Dec 15 2023
BONUS EP: Tip for Superpowers School Podcast - featuring Mark Pesce
RESHARE! Thanks to Paddy Dhanda for having Mark Pesce on his Podcast Superpowers school - we're happy to be sharing here! 🎧 Follow Superpowers School Podcast on: 👉 Apple 👉 Spotify 👉 YouTube 👉 Newsletter In this exclusive interview, Mark Pesce shares his journey of writing his new book published by the BCS, "Getting Started with ChatGPT and AI Chatbots." He was inspired by the realisation that billions of Windows users would soon need guidance on how to use powerful AI tools following Microsoft’s launch of co-pilot. The book aims to provide "rules of the road" for AI newcomers, avoiding technical jargon. Mark also discusses generative AI tools and the importance of understanding different AI models like Claude and Google Bard. 👉🏽 AI's rapid evolution requires a balance between innovation and ethical regulation. 👉🏽 Understanding various AI models and their uses is crucial for effective application. 👉🏽 Proper prompt engineering can significantly improve AI's performance and output. 👉🏽 While AI presents concerns for privacy and job security, it also offers opportunities for enhancing productivity and focusing on uniquely human skills. 👉🏽 The future of AI should be approached with cautious optimism, focusing on its potential to augment human capabilities. 🎁 You can purchase the book 👉🏽 https://rebrand.ly/3b93tly The book is illustrated by Grant Wright 🎁 You can purchase the book 👉🏽 https://rebrand.ly/3b93tly Share Mark Pesce (Author) Across a more than forty years in technology, Mark Pesce has been deeply involved in some of the major transitions points in the modern history of computing. After prototyping the SecurID card - the first 2FA device - in 1983, Pesce went on to develop firmware for X.25 networks, a forerunner of today’s Internet. At Shiva Corporation he developed software for a series of wide-area networking products praised for their ease of use and reliability. Inspired by Ted Nelson’s hypermedia system, Project Xanadu, and William Gibson’s ‘cyberspace’, Pesce invented core elements of a consumer-priced networked VR system, reducing the cost of sensing an object’s orientation by a thousand-fold with his ‘sourceless orientation sensor’ (US Patent 5526022A). After collaborating with Sega on Virtua VR, Pesce, working with visionary engineer Tony Parisi, blended real-time 3D with the World Wide Web to create the Virtual Reality Modeling Language (VRML). With VRML, Pesce and Parisi laid the foundations for today’s metaverse, culminating with its adoption as MPEG-4 Interactive Profile (ISO/IEC 14496) in 1998. Pesce wrote VRML: Browsing and Building Cyberspace - his first book - in 1995, followed by VRML: Flying through the Web in 1997. In 2000, Ballantine Books published The Playful World: How Technology is Transforming our Imagination. In that book, three children’s toys - the Furby, LEGO Mindstorms and Sony’s Playstation 2 - act as entry points in an exploration of how interactive devices shape a child’s imagination. Appointed in 1997 as Visiting Professor at the University of Southern California’s School of Cinema-Television, Pesce founded the School’s program in Interactive media. In 2003, Pesce moved to Sydney to found the program in New and Emerging Media at the Australian Film Television and Radio School, guiding postgraduates through a transition to digital production, distribution, and promotion. Shortly after arriving in Australia, the Australian Broadcasting Corporation featured Pesce on their long-running hit series The New Inventors. Every Wednesday evening, Pesce celebrated the best Australian inventions and their inventors. A sought-after commentator, he writes a multiple award-winning column for The Register, and another for COSMOS Magazine. Pesce analyzed the impacts of media-sharing and social networks in two books: Hyperpolitics: Power on a Connected Planet (2009), and The Next Billion Seconds (2011). Pesce’s 2021 book, Augmented Reality: Unboxing Tech’s Next Big Thing, critiques the design of augmented reality systems, questioning whether these devices truly serve their users - or simply stream valuable data back to their manufacturers. Pesce holds an appointment as Honorary Associate in the Digital Cultures Program at the University of Sydney. 🎁 You can purchase the book 👉🏽 HERE ⚡️ In each episode, Paddy Dhanda deep dives into a new human Superpower to help you thrive in the age of AI. ★ BUY ME A COFFEE ★ ☕️ If you enjoy the podcast, then you can donate a small amount here as a token of your appreciation: Buy Paddy a Coffee Transcription: [00:01:04] Paddy Dhanda: Isaac Newton once said, what we know is a drop. What we don't know is an ocean. As each day goes by. AI is expanding our job of knowledge exponentially. Just a year ago on November the 30th, 2022. Chat GPT burst onto the scene. Setting the digital world ablaze. Up until this moment. TikTok has been the fastest growing up, which had taken nine months to reach a hundred million users. . But chat GPT surpassed this milestone in just two months. The rise of this technology has sparked debates on data rights, misinformation, and ethical dilemmas. On the flip side, it has democratized knowledge, streamlined global communication