mergeconflict353 === [00:00:00] James: I am really excited everyone, because I believe I have finally 100% Microsoft, Frank. Just now today as he has installed Microsoft Edge and is now playing around with the built-in, I dunno if I'm gonna call it Sidebar, is my new favorite feature of Edge. What's the Bing side? There's the edge sidebar and he is now playing around with the Bing Chat. Directly in it. How's that going, Frank [00:00:33] Frank: James, you caught me in a vulnerable mood. All right. I was, I was already running a Google Chrome, beta, you know, I'd already broken outta my scary where I want to be in browser land and you're like, you can't install the edge. And I'm like, do they make Edge for Mac? And you're like, You, you are so dumb, Frank. Anyway, that was the story as I remember it. And so you convinced me Yes, to install the Microsoft Edge browsing experience on my computer and to check out, I believe I'm gonna call it the Discover area, but you're right, it's, it's Discover Chat, discover composed, discover Insights and what they do. I, I think you might have mentioned it before, but I never tried it. They baked chat, G p t or whatever, who knows what, into the Bing and. Kind of get where you're coming from, where this is kind of useful. So I guess we should talk about it. [00:01:24] James: Yeah, it's, it's really neat. This is, this is, this again, is one of my new like favorite features that has happened as of recent. And what I think was really nice about it is there's this, there's this sidebar and I've been using it at work all the time. Not, not the Bing Chat, but specifically the sidebar, because when I go and it's really neat integration, like when you open a link and outlook, it'll open. Obviously it's a link open in the browser, but like Outlook and Edge know about what's going on. So it opens like a mini outlook inside of your browser and you can like archive and respond. It's just, you know, multi window mode. Like, it, it just, you know, it makes a lot of sense. But I did find myself today we were talking about, um, Amazon. Sidewalk in our Patreon feed. Uh, go check that out. patreon.com/merge. Conflict plug. And, uh, and, and you were ordering it, but we just had a lot of questions. You had a lot of questions and I didn't have the answers. And like you said, the Amazon website that we signed up for the, for the test devices was extremely sparse. I didn't even know where to go. I didn't know where to look. I had just read an article online, so. In real time. As we're looking at that website, I decided to have a full on conversation, uh, with Bing chat on this sidebar. Just opened it up as I'm looking at the Amazon, you know, site and I'm recording the podcast in real time, so I'm just having this conversation. I don't have to open another tab. I don't have to do anything. I'm just having this conversation and you're asking me questions. So I am asking eight questions now. If we can get a oh chat, G P T. Um, plus live share integration where we can share a chat together. Now that would be something [00:03:06] Frank: I just realized. I kind of used you as Aurora from Star Trek, where someone would ask something and then she would relay it to the computer as made fun, fun of by Sigourney Weaver and, uh, whatever that show was. Galaxy Quest. Oh, that's hilarious. You are my intermediary to the computer that I didn't have the patience for. But, uh, yeah, this, okay, so sidebar is a little bit different from chat. G P T number one. The prompt is ask me anything. Wow. Cheeky stole that from Reddit. I'm like, what? What do I remember that from? Oh, that's a Reddit thing. Wow. Microsoft is embracing the Reddit. Anyway, um, it, it's. It's a little bit better, I'm gonna admit than chat G P T from a UI perspective, because it doesn't only just answer your question, it keeps the context, all that stuff that chat G P T does, but it also gives external links to websites if you wanna get into that kind of stuff. I gotta be honest, the website links feel a little bit. Advertising to me. We'll, we'll, we'll get into that psychological thing, but that's the big difference is that it's, it's giving you pointers at least into the wider internet being, being, of course that makes sense. Um, of how to follow up on it. So I find that part. Interesting because I've, I have been getting into like how, what is a good UI for these language models? Mm-hmm. And this is definitely a step. Yeah. It's a step in a good direction. I'm not sure. I feel like we can go further. [00:04:38] James: Yeah. Well you know, here's a good case of point to what you just talked is a lot of people are interacting with chat G P T, we talked about your command line tool for chat, G P T. Now the Bing chat is, it is different, right? Cuz it's integrating those search results com combination and, and feeding that back. So for example, you know, the first thing that I asked was, You can gimme a small summary of Amazon sidewalk and it wrote up a thing and it, and it did a