling a little bit so again shot GPT I give it the right prompts it suggested 20 different names for this center now truth be told they're mostly absolutely terrible uh but what we managed to do was iterate on those ideas so it was a really broad set of ideas you know ideas that I I would never have thought of myself but then we were able to iterate on them and narrow it down to things that really made sense so it sparked that that creativity that Innovation was absolutely fantastic okay so maybe let's draw down a little bit and start talking uh right so so at a sort of deeper more specific level uh chat GPT is a general purpose technology and as such it's probably going to be impacting lots of different Industries lots of different business functions lots of different occupations lots of different products and so let's maybe dig into those a little bit and so let me start with the industries and so we received lots of different questions asking how chat gbt is going to impact different Industries including Consulting writing publishing Pharmaceuticals Healthcare uh and government among others and so do you have any thoughts on how specific Industries are going to be impacted well you know that's a great question and I I can't answer that question industry by industry because I just don't know a lot about many of these industries however I can share some high-level thoughts so I think there will be General Trends to Industry specific versions of these models right as I outlined in my talk they depend on these two tricks and there can be they can be vastly expensive to go through the reinforcement learning to go through the pre-training and the guess the next words trick however I think the pie is going to be large enough for certain industries to make this worthwhile Health Care legal are the two that come immediately to mind right and so if we go to this process what do these domain specific language models buy you well they they buy you sort of customization to a particular domain so through this process of the going through trick number one and trick number two it'll come up with domain-specific terminology it'll also learn to distinguish the senses in which you're using the uh using the terms so for example when I say python in a computer science sense I'm most of the time referring to the programming language and not snake and so this customization by industry will likely have that effect and make these power of these models even more powerful in the specific verticals okay excellent uh we also had lots of questions around business functions uh and so a lot of these questions asked for example how Chachi Piti would impact Marketing sales recruiting Supply change management data analytics and Innovation itself and so for example we had a question from Tom gleeve he asked how might generative AI models and tools be used to support Innovative activities and so do you envision a a future where we can ask chat TPT a technical question here's a technical problem we don't know how to solve it go away and solve it for us and you think it could do it by itself and give us like a detailed solution to this or are we always going to need the engineers you know right in the middle of it yeah so that's yet another very interesting question so you know for all the business functions you met you mentioned I've already heard just reading through the news stories example so for example I I hear it's being used already in marketing in generating copies so so here's a campaign idea chat GPT generate five different variants for it for me and then I'll select the best so that's already being deployed today another example is in recruiting so I've heard of uh hiring managers using chat GPT to summarize people's resumes and on the job Seeker end using chat GPT to refine cover letters refine how you describe your contributions to your previous employers and so on and so forth so these impacts are very very real right at a high level these models are good at certain categories of things synthesizing information and generating information analyzing information so well let me turn the question back to you Joel you know Innovation is something that you study so what do you what thoughts do you have about this yeah yeah I it's it is a really interesting question um you know one way of thinking about Innovation is that simply the recombination of lots of different ideas into you know a new group of ideas that create some value that creates a new thing and you know by that definition who's going to have the advantage there is going to be humans or is it going to be chat GPT that can draw on all sorts of ideas from all sorts of different fields uh clearly you know it's going to have an advantage and so recombining these ideas into a new thing where I think that humans are still gonna have a massive Advantage is in using their judgment to decide which of these recombinations make any sense or at all useful right so the algorithm is going to throw a whole bunch of things at the wall some of them are going to stick and some are not and it's going to be the humans that decide what should stick what is valuable okay so we'll always have jobs I I think we'll always have jobs uh as you were saying earlier they will be different uh and some are gonna you know some folks are gonna benefit others are going to lose but we'll get to that in a second that's actually our next topic uh okay so our next topic is exactly that it's uh how what's gonna be the impact of Chachi BT on