and revolutionized productivity in the business world. The AI genie has truly been let out of the bottle. The EU is busy, developing the AI act to try and safeguard the future of humanity. Interview with Mark Pesce: AI Expert and Author [00:02:08] Paddy Dhanda: So for today's episode, I'm joined by Mark Pesce. He speaks exclusively about his new book. Getting started with ChatGPT and AI chat bots: An introduction to generative AI tools. Mark is a technology visionary with over 40 years of experience. He began his career by inventing the secure ID card in 1983. Uh, pioneering step in two factor authentication. He co-created the virtual reality modelling language. Forming the foundation of today's metaverse. He's also an accomplished author and commentator for the register. And cosmos magazine. And for a techie. He's one of the most engaging speakers I've ever had the privilege of speaking with. This episode was made possible by the amazing people at the British computing societies publishing group who have just become an official collaborator of the podcast. So, thank you so much. The Inspiration Behind Mark's New Book [00:03:14] Paddy Dhanda: So Mark, you've written this book and I'm really curious to know what inspired you to write it. [00:03:20] Mark Pesce: So the back story here is that Ian Borthwick, who's the publisher at BCS Books, had reached out to me, I don't know, probably the beginning of April, because I write a column for the Register, and the Register is published out of the UK. It is the oldest and crankiest website for news and IT. It's basically designed to be read by middle aged, cranky IT managers. And it's very... Not so much brutal, but how about honest with its opinions and free with its opinions. And I love being a columnist for it because I get to talk about a bunch of stuff. And he wrote to me and said, Mark, we really enjoy your columns. Should you be interested in doing a book for us? It's like maybe I did a book a couple of years ago. Wasn't necessarily interested in doing another one. It's like, well, maybe a book about AI and like, No. And the reason I said no is because it was all moving too fast. This was sort of April, right? Chat GPT was a couple of months old. There was a lot going on. Nothing was really static. And I was fending him off gently in emails. May came, and I was working in May at a great big client event. I do a lot of public speaking. And I'd just finished the last one of these. And these were overnight events. I would travel, go do the event, come back the next day. And I woke up. And I woke up, it's the 27th of May here in Sydney. And overnight in America, Microsoft had had a huge event. It's called their Develop event. And Satya Nadella had gotten on stage, the CEO of Microsoft, and announced Windows Copilot. And I'm sitting in bed, reading my feeds, as I do before I get out of bed in the morning, literally 6 o'clock in the morning, and I see this. And I find Ian's email, and I'm like, can you take a call? And the reason I did that was because I realized that about a billion people were about to get access to a really powerful AI chatbot, and none of them had been taught how to use it. The Impact of AI and Chatbots [00:05:11] Mark Pesce: And I thought that the best thing that I could do was to at least give people some basic rules of the road. [00:05:18] Paddy Dhanda: And if I think back to this time last year, I mean, as we're recording this episode, GPT is turning one. And if I think over that time, just how the world has changed and what has happened in terms of the development of AI, it's incredible. It's exponential. And initially I was thinking, well, maybe this AI stuff is aimed at a certain demographic. It's going to help. A certain, industry, but when you were thinking about writing this book, like, who are you aiming it at? I want to write a book that's for everyone who could be touching a computer with this in it, because Microsoft last month made the decision to roll it into Windows 10 as well. [00:05:59] Mark Pesce: A billion and a half people are running either Windows 10 or Windows 11 in the world. So that's a lot of people who need to know the rules of the road. And that's like, it doesn't need to be really dense and technical. In fact, there's an argument for not doing that because all of those technical details Are all changing very rapidly right now, but the rules of the road won't change. So if we teach people how to get started right, then we've set them up and then everything can change. You can get new technology, you can get a better chat GPT, or you can use Claude or whatever you might be using as an AI chat bot. You know, the rules of the road, you're going to do things right. [00:06:37] Paddy Dhanda: Got it. So there seems to be this gateway that's opened and the flood is coming. Everyone's going to be impacted in some [00:06:45] Mark Pesce: It's not coming. It's fully here. So chat GPT is about to turn one and depending on how you count it between chat GPT and Google Bard and Microsoft copilot About 2 billion people already have access to what I laughingly call weapons grade AI. And I only half laughingly, because it's really, really powerful AI. And Meta is busily integrating it into Facebook Messenger and Instagram and WhatsApp. And that's another 3 billion people. So, somewhere around 3, 3. 