learn more to, to, to Tom's uh, guide to about Amazon and another Amazon page and a few other ones. But to the point, there's actually like this amazing article on the, about amazon.com website from Sean O'Neill, and it was from seven days ago. You know, new, new tech, new technologies like this is hot off. You know, the things. This is really, really cool because it's integrating all of that data into it from seven days ago. It's not like there was a new model created for it, right? Um, it's just pulling that stuff in. So it was able to help summarize this stuff. And this is actually an amazing article, by the way. I'll put it probably in the show notes, uh, about it if you're interested in the Amazon sidewalk. But, um, yeah, I think that's really neat. And the cool part about it, I, I, this is why the sidebar, I think is important, why the, the integration here is so cool. Is because those learn more links, you click on it and it just opens on the left hand side. Right. It's not opening a new tab. It's basically splitting the browser into a dedicated chat G P T window with Bing chat. And then additionally, on the left hand side is gonna be all your sources and tabs. So it's kind of multiple applications being hosted. Mm-hmm. Inside of Edge, which is kind of. [00:06:14] Frank: It's the old calm days, right? We, we all used to say, we're gonna write our apps as calm controls. No one could just ever figure out how to sell calm control. So it didn't really work. Uh, you know. Okay. My one, I, I love everything you just said. My, I want to give a lot of first person perspectives, but that's not right. Yeah. I can't resize the right hand window. That's a little bit annoying. So, but I think, no, I'm failing. Can I get it to detach at all? No. No. No. Okay. So it does, they're definitely going to the sidebar analogy. But here's the thing, man. I think the world's gonna switch. It's gonna be AI first, you know, website second. So I almost wanna flip it around. I want the AI in the big middle area end, put the web off to the right where the web belongs. I don't, I like that. I don't care about, yeah. Uh, I, I wanna talk about one little thing. Do you mind if I just do a small rabbit hole? Uh, because I thought it was kind of interest. [00:07:05] James: Yeah, go for it. Because I think that, you know, the reason I want to talk about this is we haven't actually talked about the Bing Chat, uh, integration. We talked more about chat, G P T and those APIs, but I believe there's some unique things here about how it's asking you to prompt and the, the, the dials, if you will. And that's what I'm hoping that you're getting into. But if not, then just go dive into whatever you want, Frank. We don't plan anything on this podcast, by the way, people, so this is how it, this is how the, this is how it's, uh, it's. [00:07:30] Frank: We may not, we may not plan anything, but you know me so well, you you knew exactly what I wanted to talk about. I know. Uh, yeah. So I wa I was curious because on the initial onboarding experience, you have to be logged in. So sign into your Microsoft, whatever, and then press the big blue button. And I was all excited to try it out. But it was funny, they actually made me answer one question first. And. This surprised me not because I didn't understand the question. Uh, very few of the UIs I've seen have a bothered to ask this question. B thought it would be useful. C I don't, I don't know why. Anyway, the question is this. What kind of responses would you like, James? Would you like creative? Would you like? I believe I chose moderate, or I believe the last one was precise. Uh mm. What, what do you think those mean to you? What, what do they mean to you? [00:08:27] James: So, I believe it's trying to, it's trying to get some insight into what you're, what you're looking for. So here's how I perceive them. Let me, let me hit the clear button here. New topic button. So yes, creative, balanced, and more and precise. Balanced. Okay. Nice. Yeah. So creative. I think this is, if I wanted to try to like create a story or I want it to create a recipe or a lot of people have it like create lyrics or a poem, I think it's in that mode. You know, kind of this is more, maybe it's dialing into some model or something that is trained on more like, Futuristic or fantasy or something like that. I feel like balance is just like the default. It's like, Hey, this is, this is just normal. And then I think precise is, that's a good question. I feel like precise is like, Hey, I'm, I'm doing re, I'm doing research on a, um, on a topic. Um, that's a report and I need to gather more information. And maybe there's something in the algorithm behind the scenes that's like doing check, fact checking across multiple sources, like with the Bing integration, you know, that, that stuff. But my assumption is that, that the more precise would mean that ideally the results I'm getting back are more accurate as far as if I ask it about James Monte Magno, it's going to. Maybe