Stephen Meurice: Have you tried ChatGPT? If you haven’t I’m sure you’ve at least heard of it. It’s basically an eerily smart chatbot. It uses a form of artificial intelligence called generative AI. And it can do everything from suggest wine pairings to make a vacation itinerary to write a cover letter. You just ask it a question, or give it a prompt, and boom — you get your answer. Seemingly like magic.
Yannick Lallement: ChatGPT is something very special. Because it's an entity that you can converse with. That you can talk to, that you can have a dialogue with.
SM: That’s our resident AI expert, Yannick Lallement.
YL: It's literally the very first time in the history of humanity that we humans can speak to someone or to something that is not human.
SM: Game changing, awe inspiring and maybe a little unsettling? Especially if you’re thinking, “Is this thing gonna take my job?”
YL: So because this thing can converse with us, we tend to ascribe it all the virtues of being human, right? Especially being intelligent, being able to make plans, thinking about how you want to approach a problem and whatnot. But it's none of that.
SM: So then how exactly should we be using this new technology? Yannick is Vice President, Corporate Functions Analytics and Chief Artificial Intelligence Officer at Scotiabank. He’s here today to shed some light on how people are already using generative AI in the workplace, what the risks and blind spots are and where he sees this technology heading in the future. I’m Stephen Meurice and this is Perspectives.
Yannick, thanks for joining us again. Really appreciate you coming.
YL: Thanks for having me.
SM: So we're talking about ChatGPT today. And in preparation for that, I actually downloaded the ChatGPT app on my phone and I asked it to write a short bio of Yannick Lallement. It said, ‘I'm sorry, but as of my last knowledge update in September 2021, I don't have information on Yannick Lallement.’
YL: I'm not famous enough for ChatGPT.
SM: [laughs] We'll get into some of the shortcomings of ChatGPT, but maybe we'll start out with some of the basics. So last time you came on the show, you summarized what AI is for us, using cats of all things to illustrate, which was very helpful. Maybe you can now break down what generative AI in a simple sort of way? How is it different from regular AI? And your answer can involve cats or not. That's up to you.
YL: [laughs] In the past, before generative AI, AI was essentially used for two things either classifying things, meaning identifying things. An example is a cat in a picture, that's a classification. Or to predict things. For example, predict what next video someone might want to look at like YouTube does. Generative AI doesn't do those things, it doesn't classify, it doesn't predict. Instead, it creates things. It can create mostly two types of things. One is pictures. You can describe the picture you want and it's going to create many. And hopefully one of them will be exactly what you're looking for. Or it can generate text. And of course, that's the case of ChatGPT. You can say what you want it to create and it's going to generate a plausible text and you can regenerate the text as many times as you want until you have what you want.
SM: Can you give me a little bit of a primer on ChatGPT itself? I mean, everybody's heard of it by this point, but who makes it? How long has it been around? How does somebody go about using it if they haven't used it already? Is it free?
YL: So ChatGPT came out about one year ago. It's been created by a company called OpenAI. The original goal of OpenAI was to create open source AI models. Since then, they have pivoted to regular commercial models. It's just free for personal use. They have paid versions for enterprise use. What ChatGPT does really is predict text. So, if you ask a question, it's going to predict what can follow that question. And typically what follows a question is an answer, and that's the gist of what it's doing. How it works is, it works by having learned from an unbelievable number of successive word predictions. The people who trained ChatGPT essentially took the entire Internet, and that represents about 300 billion words. Imagine a giant book made of 300 billion words, and that would be the entire collected internet. And ChatGPT was trained to, given a few worlds, predict the next one, 300 billion times. So, for example, if I say, “After running up the stairs, Julia was out of.” And I ask you what the next word that might come after that, you may say, “Breath. After running up the stairs, Julia was out of breath.” And you did that 300 billion times. And in the end, you have this model that has learned lots of facts, that has learned to produce good English and so on. On top of that, there is an extra layer of human training where humans rate which answer they prefer after the model generates five or six answers. And so that's how if you use ChatGPT, you've probably noticed it tends to be a little verbose when it answers. It tends to repeat the questions you've asked when it answers. This is because the humans who decided which answer they preferred, preferred the longer, verbose answers that repeat the question.
SM: So, ChatGPT has inhaled the entire contents of the Internet and with that you can ask it a question, ask it to do something, write a letter, write an email, produce a text of some kind, answer a factual question, and it will instantly spit out that answer.
YL: Yeah, that's pretty much it.
SM: So I imagine the uses that you could put this to are unlimited, almost. From getting it to write a text to your friend, to… could you get it to write a book? Could you say, write me a 300-page book about Napoleon.