5 billion people on smartphone and on PC. Have access to these technologies as a part of their daily lives. Now, no one had them just a year ago. [00:07:31] Paddy Dhanda: You've actually used. Understanding Generative AI Tools [00:07:33] Paddy Dhanda: chat GPT in the name of the title of the book, and then you've also used the word generative AI tools. What's the thinking behind that? Because as somebody who doesn't know anything about AI, I'm associating chat GPT with AI. Like it feels like that's the only thing out there, but I know in your book, as I was reading it, you said something that really hit home to me, which was don't just trust one, like look across and compare and contrast. And that was a really useful insight because I think for a lot of us we just assume ChatGPT is the one to go for, unless you've got a particular allegiance to maybe Google or, you know, some other organization. So, first of all, why did you choose to include ChatGPT in the title? And then secondly, what do we mean by generative AI? [00:08:20] Mark Pesce: Yeah. So we, we put chat GPT in the title because that tells everyone, because the other thing that's different a year later is everyone knows what. chat GPT is whether or not they've used it and a lot of people have, everyone knows what it is. So rather than just saying AI chat bot and people might go, well, I don't know exactly know what that is. You say chat GPT, people know exactly what you're talking about, but we have basically big four, right? So we have open AI's chat GPT. We have Microsoft co pilot, which is. Basically, ChatGPT, but with Microsoft clothing on it. It's got some differences underneath, but it's basically the same. You have Google Bard, and then you have Claude, which is from a company called Anthropic. So, Anthropic was a company that broke off of OpenAI a couple of years ago. So, you can think of it as very similar, but also completely different. In other words, the way they built their engine is completely different than ChatGPT. When we talk about whether you know whether an AI chatbot is lying to you or not, whether it's making something up, which is an important point that I stress in the book, I always say that one of the best things you can do is if you're asking a question to ChatGPT, go and ask the same question to Claude if you don't trust it. Because they were trained differently. And this goes back to the idea of what is actually underneath the hood, which you asked about what's actually going on here. Because chat GPT is a website, but underneath that website is an engine that's called a language model, all right? And so, there's kind of one, maybe two language models GPT called GPT 4 and GPT 3. Now, without getting into all the technical details, let me tell you how they make a language model. Basically, you feed the entirety of the internet into a computer. And I am not exaggerating. You basically take everything that you can gather online and feed it into a computer, and then you start asking the computer questions, such as, what's the capital of Finland, which is something I make a lot of in the book. The poor Finns are going to wonder why, but it's just a good question to ask. And the first time you ask a language model this question, it's just going to spew some characters at you. So it's... going to be nonsense. And you'll say, no, that's wrong. The capital of Finland is Helsinki. And then you ask the question again, it will still just spew noise at you. And you'll do this hundreds and hundreds and hundreds of time, but eventually Helsinki will start to appear in among all the noise. And eventually it'll just say Helsinki like, okay, good. You've got that one fact. Let's go to the next fact on the list. And you do this for millions and millions and millions and millions and millions of facts. And if that sounds exhausting, it takes computers who do this 18 months at the speed that computers work to train ChatGPT. Probably somewhere around a thousand trillion of those question and answer go around. Something like that was involved in training ChatGPT. takes 18 months. But at the end of it, you can ask it a question. And what it will do is because it's been taught an answer, in a sense, it knows what answers, not so much are right or wrong, but what are the most likely answers to a question. It will be able to now generate the likely response. And this is where we call it generative AI, because what it's doing is it's using everything that it's been trained on to generate its response. That kind of response can be a single word, such as Helsinki, or it can be five paragraphs, a stolen talking about the beautiful nature in Finland. So it just sort of depends on the nature of the question that you put to the chatbot. [00:11:54] Paddy Dhanda: I really love that Finland example in the book actually because it showed just the variation in responses you can get just from a very simple question and was almost like each of these different examples, AIs have a different personality. I think you even talk about it in the book, like the, the bard version is very straight and formal. It doesn't give you too much sort of