really dial in or Frank Krueger, like really dial in on, on, on that type of data and not look at other sources. Maybe it's limiting the sources it's looking at. I'm not really positive. What do you think it is? [00:10:11] Frank: Hmm. Uh, I, I like that. Um, I don't think it's any of that because everything you described A would be wonderful. B is a whole lot of work. Oh. And I don't think they've done it quite that far, but I like where you're going. It's. I I wanted to ask you because I, I feel like with new technologies, if you don't know exactly how it works, you have to interpret ideas. And I, I'm really hesitant about which words Microsoft chose here. Um, because I think they can be misleading because I think they direct people into thoughts like you have to believing it does what you just described, but you're not far off. What I think it does, what I think it does is all of these neural networks are still stochastic machines. They're still random. They produced probabilities. They don't produce words, and so you see when they're writing, they go kind of word by word token by token, but it doesn't matter. It's going word by word. And so whenever you run these networks, and I learned this a lot when I wrote my JavaScript library for running these language models in the browser, it's half of the battle is just executing the network to get the probabilities of which word should be next. So, you know, given these three words, It has a dictionary of who knows I, I think it's somewhere around 50,000 possible words that can follow up with that. So for every 50,000, here's the probability of that word showing up. And so those three levers that you're describing are called a sampling strategies. Given these probabilities, how do I choose what? Actually to present in front of the user, I can't present the entire probability space. It would make absolutely no sense to the user. And yet you still kind of wanna explore that probability space. So here's what I think is happening, um, when you do that sampling, when you just have probabilities, you throw in. Random numbers and you sample kind of randomly using the probabilities. So things compact the way they should, but it's still technically random, and I believe precise is returning the most probabilistic result experimental. Or, uh, what they call creative. Creative. That's, that's, oh god, what an overloaded word. Uh, creative is returning a more random response, and balanced is somewhere in between. So I believe the lever that they are moving is the randomness generator when they're actually sampling the neural network. [00:12:42] James: Yeah, that, that probably makes some sense. I'm in, I mean, I am interested, I don't know if they have an article. Uh, actually let's just ask it. Uh, um, what is, I love it. The, how does, how does the Bing Chat, uh, tone of response work? Question mark. Let's see what it says. Mm-hmm. Um, cause that's what it's asking. It's asking for the tone of that, of of, of, uh, the, the response again. Oh, and this is kind of what's interesting about it is it, uh, it has this, it says, but how does it work? It says, Bing Chat has a feature, allows you to change it. Tone response. You can choose between these modes. By default, it returns balance, toggling. But how does it work? But how does it work under the hood? Does it know that when I say under the hood, [00:13:34] Frank: probably does, it's a common enough expression. Interesting. [00:13:39] James: That's, oh, it actually said, okay, so it says, and it doesn't really gimme a lot of information, Frank. [00:13:54] Frank: Well, I'll, I'll say this much. All of these neural networks have that setting. Okay. So Microsoft has to use one of those settings, and the fact that they're bringing up that prompt just makes me think that you're choosing that setting. Yeah. Fun fact though, most UIs for this, just default to the precise, which is give me the next most probable word. And that's usually. We used to avoid that in the past. Uh, did you ever do many experience with G P T three before it was chat g p T? Did you, I know we talked about, did you ever play with it much? [00:14:27] James: I did not, no. I was kind of pretty hands off. I've been actually pretty hands off with everything, including all the image creators, everything like that. Cause I just felt like I wasn't, I, I don't know. I just, I didn't dive in [00:14:38] Frank: Basical. Yeah, well the, the old problem with those ones is greedy sampling, which is taking the most probabilistic, next word, the most probable next word, um, I, I, is that it would get into nasty, ugly loops with itself. It would just start repeating itself over and over. Um, it would just, it, it was bad. It was bad. You could never return the most probable next word. Um, but then with the way chat, G p t was trained, oddly enough, most people defaulted to the next most probable word. And that's the way I've seen most UIs work. That's the way my UI works, is the way most of 'em run, because none of