Full transcript of “A Non-Techie’s 10-Minute Guide to Using GenAI”. Click a timestamp to open the video at that moment. Need another video? Use the YouTube transcript generator.
0:00 ETHAN MOLLICK: The thing about using generative AI, like ChatGPT is the first way you try and use it is almost always going to lead you into frustration. So the first thing people do is they treat it like Google. They try and Google something and find out a fact, and actually, it makes up a lot of facts. And it's not a search engine. Very unsatisfying. People often walk away. Maybe, they try and ask the system to have a personality. They'll banter with it. They expect it to be like Amazon's Alexa, and it doesn't have a real personality. And they walk away disappointed. Maybe after that, they start trying something maybe related to their job. Maybe they ask for something really specific. And they get a paragraph that's OK but not very good, and they're kind of like disappointed. Maybe they walk away then. All of these things are actually misleading. They're not the way you actually want to use the system. To use the system, you have to treat it like a person in some ways.
0:45 Like an infinitely helpful intern that also happens to lie sometimes and get confused and really wants to please you. And if you start thinking about it that way, it unlocks a whole realm of possibilities about how to use the system to make your life better. [MUSIC PLAYING] So what can this thing do for you? Let's do a couple of examples. Let's say we want to do some market research. Common kind of thing to do is let's just tell it look up the market for running shoes. Give me a chart of the major players and their advantages and their market sizes. And this is now using Bing. So Bing is a version of ChatGPT-4, which
1:31 is the most advanced model. But it only works that way in creative mode, which you see I've selected here. And the advantage of Bing is it's connected to the internet, so it's actually looking up results. Now, you can see that it's actually creating the entire chart that I might have wanted to have for my market sizing here. And not just that, it's giving sources and information for each of these things. Because the system doesn't actually know what's true and what isn't, you do want to fact check everything that you see. What I would do is check each of these sources, and this is where Bing makes it nice because I can actually click the links and see what the sources are. Fact checking though takes me a lot less time than if I was trying to write this all from scratch. So we're getting a whole report here about the future of shoe brands with all kinds of really neat insights and stuff you can work from. I would never just send this to someone, but as a starting place, it's probably saving me hours of work right now. That's the way to start thinking about using the system.
2:17 Hallucination is a fancy word for the AI makes something up with complete confidence. It doesn't know things. It will tell you stuff, and its goal is to make you happy. So it's going to tell you all kinds of things. It also makes stuff up if you ask it to explain its own reasoning. It doesn't have reasoning. It'll just make up an answer that sounds good based on the answer it gave you. So it's easy to get this thing to lie, and a lot of people do that. And then they're like, oh, this is useless. But if you're not trying to maximize lying, and you learn what these rules are, it's very easy to have it produce mostly accurate answers. You still want to check your results. [MUSIC PLAYING] So there's a mindset you want to have when you're working with these generative AI systems, and that mindset is based around the idea that it's an interaction. You've brought an alien mind to the table. How are we going to interact with that alien mind to get the most results out of it? So let's say we want to make an argument on behalf
3:04 of continuing a hybrid work policy at our workplace, and I'm going to use Bing here. So the first thing I often want to do is get the system to learn some stuff. You want to provide context to it about who it is and what it's trying to accomplish. How do we provide context? So a few different ways. We can tell it who it is, and we can also give it a context of knowledge. So let's first ask it to get some information. Look up why hybrid work is good and the arguments for hybrid work. We're kind of seeding it with context that we want it to understand. Now, the next thing I'm going to do is give it a persona. I'm going to say who it is. So I'm going to say you are the manager at a Fortune 500 company, and you want to write a memo to the senior leadership
3:53 team about why hybrid work is good. So now we're going to see what happens. If I don't like this answer, and it looks like it's going to be OK so far, I would actually just stop the response. Either ask it again or do a clean sweep and start over. One of the ways you're iterating is on the whole conversation. You get to a dead end? Just restart. It's memorizing stuff. It's getting stuck on a path. That path could continue. Now, you can see this actually is a pretty high-quality memo piece it's already doing. Dear senior leaders, I'm writing with you with my own perspective. It's pretty good, but I would be reading this over carefully and saying, OK, are the facts correct? And I may want to make some corrections. I think some more specificity might help. I would say, we are a manufacturer of batteries. Redo it with that information.
