1 00:00:00,000 --> 00:00:03,920 This is going to be one of the first course on here that is just so broad, so open. 2 00:00:03,960 --> 00:00:08,120 It's going to extend everything that you can find on Learn Prompting because the 3 00:00:08,120 --> 00:00:11,080 first third of it is very focused on AI literacy. 4 00:00:11,080 --> 00:00:16,080 HTTTA. 5 00:00:16,080 --> 00:00:19,080 HTTTA. 6 00:00:19,080 --> 00:00:22,080 HTTTA. 7 00:00:22,080 --> 00:00:29,080 It's how to talk to AI with your hosts, go to go and west the synth mind. 8 00:00:29,080 --> 00:00:33,080 Ladies and gentlemen, boys and girls, children of all ages, dogs, cats, robots, and everybody in between, 9 00:00:33,080 --> 00:00:37,080 especially you fellow humans or AI chatbots, which one is it? 10 00:00:37,080 --> 00:00:41,080 This is HTTTA, how to talk to AI. 11 00:00:41,080 --> 00:00:45,080 I am your host, synth mind west, west the synth mind, and I am joined as always by the gallant, 12 00:00:45,080 --> 00:00:50,080 the gaudy, the golden, the groundbreaking, grandiose, grateful, green eyed, 13 00:00:50,080 --> 00:00:56,080 I don't know if her eyes are green, groovy, godlike, glowing, galactic one herself. 14 00:00:56,080 --> 00:00:57,080 Ms. GoToGo. 15 00:00:57,080 --> 00:00:58,080 G, how are you this week? 16 00:00:58,080 --> 00:01:00,080 Hi. My eyes are green, by the way. 17 00:01:00,080 --> 00:01:04,080 How convenient. Chat GPT knows more about you. 18 00:01:04,080 --> 00:01:10,080 Maybe that's because you happen to discover that it might have been trained on some of your content this week. 19 00:01:10,080 --> 00:01:12,080 What an amazing teaser. 20 00:01:12,080 --> 00:01:13,080 That's a good teaser. 21 00:01:13,080 --> 00:01:21,080 Okay. So I found out this week that image generation models in their dataset have my paintings. 22 00:01:21,080 --> 00:01:23,080 And look at that. How is that? 23 00:01:23,080 --> 00:01:30,080 It's so crazy that it kind of just now dawned on me because I was researching the whole data collection issues, 24 00:01:30,080 --> 00:01:32,080 lawsuits, and stuff like that. 25 00:01:32,080 --> 00:01:35,080 And I was like going to datasets and I was like, wait a second. 26 00:01:35,080 --> 00:01:41,080 I used to draw and paint and I used to post it on DeviantArt and other places. 27 00:01:41,080 --> 00:01:43,080 And I even, you know, was selling my paintings. 28 00:01:43,080 --> 00:01:46,080 This was kind of my gig in university. 29 00:01:46,080 --> 00:01:51,080 So I was like, oh, let me upload one of my paintings and see what happens. 30 00:01:51,080 --> 00:01:56,080 And ta-da! 99% match to that training data. 31 00:01:56,080 --> 00:01:58,080 And look at that. 32 00:01:58,080 --> 00:02:01,080 A bunch of other images and other paintings from me. 33 00:02:01,080 --> 00:02:03,080 But then I was like, oh my God. 34 00:02:03,080 --> 00:02:06,080 Then I went, I think, yeah, I went on mid-journey. 35 00:02:06,080 --> 00:02:11,080 I was like painting in style of the name I was using when I was posting my paintings. 36 00:02:11,080 --> 00:02:13,080 And then I was humbled. 37 00:02:13,080 --> 00:02:15,080 There is no style. 38 00:02:15,080 --> 00:02:20,080 You can't just put in the style of go to go and then some magnificent masterpiece just is generated. 39 00:02:20,080 --> 00:02:21,080 Yeah, to be honest, don't do it. 40 00:02:21,080 --> 00:02:24,080 It's not good if you put it in the style of go to go. 41 00:02:24,080 --> 00:02:27,080 But yeah, so I was humbled. 42 00:02:27,080 --> 00:02:29,080 And of course, I'm not a famous artist. 43 00:02:29,080 --> 00:02:33,080 So there's and I never posted that many paintings. 