society and of course you know one of the things that sort of touches all humans that all humans do is we learn we teach and so I thought I would start with a question on on education on teaching and so Masuma kengura who is the HR executive officer at the Peel District School Board asks how will chat GPT impact pedagogical practices for educators and how can Educators keep up their competency all right hi Masuma thank you very much for your question so as a fellow educator this is obviously something that's near and dear to my heart and something that I I I I think about uh on a daily basis right so so Joel I've been scouring the internet and on social media trying to figure out what people have been doing in in this respect and I think I can categorize the reactions into three broad categories right the first category is perhaps the most conservative these are the professors the Educators that say no you cannot use it if you use it we'll catch you and we'll consider it cheating and will you know run you through the plagiarism or cheating process that's position number one position number two is a more uh you know use it you can use it you cannot use it but if you use it cite it in the same way you would cite a source yeah right and in the third Camp you have the people that have truly embraced it this is like you must use it uh they are professors that have designed uh assignments around chat GPT um uh sample one is uh get chatgpt to make something up and then to fact check it another sample assignment is you know do your best to try to convince chat GPT to generate some bad bad stuff some false have missed some in some information now with respect to fact checking it's actually a really really good exercise in this sense that chat gpe GPT generates text that is so convincing that through the process of fact checking you're actually learning a lot all right so I I think personally I am in the third camp we must embrace it I don't think the first Camp is tenable uh and I think the neutral the middle camp is gonna swing one way or the other moving forward so in my opinion for bidding chat GPT is like saying you can't use a calculator where you can't use spell Checkers where you can use the sore eye right so the final comment I would make and it's something that most Educators do not think about students are using it but you know what you can use it too we can use it also we can use it to generate assignments we can use it to generate exam questions hey we can use it to even comment on student essays right so I it is this something you've had experience with I I have to admit that I have I've used it for pretty much all of those things not always with success but it's yeah I certainly have yeah as a brainstorming to us we've talked about right absolutely it has helped me come up with exam questions and uh you know and then again you iterate on it and and get to something that you like but it allows you more breadth of thinking about the issue yeah yeah for sure uh okay we've also received lots of questions around copyright and what all of this new technology means for copyright and so Francis Ranger asks what are the implications for copyright that we need to be thinking about right now uh as creators and in leveraging AI Tech yes like Photoshop you know Photoshop is a very powerful image manipulation tool right so when I download Photoshop and use it to touch up an image or create new images from scratch I own the copy right so perhaps it'll be something like that I don't know what do you think yeah I uh no I agree with that for sure I I think right now the way it's structured is that copyright can only be given uh to something else created by human ah okay and so de facto uh if it's created by a machine then it can't be uh copyrighted but I think you know you know things are always there's always going to be gray area and I actually think that probably 98 of the time we're going to be in that gray area where it's going to be maybe mostly created by gbt but with some human influence in the process and the question is at that point can it be copyrighted or not and so uh Richard writings asked a question very much along these lines I think he he says who owns the copyright of the final product created by an employee's use of AI yeah I mean that's a great question I I think as you said it'll depend on the relative contributions of the human and the uh and Chachi PT so um it's going to be a gray area and I guess this that the courts will have to eventually sort it out right I I suspect that's exactly what will happen I it I mean I I have to say that this is probably an area where we should be at least a little bit worried so specifically one of my worries is uh imagine that we allow things that are being developed mostly by Chachi BT with just a little bit of human intervention to be copyrighted right and so as we were saying earlier this fundamentally lowers the cost of content creation uh and what that means is that we might have a ton of copyrighted content going out there into the world and effectively colonizing the creative space yeah and if so much of the creative space becomes colonized it's going to become really hard for people to innovate in the future to be able to come up with new creative things yeah we're already seeing that for example with music right there's lots of lawsuits because this new song has a similar beat to some song from 20 years ago and I won't get into specifics but and so that is something I worry about yeah