4:41 So now I want some specificity. We'll see if it's better. It might be better. It might be worse. The point is it's almost costless for me to do this. While if I sat down to write a memo to senior management, it would take forever. Even if I don't use a single bit of the text here, it still has pushed me to think about topics I wouldn't have thought about otherwise. I'm going to end up being more creative. I'm going to have less problems in the way here. And you'll now notice it's talking about the whole idea-- it's now starting to weave this battery perspective into this. I might want to go one step further. I'm going to stop this here, but I could say, now, tell me the arguments I would likely hear against this point in a meeting and how to answer them.
5:28 And so now we'll go one step further. Let's figure out what the counter arguments are and what that would happen here. So here are some possible arguments, and it's saying-- it's giving the argument and the answer. So now we're prepping for this whole thing. And by the way, I've got a fictional case here, but I could give the names of the people. It would look them up, and then tell me about the individuals involved and who would likely make that argument. Not everything here is going to be right, but we're working through a process of working with this. [MUSIC PLAYING] There's a whole bunch of different kinds of ethical implications. On one hand, it's trained on the entire internet, which is a mix of amazing and sort of cesspool of nonsense. People aren't being paid for the training data. Everything we've ever written is feeding the model. How do you feel about that is an open question, and legally we're not 100% sure of how that all works. The second ethical level up is that the models probably learned various biases about the world
6:14 as a result of all of this. So if you ask it to talk a story about an entrepreneur, it's most likely to talk about a male entrepreneur. And you see those kind of biases sneak in, and they can be very subtle and difficult to detect. The third level is, on top of all of this, companies like OpenAI have worked pretty hard to create guardrails. That's why the system won't do certain things, or sometimes will refuse to do things, because it will not let you make an offensive statement or it'll try to prevent you from doing that. Some people chafe against those restrictions, but that's an attempt to make the thing be ethical. And then there's a final level, which is called alignment, which is the idea that as these AIs get more powerful, are they doing things that help humankind as a whole or hurt humankind as a whole? There's a lot of levels of ethics to deal with. The truth is that when you're using the system, it's hard to separate out what's happening. And you have to make a choice about whether you're
7:00 going to embrace this or not, but there's a lot of complex ethical discussions happening. [MUSIC PLAYING] So let's create a product. Let's do a watch. We want to create a new watch, so let's say, first thing, Bing look up the trends in mechanical watches for 2023 and list some. So it'll do that for us now. So again, we're giving context. Because the process has a lot of randomness, by asking the AI to go step by step-- and you can even just say step one, do this. Step two, do this. Step three, do that. That is the most effective way to prompt. So it's talking about some watches that fit various trends, so we could say, create personas for five different buyers of watches.
7:52 Make a table describing them and what they are looking for. So now you can see we've got Beth, Carl, Dana, and Eric and their budget for watches and their motivations. Let's say, conduct a fake to a focus group with people in the Carol category. And again, we're taking it step by step, and the Carols like this one better. I like that it's a collective set of Carols that are working on this problem. And then we can say, OK, now, describe the watch-- a prototype watch that would make Carol happy and pass that detailed description
8:43 to Dolly, which is the image maker, to show me a picture. We could also say while we're doing this, come up with the proposal for a website-- oh, it's not bad-- selling this watch to Carol. And so these are not bad pictures. And so let's do that here. So now it's going to write a proposal that we might want to give to our website designer. It's creating a whole timepiece thing called Vintage Timepieces, and here's what we're going to do with this. We're going to take this information, which it's giving us, and we're going to paste it into GPT-4 and ask GPT-4 to create us the home page. So we'll just take this here, and I'll copy this over. The context window on Bing, it only
9:28 takes a little bit of space, and so it's not as good for coding. So we don't need the internet anymore, so we will say-- we will ask it to create this website. Create the HTML and CSS. So now it's going to create the website for us, and we can just literally paste this in. And here we go, it's already coding up for us. Nice way of skipping website design. If you don't know how to use this, by the way, just ask it. It'll tell you how to create a web page and how to host it and where to host it and do all that work for you. So all of this is novel, and we've created a product in three or four minutes with a web page and done fake market research on it. It's pretty cool. [MUSIC PLAYING] To be very honest, almost everyone
10:15 I know who has used these systems, you bounce off it until it clicks for you, and then you have a very sleepless night or two. And I don't have an answer for the sleepless night. I think you kind of-- that seems to be what happens. You're like, oh, actually, this isn't a toy. This actually does matter. This is actually going to affect my job or jobs of people I know. And then you're like, what does it mean? And we don't have an answer. So what you do is you double down, and you learn how to use these tools. And then you think about what capabilities does that give me that I didn't have before? What parts of my job that are boring and awful can I offload? What things are tedious that I don't have to do anymore? And that model can get you through that period. [MUSIC PLAYING]