us knew how to present that option to users. So anyway, I think we can finally get out of this rabbit hole. I just thought it was very funny. A, the words they chose B, the setting altogether C, the fact that they don't really explain it. [00:15:33] James: Yeah. Yeah. It, it is, uh, it is, it is fascinating. And I am curious, like when, if I was to go to, you know, bing.com and then there is also a chat button at the top, you know, is it, is it exactly the same? I think it is. That gives me those, those options there as well. So that's kind of cool. Um, you know, they're, they're unique. They're the same. I think that the main Bing chat when you go to bing.com, I think slash chat, I think it may give you additional stuff, but I'm not really positive. Yeah, [00:16:03] Frank: yeah. Can I talk about a feature I like? Yeah, I'm ready. Okay. Uh, after I asked it a question, it suggested follow up questions. I'm not in love with the follow up questions it offered, and yet, uh, in my brainstorming of how to make better UIs for this thing, I immediately realized, yeah, automatic follow up question. Generation should be a part of it. Don't make me type that, you know, if I want to drill into something, I almost feel like every. AI response should be a hyperlink to just formulating a question around that word. Ooh, there's an app idea. Anyway, uh, I just like that Bing Chat does that. So I got, uh, three recommended, uh, follow up questions. None of 'em really apply to what I care about, but I also asked a very poor question with very little context, so that's fine. [00:16:55] James: Yeah, I do like that as a, as a. As sort of a, I maybe don't know what to ask at next, right? Mm-hmm. Because I think there are people that are honing their skills and people are starting to improve upon here, uh, and get used to it. The more you use it, the more you are gonna dial in. But if you're given this sort of, you know, Helping hand along the way, it's going to be like, oh, I didn't know that. And that was really helpful inside of the, um, sidewalk conversation I was having. It literally was like, uh, um, do you want to know more about security questions about it? You know, and I'm wondering how, how do you think it's generating those, um, prompts in general? Do you think it's sort of reading those, uh, articles and coming from chat G P T and it's, it's kind of. Does it look in those sources for like next viable information and craft something off of it? Or is it more like, oh, other people have asked similar things, X, Y, Z, or a combination of it. [00:17:55] Frank: Uh, not being, being, I, I could see them using your last solution there of being already having kind of a questioning system. But that said, it is trivial for these networks to generate their own follow up questions to their own prompt that they're very good at that. In fact, uh, they may not always. Be in the right order or anything cuz it, it's, uh, it doesn't know what, you know, it can't read your mind yet, yet. Um, but when it can, it'll ask the best questions. So I would go with, um, either it's generating the questions, which is the method I was gonna use, so therefore must be right. Right. James or so, uh, being has had, uh, uh, research for how long now? 10 years. I'm sure they had an internal engine that can also do follow up question. [00:18:43] James: Yeah, I assume it would as well. Yeah, I do like that. That is a nice option there. It's pretty, pretty cool. It's pretty fast. I'm pretty impressed. [00:18:52] Frank: Yeah, the, the speed is definitely something I'm, cuz I have my own little command line one, I, I know exactly how fast it returns responses. And you were mentioning when you were using it, you're like, this thing feels like it's speeding up. And I, I was wondering if, um, they do a tiny bit of user throttling. Like if you just ask it a one-off question, you don't get high on the queue. But maybe if you're continuing a conversation, maybe it puts you a little, a little higher priority on the. [00:19:17] James: Yeah, I am appreciative of a few things. Like I just asked it about the big image critter, which I wanna talk about here in a little bit. I asked it like, how does it work? And it gave me a thing, and you're right, it says, you know it's, it talks about Dolly and a bunch of stuff in this statement, but it also says, what is Dolly? What are some other AI image generators? How can I use image generators by being image creator, blah, blah, something else, right? That kind of doesn't wrap around, but you can hover over it, I guess. But to your point, right, those are pretty useful. Blocks of text. Now. I think what's really fascinating about this is that Chad isn't the only mode. There's something called compose as well. Yeah, [00:19:57] Frank: and I'm gonna let you run this one because I have very little experience with the compose, but you were, you were making some good arguments for it. [00:20:04] James: Yeah. So, In general. So, so there's a few things. I think that, because