fluff. It's very much you asked for this. This is what you're going to get. ChatGPT is a little bit more sort of informal. It's like your friend and it's using certain words to. Yeah, yeah [00:12:29] Mark Pesce: very American in that respect, I guess you could say. It's just, it's a little. Perky co pilot is very Microsoft. It's helpful. It's also quite prolix. And it's really funny given that chat GPT and co pilot are kind of working off the same base. You can say what's the capital of Finland to chat GPT. It will say Helsinki. If you ask co pilot, it will give you two to three paragraphs about Helsinki, its population, its location, its natural resources, its biggest businesses, and perhaps some tourist attractions might even throw in a photo. Alright, and it will also give you a whole page of references, which is good, because then you know it's not making things up. One of the things that I point out in the book is that it's almost always a good idea, whenever you're putting a question to an AI chatbot, to follow it with Be Brief. Because you'll generally save yourself and the computer 30 50 percent of its time when it's giving you the same answer, the answer that you're looking for, but it's not padding it out. [00:13:28] Paddy Dhanda: I'm gonna take that advice and use it on humans because my brother in law he can talk for the world and I think I need to Apply that rule to him as well. Sorry Hanj. Oh, i'm just putting it out there so mark in terms of generative AI, you seem like you're very excited by it. And, you know, just from hearing your tone of voice, your enthusiasm for AI, is it a good thing? And can you tell us like, why, why is it going to be so great for the world [00:13:54] Mark Pesce: It is a mixed bag. Let me be very clear on this. It is exciting because there's a lot going on. There's an enormous amount of research. Just sort of another tip for the listeners. I have been doing my best to stay ahead of the papers that are coming out. And it's kind of like sitting in front of a copier that's going crazy, just spitting out new papers all the time. What you can do is you can take the PDF of a paper and you can hand it to a chatbot and say, Here, read this and summarize it for me. And it does it. Excellent job at that. And so I could find papers and I'm like, All right, there's all this math. Okay, I'm not gonna try to get across the math. What are the key points I need to know about why this innovation is really important? And that's really been helpful to me to understand that it's not really just what's going on at open a I or at chat GPT. What's happening is that the entire field of artificial intelligence. And let me remind people just a year ago, artificial intelligence was a joke. Artificial intelligence was you shouting something at Siri or Alexa multiple times and it not understanding what you wanted. That was how low our expectations were for artificial intelligence. a year ago. That is not the case anymore, right? It is completely different. You know, OpenAI released a new interface for the app with ChatGPT just last week that allows you to have a full conversation with it. You pop your AirPods on and you just start talking. It talks back to you. You're fully in conversational flow with it and it never really gets it wrong because it's good enough AI. So, The reason I'm really excited, and you ask people in the field why they're excited, they're like, well, this time it works. They don't have to, they don't have to gaslight us anymore about when they're going to have AI. There's kind of good AI. Is it fantastic AI? No. It's still really buggy. It still does all sorts of things we don't understand. It still makes things up. We get that. But it's still really useful, even as buggy as it is. And I think that's why I find it infectious and interesting. But because it's buggy, and yet people are still convinced, such as Microsoft, that they have to put it absolutely everywhere, that's a really good reason to have a book out there telling people, here are some things you should do and some things you really shouldn't. The Evolution of AI [00:16:07] Paddy Dhanda: You mentioned, a year ago, things were really buggy. So if we rewind back to the early days of AI, could you give us a brief history in terms of where it all started and how we've got to where we have and why the acceleration in this last year? It's just been incredible. [00:16:24] Mark Pesce: So, the term artificial intelligence goes all the way back to 1956. There was a summer workshop at Dartmouth University, seven or eight people, and there's a really famous photo of them. And they got together and sort of said, okay, we understand learning, we're going to teach computers how to learn, and they're going to be learning like human beings five years 10 years max. This is how much hubris they had and also how easy they thought the problem was going to be to solve. In fact, the biggest thing that we've learned about artificial intelligence or from artificial intelligence is that we have a very poor understanding of what human intelligence is and therefore have had great difficulty to endow Computers with anything like intelligence. So that's taken a long period of time. There have been multiple what