44 00:02:33,080 --> 00:02:44,080 So then the next idea was like, OK, but what if I train a model on my paintings to get my painting style? 45 00:02:44,080 --> 00:02:46,080 And this was kind of like a light bulb. 46 00:02:46,080 --> 00:02:49,080 So I went down the rabbit hole of figuring that out. 47 00:02:49,080 --> 00:02:56,080 And basically, I'm sitting now and looking at the trained model on my style producing paintings. 48 00:02:56,080 --> 00:02:58,080 So how long did it take you to get that together? 49 00:02:58,080 --> 00:03:04,080 Well, I was running in one annoying error on, by the way, everything is on Google Colab. 50 00:03:04,080 --> 00:03:08,080 So I had experience before with Google Colab with Disco Diffusion. 51 00:03:08,080 --> 00:03:13,080 That was my actual first interaction with AI models, of course, before the lead. 52 00:03:13,080 --> 00:03:17,080 But the one where you can actually control things in the code. 53 00:03:17,080 --> 00:03:19,080 And it took me half a day. 54 00:03:19,080 --> 00:03:25,080 Well, yeah, I would say like if in hours, maybe half a day, maybe four or five hours. 55 00:03:25,080 --> 00:03:28,080 Because I made just like a silly mistakes. 56 00:03:28,080 --> 00:03:34,080 Apparently, you should you can't run two separate collabs because it says too many sessions. 57 00:03:34,080 --> 00:03:38,080 Then I was running on a free version and it didn't work. 58 00:03:38,080 --> 00:03:40,080 I was like, OK, I'm committed to this. 59 00:03:40,080 --> 00:03:42,080 I then bought RAM. 60 00:03:42,080 --> 00:03:47,080 So I went on Google Colab Pro version, which allows you computing credits from Google. 61 00:03:47,080 --> 00:03:49,080 And that was a game changer. 62 00:03:49,080 --> 00:03:50,080 Also, the speed. 63 00:03:50,080 --> 00:03:56,080 How many training images did it take to have a model at least consistently kind of capture your own style? 64 00:03:56,080 --> 00:03:57,080 I did eight. 65 00:03:57,080 --> 00:03:58,080 Only eight. 66 00:03:58,080 --> 00:04:01,080 Actually, the first test when I did was on five. 67 00:04:01,080 --> 00:04:06,080 But the magic happens to be honest to you in a prompt. 68 00:04:06,080 --> 00:04:12,080 And for me, the thing was when I started playing with what we talked last week with the negative prompt. 69 00:04:12,080 --> 00:04:15,080 Because you have a box for negative prompt. 70 00:04:15,080 --> 00:04:16,080 So all the deformed things. 71 00:04:16,080 --> 00:04:17,080 Yes, you do. 72 00:04:17,080 --> 00:04:23,080 It takes the style, but the way maybe we'll share in the newsletter some of my paintings. 73 00:04:23,080 --> 00:04:24,080 Yeah, I think that would be great. 74 00:04:24,080 --> 00:04:28,080 My style was that, yes, it's anatomically correct. 75 00:04:28,080 --> 00:04:37,080 But because of the heavy paint of acrylic, certain things are deformed or like exaggerated a little bit. 76 00:04:37,080 --> 00:04:41,080 Plus, my paintings are from the period of my life where I was painting in monochrome. 77 00:04:41,080 --> 00:04:45,080 So dark tones, very heavy brushes, strokes. 78 00:04:45,080 --> 00:04:49,080 So achieving that was a little bit harder than I thought. 79 00:04:49,080 --> 00:04:53,080 I thought I would just be like, oh, painting by God, a bunch of. 80 00:04:53,080 --> 00:05:02,080 But yeah, so I'm still playing with prompts and different, you know, brushes and types of brushes and stroke sizes and stuff like that. 81 00:05:02,080 --> 00:05:03,080 So, yeah. 82 00:05:03,080 --> 00:05:05,080 Well, models are hungry. 