YL: So you can, but it's not going to be very good at it. There reason is that it has what is called, technically speaking, a context window that's a few thousand words. And so the next world that is produced is in the context of those last few thousand words. But a book, a novel is much longer than that. And so it will quickly lose track of what happens before those few thousand words. If you read large novels, you know, War and Peace, it's millions of words. But when you read the end of War and Peace, you still have in mind the beginning, right? You still remember what happened from the beginning of the story. ChatGPT does not have that ability to remember, it is only the last few thousand words, anything that came before that is forgotten. And so you can produce short text and that's going to work very well. Novels, it's essentially not going to work.
SM: So what is the thing that it's probably most useful for for a regular person?
YL: It's pretty good at answering general knowledge questions. It has learned more knowledge than any human has in their heads by far. So anything you would do with Google, you know, find information about something, you can do with ChatGPT. And instead of having a list of documents to comb through to find your answers, you will have the answer directly from ChatGPT. There is one big caveat when you use it this way, it's that the answer is not always factual. It is most of the time factual. And when I say most of the time, it's really nine times out of ten, something along those lines. But you always have to take it with a grain of salt because of course when you see the answer, you don't know if it's that one time where it got it wrong. And the reason is the way it's been trained. It's been trained to predict the next word and the next word after that and the next word after that. So it produces text that is plausible in the context of what it has learned, right. But it doesn't have a model of the world. It just has a text model. And so there's many more different texts than there are facts, right? You can say Mars is green, Mars is red, Mars is blue. That's three pieces of text. Only one of them is a fact. And also it's not able to quote any sources, right. It will answer your question without saying where the answer comes from. Whereas Google, it's almost the opposite. Google will give you the references and then you deal with the references and you find the answer in there. So GPT will gives you the answer without any of the references.
SM: I'm sure people have heard all sorts of opinions on ChatGPT by now, good and bad. Maybe we can start by talking about the benefits of using ChatGPT in the workplace. Can you give me some specific examples where generative AI like ChatGPT is currently helping people to do their job?
YL: So it depends on the kind of job. And one thing that we have to remember is that it's a tool. So it's not a person, it's not smart like any human being. It is able to produce text in response to text. And so really it's a very specific tool. That being said, that has lots of uses. We did a survey at Scotiabank recently among the employees that expressed interest in ChatGPT, and we asked how they are using it for work. Roughly one third of the respondents said they were using it for work. And out of the respondents that are using it for work, 86% found it useful. So what do they do with it? They do things like creative writing, so creating social media posts, editing them to make them shorter or longer or more formal, less formal, simpler, punchier, even funnier. All of that, ChatGPT can really help you do. It's not going to create an interesting social media post for you from nothing. You will have to guide it and be very specific in what you want, but it will help you produce it faster, for example. Other things it’s really good at is analyzing and summarizing complex text. So for example, we are a bank, we deal with a lot of regulations, and regulations are typically very long and difficult to understand. So, it's very good at summarizing, extracting the key points. So at least you can know if the text you are looking at is relevant for you or not. And then you can dive into it deeper. It's very good with helping you with Excel formulas. Excel formulas are essentially a programing language, but most of us who do use Excel are not software developers, so we don't necessarily know how to program those formulas so easily. But for ChatGPT, Excel formulas are just a language like English or what have you. So it's very good at translating English into an Excel formula. And then we got a lot of really creative ideas, too. One of my favourites is one of our procurement officers, she deals with specifically AI related products. She uses it to program ChatGPT to simulate being a vendor in a negotiation with her. And so she can have a session with ChatGPT where she explains, you know, ‘This is where I'm at with this vendor. That's the context. That's what happened. What is the vendor going to get back to me with and how can I prepare for that?’ Again, very, very specific. It does not replace anyone. It just helps our procurement officer actually be ready for next time she talks to her vendor.
SM: And that seems to be the undertone of a lot of what you're saying. It's a tool. It's not meant, at least at this point, to replace people in doing their jobs. Is that how you see it?
YL: Yeah, that is very, very much how I see it. ChatGPT is something very special because it's an entity that you can converse with, right? That you can talk to, that you can have a dialogue with. It's literally the very first time in the history of humanity that we humans can speak to someone or to something that is not human, right? So, because this thing can converse with us, we tend to ascribe it all the virtues of being human, right? Especially being intelligent, being able to make plans, foreseeing the future, thinking about how you want to approach a problem and whatnot. But it's none of that. Even though it can have a conversation, it cannot do any of those things. It can just produce text in response to a prompt.
SM: Let's get into the potential downsides of using ChatGPT in the workplace. What are some of the concerns that either you in your role or businesses generally have about introducing ChatGPT in the workplace?