they call AI winters. my own career, I distinctly remember two of those AI winters where we're coming out of the third AI winter and God, have we come out of the AI winter into an AI spring. Now, the thing that changed was in 2017, researchers at Google invented a technology that they call the Generative, pre trained, transformer. If you take the initials, GPT. Alright? So they, they used this originally to do language translation. If you read the original paper, they were doing translations between German and English. Language translation is hard. It was one of the original problems that the folks who were working on artificial intelligence back in 1956 were trying to solve. Because language is... indistinct, it's vague, there's a lot of idioms, how did one things don't map neatly from one to another. But they built a German to English English to German translator using this new transformer. And it worked really, really well. And So that was one reason why you now have, for instance, simultaneous translation as a feature in tools like Skype or other sorts of programs. It's because there's a transformer in there. It wasn't until OpenAI, who didn't invent any of this, by the way, but got a hold of it and said, we should explore this. They started then building much larger models. So you didn't just feed it. Some English and some German so it could translate it, but you fed it a lot of written information and GPT 2, which is kind of the very first version of that was dumb as a brick, but showed a lot of promise and like, okay, this seems to be working. Let's throw a lot at it. And they spent the next couple of years, sort of the early years of the pandemic, really collecting a lot of data, really training it. And that's where they got GPT 3 from. And GPT 3 was. basically the foundation for them to make something like ChatGPT possible. They improved it a bit between GPT 3 and releasing ChatGPT, but it's basically the same foundation and they had fed it enough. English, or enough data in that sense, right? Enough language. And trained it long enough that it wasn't just about translating one language to another, but it was about being able to give rich, detailed answers to requests. And the thing that shocked them, and has shocked everyone else, I spent a lot of time in the book talking about what we call chain of thought. The Mystery of AI's Problem-Solving Abilities [00:19:38] Mark Pesce: prompts where you lay out a word problem and you say, okay, you know Jane has six apples, gives three to June, buys four at the market. How many does Jane have? And you can show the chatbot how you solve that problem. And then you can say, okay, here's another problem. Bob has three quarters, gives four to Bill and then goes and gets six more. What's the answer? And it turns out that the chatbot can solve that problem. Now, you want to know why a chatbot can solve that [00:20:07] Paddy Dhanda: Yes, please. No one knows. All right. This is the part which is, it's not spooky. It just tells us that we've built something that is so dense with data that we don't actually understand how all of the connections are working inside of it to give it some of the qualities and the capabilities that it has, but the AI said, Oh my goodness, it's doing all these things that we didn't. [00:20:31] Mark Pesce: specifically program into it, it's really quite good. And that was really, I think, when they started to think about building a tool like chat GPT. [00:20:39] Paddy Dhanda: Wow. That's incredible because I heard a story, what was it? Facebook, when they were previously known as Facebook, got two of these bots to talk to one another. And then all of a sudden they kind of went a little bit out of control. They started to invent their own language that humans couldn't understand. And did someone pull the plug or something like that? Did you come across that story? [00:21:00] Mark Pesce: did come across that story , they wrote a paper about it. It wasn't so much that the humans didn't understand it, basically they were making a digital language. And the thing is that language models under the hood. They don't talk in English. They talk in these things called tokens. Alright, so in English, generally a syllable in English, it's a little bit inaccurate, but around a syllable will convert to one single token. And everything that's being stored inside of ChatGPT is in tokens. It's not in English. It gets converted to tokens on the way in and comes out. Unexpected Behavior of Machines [00:21:30] Mark Pesce: As tokens and gets converted back into English. And so I think what had happened is probably these two computers like we don't need this English. We're just gonna talk in tokens and they just started passing tokens back and forth, which is kind of how the Internet works. It was surprising because no one had expected that behavior from those machines. We are entering a period now where machines are showing qualities that we didn't explicitly program into them, which A year ago we would have called bugs, and maybe this year we're calling features. [00:22:00] Paddy Dhanda: That's always a fine line, isn't it? Between a bug and a feature. I have to say. Personal Experience with AI Technology [00:22:04] Paddy Dhanda: At