it's called Bing Discover, so there's an, there's another one too that we'll get into, uh, that, that's really, really fascinating. It's a totally different mode, but it's about sort of interacting on and, and the website. But imagine compose, like I've, I've had a conversation, I'm getting some context, and now I want to. Um, summarize something, or I want to create a blog post or generate an email. Maybe I want to generate an email that says, give me the, uh, write me, write me an email, uh, about the top, you know, locations, uh, in Finland, uh, where people use the sauna, right? And so mm-hmm. Right. You could, because you could ask it that, but it's going to basically have a conversation with you. What you want it to do is, Uh, do some of the heavy lifting for you in different formats. So now we're not in a chat mode. Now we're in a creation mode. Alright. I think now we're sort of sh that's why it's called compose, right? Because we are in the chat mode. Where it was creating texts and responding to us, but it's having a conversation. Whereas the compose mode is sort of like Dolly or the other image creators where you're having it, giving it a prompt and you're giving it some settings to tweak, uh, and then it's gonna create something for you so you can ask it to write about anything. So I specifically asked it to write about Amazon Sidewalk and how developers may use the technology and build devices and include security concerns. And you can give it a tone. So kind of like we were talking about earlier, but there's more, there's profess. Casual, enthusiastic, informational, and funny. I put professional, and then you can give it a format. So do you want a single paragraph? Do you want an email? Do you want a blog post? Do you want a list of ideas? And then you can give it a link. So short, medium, long, and it'll generate a draft. And it generated an absolutely stunning, uh, draft of everything Amazon sidewalk. So imagine you just had a conversation with, uh, the, the Bing Chat. The chat g. And you're getting that context of what's important that you might wanna report on or what might a summary of, and then I think here, the fascinating part is you can then go into compose, come up with a better prompt now that you have more information from the conversation, and then have it output a bunch of stuff. [00:22:27] Frank: Yeah. Uh, I kind of love this now that I know what it is, it took, it took me a few seconds to understand what was going on there, so thank you for the exclamation. Uh, but I, I, I gotta take it up a level because. The way I see it, this is just one in a mini series of apps you can basically write with this kind of thing. And they created a composition app for writing some text stuff. And it's neat. Um, it's, it's almost like advanced ting. I always make the madlibs joke. But, uh, these networks, if you just tell them roughly what you want and then prompt 'em, uh, then. Generate roughly what you want. And so it's almost like this little outlet here is generating a pre promptt something that goes a little bit in front of yours so that the network knows exactly what to generate. And it's neat. Uh, the formats they have, the outputs they have are paragraph, email, blog, post idea. Those are fine. It's a good starting point. Yeah. Um, you can, you can see others coming up though, like see sharp. Sharp code. You know, there's a lot of things you could add those into, and that's why I'm saying this is almost a general UI for just a another UI into the engine. I do like the ideas one. This was the first time, you know, we talk about this so much on the show and everything. You forget that you sometimes live in a little bit of a bubble, and not everyone is playing with these as, as everyone should be. And a friend was designing a little video game. And we were just kind of like joking around, thinking of different weapon types, you know, different missiles, different guns and everything. And I'm like, I'm just gonna ask the ai, like give me 50 different weapon types for an action space game. And oh my goodness, it came up with, you know, a million ideas, 50, you know exactly how many I asked for. Of course it's a computer, but, uh, I, so that's why I find it funny that there's an ideas, a bulleted list that can be the output of Bing Compose that is doing exactly. Exactly that use case that comes up quite often. [00:24:27] James: Yeah. And you could see this, for example, being used by, um, you know, to start a draft. Right? Uh, I think that M K B H D was talking about specifically like, Hey, you know, I, you know, start my, my, I have an idea. I have it generate a script for me as a starting point, and then I go and do the validation of the research and make it my own. Or maybe I want it to be the one that creates some of the, the descriptions in my YouTube. Um, You know, um mm-hmm. Video description, summarize and stuff like that. Yeah. Or I like the list ideas like you're saying is we do a podcast and we never know what we're talking