so it's a little bit like a flag planting right we we saw it in the dot com error so where people were essentially flag planting ideas of the internet um but I think the key difference here is that back then flag planting had real costs following patents has a real cost associated with it right but now the game has completely changed this equivalent of flat planting has essentially become free what's the world going to look like then yes yes I think that's exactly it that's exactly the the concern um okay so so you know one other question around uh Society uh I really liked how you ended your talk you talked about that overall these Technologies are going to be good they're going to generate wealth they're going to make us better off uh but there's also gonna be lots of challenges uh and and that it's going to take Academia it's going to take industry it's also going to take uh you know frankly people and government to to make sure that everyone benefits from this uh and so one of these sort of potential challenges is around skills and inequality and so we had a number of questions on this so Vanessa Iverson for example from the Cooperators she asks how can knowledge workers prove their worth uh as AI disruptive thinking and create a space and John herties asks who will benefit and be disadvantaged by the proliferation of Chachi BT and other generative AI so fundamentally like who are going to be the winners and losers yeah so this is a very important question we should think very carefully and deeply about right so I am optimistic about the future benefits of AI it will broadly benefit AI uh AI broadly benefits Society on a whole however not going to sugarcoat it there are going to be winners and there are going to be losers there are the winners are going to be those with the skills that are complementary to AI so exercising good judgment entrepreneur skills presentation skills things like that the losers are going to be tasks that are going to be easily replaced and automated right so although Society overall will be better there will be an uneven distribution of the benefits and it'll be up to us collectively to figure out um that these benefits just don't accrue to the few right so I yeah I I couldn't agree more um and and so we need to make sure that everyone benefits from this and and but you know I think the key question is how yeah and and maybe again as you mentioned earlier there has there's a role for government maybe this is precisely where government comes in and and and and puts in in place the the right policies programs Etc to make sure that all Canadians benefit yes absolutely right so we need to have a broader uh the discussion because there is going to be tremendous disruption yeah right so they're going to be winners and losers and we're going to have to have a very um very detailed discussion about the role of government and uh in caring for its citizens and caring for the welfare of the population and making Investments for the future and everything so everything along those lines yeah I I yeah I I agree I mean I think government's going to have a really important role uh you know one of the areas of course is going to be retraining as people are displaced from their old jobs uh another area frankly is is going to be redistribution uh and the only way that's going to work of course is if we potentially so we may have to reform our taxation system to make sure that the really big winners from this uh that we are able to tax them so we can actually pay for all these things right yeah well you know um I understand that regulation and taxes are sort of the main two levers of government all right but I think we really need to be very careful here of uh unintended consequences right but these are exactly the types of conversations we should be having right around our dinner tables in the boardroom in the classroom in the laboratory around our virtual Zoom water cooler these days right and so I look forward to having these discussions with all of you well Jimmy I have to say I really enjoyed having this discussion with you thank you yeah um and and thank you to all of you for all your amazing questions lots of great questions out there and I'm sorry we couldn't get to all of them you know there there's just there were too many uh so that fundamentally concludes the Q A's uh part of this discussion and I'm gonna pass it therefore back over to Sanji thank you Jimmy and Joelle thank you so much that was a fantastic session at watt speed we're looking to build professional and executive development programming on this topic that explores Ai and related Technologies but we want to hear directly from you are our programs on this topic of interest to you as a leader or to your teams uh so we'll take we'll take a minute to see a poll question on your screen in just a second we'd appreciate if you take a second or two to answer the poll and let us know what your thoughts are on that great thank you very much the next event in our Tech Horizons webinar series is called geopolitical turmoil and its implications for the technology landscape featuring University of Waterloo Professor Dr Besma momani details for that session will appear on your screen shortly and will also send you an email invitation once again thank you so much to everyone for joining us today I hope you have a great day

Channel: WatSPEED - University of Waterloo

Views: 13,920

Keywords:

Id: Dfb-DL36RIk

Length: 61min 8sec (3668 seconds)

Published: Thu Feb 16 2023