first, I was a little bit hesitant about this whole technology and it seemed fun and it seemed cool, but I was just asking it silly things like make up a song about this or, you know, tell me a great story about this and It's, it's interesting how my habits have changed. I was on a call just the other day and my colleagues were trying to brainstorm a template for some sales collateral and we were going back and forth and everyone was giving their opinion, but it wasn't quite landing and we weren't really getting anywhere and in the background I was just there talking to chat GPT and within like seconds, I've got this starting point. And I pasted in the chat and everyone went, Oh, it looks like Paddy's got the answer. It's not perfect and it's not probably the actual thing that we were going to go with. But the fact is it gave us a great starting point instead of this blank piece of paper that we were just not making any headway on. And it's starting to become more and more of a go to now in terms of the things I do. In terms of. People out there that are a little bit afraid or perhaps not been using it so much. AI's Limitations and Strengths [00:23:10] Paddy Dhanda: What advice would you have for them? Because it is scary if you've never tried this sort of stuff before. [00:23:16] Mark Pesce: I mean, it's scary when you think that it doesn't have any limits, and it does have limits. It is a very. Let's put it mid range thinker. It is not broadly expressive. Sometimes it can crack a good pun, because it's managed to find a good pun somewhere in its database. But in general, it will give you what we would call mid range content. Now, when you're just trying to get a form for your ideas, all you need to get started. Everything that goes in at that point then is human creativity. But I think it allows us to see a new kind of relationship. And I think people are thinking that the relationship is going to be just a take relationship, that the computer is just going to take everything, right? And in fact, where we see this going already is it's much more of a play, where we're going to the computer to help us do this thing that is a little bit too routine for us, right? But the computer is really good at. Cause it's read a lot of stuff and it can help do that for us. And then we can fill in all the interesting bits. I have a friend who works at data 61 data. 61 is a science organization here that does it. And he had to do a great big report. I think it was an ISO standard that the report had to be in. And so there were lots of specific rules about which parts went where and what went into which parts. He fed the ISO standard into CHAT GPT and said, okay, I need you to now spit out all of the boilerplate for this. All the parts that I don't need to worry about that just need to go in here because they need to go in here because it's this kind of document. And it did all of that. And then he was able to focus on the bits where creative thought and his creative input were required. [00:24:51] Mark Pesce: We live in a bureaucratic civilization where there's a lot of paperwork. We have to cross a lot of T's and dot a lot of I's. That's just the nature of business in the 21st century. It's also part of the nature of digitization in the 21st century. And I think that a lot of what we're going to be using AI chatbots for is to make sure that that paperwork is actually dealt with and not just kicked into the never never. AI's Impact on Legal Profession [00:25:15] Paddy Dhanda: niece who's training to be a lawyer and when I first introduced chat GPT to her. I think she was blown away. She was like, what does this mean for my job in the future? I'm like literally on my path to qualify in this profession that is being massively disrupted. a lot of people out there that are worried. [00:25:36] Mark Pesce: All right, so next time you see her, this is a story that I tell in the book. There's a lawyer who was researching a federal lawsuit against an airline and went to CHAT GPT and thought that CHAT GPT was rather similar to LexisNexis, which is a big legal research system that they use in America to get research in American case law. And said, okay, so here's the case that I mean, can you give me some references, and ChatGPT dutifully gave him a whole set of references and citations, which he then put in the brief, which then was submitted to the federal judge. ChatGPT had made all of them up. The federal judge! was not amused. And he basically had to grovel in front of court and said, I'm sorry judge, I thought this was like all the other tools I'd ever used. I had no idea this could happen. So, when we're thinking about those kinds of things, I don't think an AI chatbot is great. going to be the kind of thing that's going to take that kind of work away. It's not well suited to that. But it will be very good at doing all of the boilerplate that we get paralegals to do. But that doesn't mean we're not going to need paralegals, because paralegals will still be setting those programs up and then inspecting the output to make sure that the output is actually suitable. So this is what I'm saying. What we're actually learning now, even just one year in, is that AI is not about mass unemployment. It's going to