about. So like, what if I asked it for top podcast topic ideas for new features on.net seven, and what if it told us, um, how about how to use the new global using feature to simplify code benefits, a new minimal APIs new C sharp features like record STRs, file scope names, faces, interpolated string handlers, how to migrate your code to donnet seven and potential challenges and solutions like it just did. [00:25:26] Frank: How do I upgrade my coat to Yes. We'll, we'll do an episode on that one. Yeah, [00:25:32] James: I like that. [00:25:34] Frank: Yeah. So, okay. Uh, I, I totally dig, compose. It's something you can get chat to do if you prompt it enough. So this is just reducing the amount of prompting you do. And I, I'm just gonna repeat myself because I think it's important. This is the future everyone. Um, if you're an app developer, not now, we're all fine for now, but in 20 years a lot of apps are gonna be prompts for an AI and you might as well get good at it now cuz it's not going anywhere. There are bubbles, bubbles will burst and all that. But this is a foundational technology and I like to see these tiptoes into apps written for it. But I think we can go much, much, much, [00:26:14] James: much further. And the, the other thing that is in this tab as we close out this tab, and I have one more thing to talk about after this, is there's this insights tab, and this is kind of fascinating and I haven't really used it until now, but this is a little bit more contextual chat and compose. Those are just, you're giving it stuff and then it will open the browser on the left hand side, insights is more, give me insights on the page that is open right now, so. As an example, if I open monte magna.com, it will show me the top sort of hits. For what it is finding relevant on this page. So it's finding my bio, my GitHub page, my new GI feed, my Twitter account, and my YouTube. And then it also gives me related searches to this and then about this site and then analytics. So it gives me monthly traffic. Right. And mine's been going down every month. That's great. And where it's coming from. And it's coming from. Mm-hmm. Is it coming from Search engines gonna stop blog as much, but I think what's interesting here is, let's say I go to In Gadget, and let's say I tap on an article. Let's say I'm talking about the Unni Volt 12 taking pizza party indoors. Now, this is fascinating, okay? Because what it does is it's not giving me information about En Gadget, it's giving me information about what's in the article on En Gadget. So it's going and finding me q and a. It's finding me key points inside the blog. It's giving me more information and then it'll gimme information on the site, right? So, you know, if you're looking for more resources, it's kind of like. I find an article online, but then I really want to go do more with it, and I could just ask Chat, beach Chat, G p t, or to x, Y, Z. But this thing is kind of doing a big summary as if you had binged for it, basically for all intents and purposes, but kind of organizing it nicely, I think. So that's kind of cool. [00:28:06] Frank: Yeah. Yeah. This is, this is what we need too, is the passive. As much as I love the prompting and the prompting is fun. Mm-hmm. Um, we need the passive here. It needs to just be the little inte sense that pops up. All the time. And that's, that's just what I'm gonna call, I'm gonna call it web intes. Yeah. [00:28:25] James: Um, it is, it is web IntelliSense. I, I like that, that's what it really should be called. I love it. [00:28:31] Frank: Uh, no one else other than programmers would like that name, but, uh, you know, people have tried this for a while. I remember sidebars, we used to write sidebars for browsers, and we pretended to do this. We're like, we're gonna give you relevant. Search links to our advertisers with very popular shopping items to get through people we have deals with, and those have never worked out. So it's funny, I would almost trust an AI more over like a company giving me a suggested sidebar like this. I'm like, please, and, and, and tell me how you're pre biasing the network. And it's initial prompt. So I know exactly what I'm getting. But it's funny that this, this idea is, But it's coming back around and it's funny that I kind of trust the AI [00:29:14] James: more. Yeah. All right. Now here's the last thing that we talk about in our Bing Not Sponsored by Microsoft episode, uh, at all. How do you feel about bing.com/create? We were talking about Discover, and now we're creating frank [00:29:29] Frank: bing.com com slash create. Yes. Ooh. Create an image with the bing. All right. All right. Um, I would like to see James Monte Meg, no, I know how to spell your name. [00:29:48] James: Toes, yeah. Mango. [00:29:52] Frank: Mm-hmm. [00:29:53] James: It's generating, it's just generating a photo of me. It's not gonna be great. I did try that. Um, so did you Yeah. What the being image creator is specifically is, uh, image creator using Dolly in a