YL: So there are two main concerns. The first one is what we call the third-party risk. ChatGPT is an application that is developed by a third party, by OpenAI. ChatGPT of course is not the only one. Google has similar things, Microsoft does as well. And we need to be sure that anything we type into ChatGPT will not be leaked, right, will not be out in the public at some point, especially, of course, if we want to use confidential data. So for now, the approach we have taken at the Bank is very conservative. You know, we are a bank. We don't want any of our data to risk being leaked. And so we allow using ChatGPT exclusively when we use public data. So, asking for questions in general is perfectly acceptable, but pasting any bank data into it is not going to be acceptable. But again, this is a third-party risk that happens with any software that we use from any vendor. And then on top of that, there are risks that are connected uniquely to that specific technology, to the generative AI technology, and it's the risk of hallucination. If you ask a question. Most of the time the answer will be correct, but not always.
SM: Just so I'm clear, hallucinations are when ChatGPT doesn't know something, so it just makes it up?
YL: Yes, they will always produce text in response to the question. There is just no possible guarantee that the answer is correct.
SM: But it will say it with the utmost confidence.
YL: Yes, and that's very confusing for us humans because we don't act like this with each other. When we are not sure, we tend to say it. So part of it is employee education. We have told everybody that yes, everyone is allowed to use ChatGPT, but essentially there are two guardrails. Number one is we can use exclusively public data and number two is you must vet the answer. Whatever answer ChatGPT gives you, you have to double check.
SM: Okay. Are there any ethical issues with how ChatGPT is trained?
YL: The problem is that the corpus is so large that nobody can review that corpus and see what's in it. Again, it's essentially the entirety of the Internet. OpenAI don't disclose what the corpus is exactly anyway, so nobody can double check. It turned out recently some researchers found that some content from 4Chan was used in the training corpus. So if you don't know what 4Chan is, it's a website you don't want to visit and it probably shouldn't have made it in the training corpus. But that's going to create some dilemmas for ChatGPT when answering specific questions, for example.
SM: And when you say corpus you mean the data set it uses to train, right?
YL: That’s exactly it, yes. Another concern, potentially more pressing, is that any text that has been used by ChatGPT for training was used without the permission of its creator. So, ChatGPT, for example it has ingested the entirety of the New York Times archive that is available online. But The New York Times never consented to that. And of course, it's not just The New York Times. So, this is playing out in courts at the moment. It's not clear which way it will go. But the same concern exists for generative art model because all of the artists whose work has been used to train the models, none of them consented to their work being used this way.
SM: Yeah. So some massive copyright issues that are going to play out eventually.
YL: Yeah and it's not clear which way it will go.
SM: So I mentioned earlier in the example of trying to get ChatGPT to write a bio about you, and it said it didn't have any information up to, whatever it was, September 2021. Is ChatGPT continuing to learn or is it stuck at that point in time where nothing that's happened since then is included in its answers?
YL: It is stuck at that point in time. If they want to bring it up to date, they essentially have to retrain the model one time again from scratch. So maybe that's going to be, you know, the next version of ChatGPT. But the current one is frozen for good.
SM: Okay, so how do you see the future of ChatGPT or generative AI in terms of the world of work?
YL: So, you're asking me to predict the future?
SM: Yes, of course.
[both laugh]
YL: The answer is not clear yet. My personal prediction is that we are going to see a lot of small applications that are useful for various specific jobs. So, for example, you can have QA bots that will answer contact centre agents’ questions. So when they're on the phone with a customer, they don't always have all the answers, especially if the customer is asking for something a little unusual. We also have been looking at many other types of use case. One is this vendor simulation. You can imagine having a mini user interface for procurement agents where they can use that to train and get ready. All those small applications, I think, will get to their own mini piece of software, and each of them will be targeted to a very specific group of employees. That's how I think it's going to play out.
SM: So ten years from now, is ChatGPT or one of its competitors going to be everywhere? Is it going to be part of our daily lives?
YL: I think it will be. It's already starting to be part of our daily lives. It's going to get better and better. It's going to be integrated in more and more applications like the Microsoft Office Suite and so on. So, yes, I have a strong prediction for this one. Let's meet again in ten years.
SM: [laughs] Okay, I’ll put it in my calendar. Until then, Yannick — thanks so much for joining us again today.
YL: Thanks for having me. Pleasure to be here.
SM: I’ve been speaking with Yannick Lallement, Vice President, Corporate Functions Analytics and Chief Artificial Intelligence Officer at Scotiabank. And as we mentioned at the start of the interview, Yannick’s been on the show before with a great explainer about how AI works in general. You can find that in our feed and we'll link to it in the show notes as well. The Perspectives podcast is made by me, Stephen Meurice, as well as Armina Ligaya and our producer Andrew Norton.