turn us all into people who spend a lot of time eyeballing and checking the outputs of AI. [00:27:08] Paddy Dhanda: And that's a really interesting point. AI's Hallucination and Verification [00:27:09] Paddy Dhanda: I heard the other day I was at a talk and we had a, an expert in AI and he mentioned one of these reports I think it might've been the Oxford dictionary, they have like a word of the year and the word of the year was a hallucination and. [00:27:24] Mark Pesce: Hallucination. [00:27:25] Paddy Dhanda: Yeah, tell us more about that because I think that touches upon this story you've just told us about and it's definitely not hallucination of humans I don't think we're talking about here are we? [00:27:34] Mark Pesce: So we talked about this transformer, which is at the heart of what a chatbot is doing. And what happens is you put your question to the chatbot, your prompt to the chatbot, and it goes into the transformer. And the transformer is basically basically looking at all of the data that it's learned while keeping its eye on the question that you've asked. And what it's trying to do is it's trying to find the most likely, statistically likely response to the question that you put. And it's going to generate that response. But the chatbot has no sense internally whether that's true or false. And you've got to remember that the internet is filled with Well, lies, let's just put it out there, and craziness, and rants, and all sorts of misinformation, as well as Wikipedia, and The Times, and The Guardian, and all of these great news sources. But it's a mix, and the chatbot's been trained on all of that. And so, The answer that it generates is what it thinks is the likely answer, but may have zero basis in fact. And as near as anyone can tell, that is an artifact of the way these systems work. In other words, that is not a bug, it's just the way these systems work, because they will be good at generating the most likely response, but the most likely response may not be the correct one. [00:28:53] Paddy Dhanda: how do you verify if you're being told the truth because Our go to would then be the internet and if the internet's wrong, then I'm guessing some human somewhere [00:29:05] Mark Pesce: is particularly a problem if you're asking it a question where you don't have enough expertise to be able to sniff out that does not feel right. But maybe, maybe, you know, in your gut you're like, really, Chat GPT? So as soon as you have that feeling, the best thing to do is to ask the question a different way. Rephrase it. Be less vague. Be extremely direct. Give an example or two, if there's an example or two. That helps ChatGPT generate a more correct answer. And if that more correct answer is exactly the same as the answer it got you last time, okay, maybe it was the right answer. If you're still not convinced, if it's ChatGPT, go and ask Claude. All right? Because again, trained entirely differently. So if you're getting exactly the same answer from both of them, again, higher probability that yeah, okay, you don't necessarily believe it, but they're both saying the same thing. Might be right. The big thing here is that until last year, we never had to consider that a computer would make things up. It was impossible. Computers didn't make things up. They would have bugs, but they couldn't make things up. Well, actually that now has to play into our relationship with a computer is that actually a computer just makes things up from time to time. And so we need to now think about how we can actually run the ruler over it. [00:30:28] Paddy Dhanda: And you touched upon a really important part of ChatGPT and these bots. Art of Prompt Engineering [00:30:31] Paddy Dhanda: And that's the way we ask the question and those prompts. There seems to be this almost like a profession developing, which is around prompt engineering. Like, you know, people are selling prompts. They've spent time and effort working these through crafting these questions, and then they're selling them online. Could you tell us more about, like, what makes a good prompt? Are there specific words we should be using? Is there a specific structure that I should be using? [00:31:00] Mark Pesce: So, okay. This is, this is a very big field. And the reason it's a big field and also a very messy one is because we're all playing with English. And English is a big, fun, messy language, which is imprecise in all sorts of really wonderful ways. Which also makes it. terribly difficult to sort of get a handle on, right? That's kind of the beauty of language and English in particular has a lot of slipperiness around it. And so that's what makes prompt engineering much more of an art. And I would probably say a black art than a science. However, what we know recent research published, I think only three weeks ago, it turns out that a chat bot we'll pick up on the emotional tone of a request made to it. And so, If you want to improve the reliability of your chatbot, you will end your question, your prompt to the chatbot with, this is very important to my career. You put that in, statistical improvement across the board with any chatbot, probably because it's read enough information where people are pleading about getting good data