very. Nice format. It gives you four results back. You can share, you can download, you can do a bunch of stuff. I've been tweeting some out recently. Uh, I just did one. Uh, here's, here's a, here's a good one for you. I like people, people's really entertaining cuz they have way too many fingers. Mm-hmm. Um, but this one here, and now you can share it with people. This is what I really like. So I just sent you a link on the Z zencaster. But I put visual studio code presentation with a code editor in front of them showing tips and. And it generated, um, a, a person with glasses and a lanyard on cuz they're presenting, which is funny, uh, pointing at a screen that has a big code editor on it as if they're like in a like user group or something like that. And that's like one of four that did a bunch of them, which I think is great. Yeah. Why are fingers so messed up? I watch, there's a Vox video on this. I don't understand, but there's another one like, so this is really cool. And I did one that was. My favorite one so far as I'm getting better at this is a cute purple robot that is programming code and racing away in a sports car, modern and vibrant colors, and it, and it generated one of the most adorable don net bots that I've ever seen in my life. I'll just share that with you. I put it on Twitter already, uh, but it's so freaking cute. And now this is generated Frank, like this. This isn't art that exists on the internet. This is a unique. Art that was created for me. Is that true? True. Yeah. [00:31:28] Frank: I don't understand how this works now. Okay. I'm gonna say yes. People, there are caveats and hashtags and everything. Probably these networks could reproduce existing artwork if you prompted them correctly and said, you know, I need a Mona Lisa and the color of Da Vinci. You know, like give it exactly what you want. I could probably do. But in this case, your crazy little icon in the car, your robot, I mean in the car is, yeah, that's unique. It was a hard step getting the networks to be able to do this. We could always get the networks to regenerate images in the data set. And we could get them to generate images that were a little bit outside the dataset, but then you start getting weird things like people with four fingers or Yeah, four fingers or six fingers and things like that. So there is this real trade off you make with neural networks of memorizing the dataset versus being. Creative. Creative, and you might have heard of the diffusion networks. That was the big breakthrough that allowed us to create these more creative networks that could compose more complicated scenes, not just things that's seen before, but have a concept of a robot, have a concept of a car, put the robot in the car, make it vaguely. I mean, there's tons of mistakes with this image, but it did a, it did a pretty darn good job, in my opinion. So, yeah. This is Dolly. This is classic Dolly. Can I explain to you the finger prop? [00:32:59] James: Yeah, I would love to, [00:33:01] Frank: uh, again, this is, I don't know, I don't work at Open ai, everyone. Hi, I'm living on island. Um, uh, it's just the data set. Uh, they chose not to bias the data set toward people. Hmm. And what are people most interested in, James? [00:33:20] James: Yeah, I'll answer. True. [00:33:24] Frank: And so the people want to generate people with these things, but they were afraid of all the pornographic uses, all the honestly terrible scenes it could generate because it. Creative, deep fakes. Um, it can do deep fakes very easily. It's trivial for these big networks. And so they made an active, conscientious, conscientious choice not to, I mean, it can do people, but they didn't bias it toward people. They didn't make sure that it did people perfectly all the time. They could have, they definitely could have, because now there are pornographic networks that can do the exact right amount of fingers on a person among. [00:34:04] James: Yeah, if you just put in like James Monte Magno presenting at a conference, it doesn't look like me. Um, but it does look like. Kind programmer bro. Uh, programmer bro. Yeah. Um, presenting and I, I don't, they're all wearing suits though, so I, I guess I said at a developer conference, maybe it would, oh, lemme do that at a, I got two [00:34:26] Frank: t-shirts and a striped shirt. None of 'em are really your style though, so they're definitely lacking knowledge of you. I'm gonna do James [00:34:33] James: presenting at a developer conference in the future. Now I have nine 19 Boost Credit Sloth, so I'm gonna use the. That's my favorite video boost credits, that [00:34:42] Frank: it's Microsoft. I, I assume the whole thing will be Xbox points at some point. [00:34:46] James: I did this one. I'm using them now on my YouTube thumbnails, but as background behind it. Okay. To gonna do some stuff. So I did one on dev tunnels and, um, it's cool when you share an image, it shows, shows you the prompt. So this one was a, a, a program that shows a tunnel transforming blah, blah, blah into it. Um, and. And that was pretty cool. I use that as a background tunnel. Yeah. Now if I do a pro, a developer conference in the future, I don't really get the future aspect of it, but I am wearing less suits in this instance, which is good. [00:35:20] Frank: It, yeah. And again, these things are so castic put the same prompt in it gave you four images, it'll give you another four and another four and another four. Um, it, you can get 'em to reproduce the same image, but it's a little tricks. And by Trixi, [00:35:37] James: I mean very simple, but here I think the image, the image of the people are actually pretty good. Uh, very fascinating of these. But now I'm gonna say a monkey presenting a developer conference in the future outside over a mountain view and. That is probably, hopefully gonna be absolutely delectable. Now this is fun because I don't have to pay for anything. It's all happening in real time and it's all just built in, uh, which I, I absolutely. So [00:36:03] Frank: you don't have to pay for anything, but they have boost points. Does that just make it [00:36:07] James: faster? Is that what makes it faster? Yeah. This is monkey. Oh yes. [00:36:12] Frank: This just your first image generator. James. We've been, we, we all went through our image generation phase a month ago. Oh. [00:36:19] James: This one's even [00:36:22] Frank: this one. That is a cute monkey though. Uh, I will bring up a point that you're discovering. Uh, these networks are terrible at text. [00:36:28] James: They're bad, real bad. [00:36:31] Frank: Uh, so I'm, I'm gonna self-promote here. Hi. Go to pre clam.org. I just wrote an article chat. G p t turns out is capable of generating s V G images. Mm-hmm. Uh, scalable, which is nice for UI stuff, but also, uh, it gets text, right? Because SVGs represent text as text. And so you can be guaranteed that your, uh, image, whatever you wanna call it, will actually have the correct text. [00:36:57] James: Oh, that's cool. I like that. Yeah, I did one cuz a lot of my stuff I've been doing is for programming stuff, so I've been talking about like a code or an editor or whatever, and it's just a bunch of shenanigans, but it looks cool. I, I think it looks like something outta hackers. So like, there's a, there's like a, a mainframe, it looks like I put a, I like, we use words like neon futuristic. Okay. Like, it does, like, you know, it's, it's really into that stuff. Um, um, yeah, I think it is all [00:37:24] Frank: over. Think of it this way, um, chat, G p t based on G P T three had 4,000 tokens of context. So the, the sentence you typed in, call it roughly two and a half tokens per, uh, word. Did I say that right? Yeah. Um, we're giving it so little information for what it can take. Yeah. So you're giving it what, a hundred words, 10 words, 50 words? Mm-hmm. You could give it 2000 words of prompt and. And still have 2000 characters left over for it to generate. It's a grandson between the input and the output. Yeah. And so you can really prompt yourself into oblivion with these things if you want to. And that's what the chat is doing, by the way. It's just feeding that back buffer and making it bigger. Love it. I should say, God, gosh, sorry. We did, we did, we did an episode on G PT four, so I already talked about this, but I always use the G P T four version of Che j p Tino, and that thing can handle up to 32 k a token. So, oh, you can really go insane with the amount that you can, uh, prompt these with. [00:38:31] James: I am this, you're saying that these images don't exist. Correct. Look it. Yeah. I wanted to do something a little bit last. You can test [00:38:41] Frank: it yourself. Take one of these images and do a Google reverse image search for it. [00:38:45] James: Oh, that's a good idea. I did a, it's, it's very easy to test. I did a sauna outdoor near a cabin, snowing at wintertime and is very good. I'm very impressed by all the, yeah, well some of them are really photo, oh, and I guess you could specify photorealistic or Renaissance era. Oh. It gives you the tips and I like that. That's kind of. [00:39:07] Frank: Oh, James, after the show, sorry, everyone, I'm gonna show you an app I've been developing that really helps you do this kind of stuff. [00:39:13] James: Well, I think with that then we should just end the show. Also, I'm just gonna start generating images and this probably hasn't been fun for anyone, but I'll put some of the links into the show notes, so you, you better put all [00:39:22] Frank: the links in, let people follow [00:39:23] James: along. I will do not, do not worry. Everything I shared with Frank, I will. Into it. Uh, you'll see it in the show notes below. Uh, let us know what you're been generating with chat, g p t, Bing Chat, creator, composer, all these things. Let us know at Edward Merge Conflict fm, but this is gonna do for this week's merge conflict zone. Until next time, I'm James Mat Magno. [00:39:44] Frank: And I'm Frank Krueger. Thanks for listening. Peace.