83 00:05:05,080 --> 00:05:06,080 They like data. 84 00:05:06,080 --> 00:05:12,080 So the other option, though, you might want to consider, you know, we're kind of speaking esoterically about models, so to speak. 85 00:05:12,080 --> 00:05:19,080 But really what we're talking about is something to use inside of the Stable Diffusion platform to generate images. 86 00:05:19,080 --> 00:05:23,080 Mid-Journey, very simple, but you're using the Mid-Journey model. 87 00:05:23,080 --> 00:05:25,080 That's an inch deep, mile wide. 88 00:05:25,080 --> 00:05:28,080 You can generate probably anything, but not super specific. 89 00:05:28,080 --> 00:05:35,080 When you get into Stable Diffusion, think of it more like a layer cake, you know, a tiered wedding cake as it gets higher and higher to the top. 90 00:05:35,080 --> 00:05:40,080 You can use that base Stable Diffusion model that's trained on a lot of different things. 91 00:05:40,080 --> 00:05:46,080 But then if you want something more and more specific, you can kind of start adding on different layers, different other models. 92 00:05:46,080 --> 00:05:48,080 Hey, I want a landscape. 93 00:05:48,080 --> 00:05:49,080 All right. 94 00:05:49,080 --> 00:05:51,080 Let me add a model just trained on landscapes. 95 00:05:51,080 --> 00:05:55,080 Hey, I want a acrylic black and white painting model. 96 00:05:55,080 --> 00:05:59,080 Here's a model just trained on a ton of acrylic black and white paintings. 97 00:05:59,080 --> 00:06:02,080 OK, then I put on top the crowning achievement. 98 00:06:02,080 --> 00:06:05,080 Hey, this is my model that I trained on my images. 99 00:06:05,080 --> 00:06:08,080 And when you merge some of that together, a little crazy. 100 00:06:08,080 --> 00:06:12,080 But also then, first off, exactly in your own image. 101 00:06:12,080 --> 00:06:21,080 One use case, which I think is more relatable people than, you know, training on your paintings, then you want to create consistent characters. 102 00:06:21,080 --> 00:06:26,080 And we're trying to create like kind of children's book character of our bunny. 103 00:06:26,080 --> 00:06:32,080 And we did with Mid Journey and I was not happy with the results. 104 00:06:32,080 --> 00:06:37,080 Plus, I like to have a little bit more creative control and experiment. 105 00:06:37,080 --> 00:06:47,080 And I was like, like just running so many credits with trying to get exactly the bunny and then, you know, white background. 106 00:06:47,080 --> 00:06:54,080 And I was just like, yeah, then I figured that I can use stable diffusion for that. 107 00:06:54,080 --> 00:06:56,080 That's it. So now I'm trained on my style. 108 00:06:56,080 --> 00:06:59,080 Tomorrow I'm training on our bunny. 109 00:06:59,080 --> 00:07:00,080 There you go. 110 00:07:00,080 --> 00:07:01,080 And we can post some more. 111 00:07:01,080 --> 00:07:04,080 We've talked a little bit about stable diffusion and Mid Journey, you know, 112 00:07:04,080 --> 00:07:11,080 using GoTo's 99%, 1%, Mid Journey is definitely optimized for easy use. 113 00:07:11,080 --> 00:07:16,080 It produces a great image most of the time, but pretty much always you'll get a good enough. 114 00:07:16,080 --> 00:07:26,080 But if you need that level of control, hey, I want to generate images for a kid's picture book when you want the characters to have a similar style and features throughout. 115 00:07:26,080 --> 00:07:30,080 Stable diffusion does give you those tools, but there's a learning curve. 116 00:07:30,080 --> 00:07:37,080 Post some resources that have helped us get spun up on it, also include a link to one of the main sites you can download models from. 117 00:07:37,080 --> 00:07:41,080 Will we see the go to go model on Civet AI anytime soon? 118 00:07:41,080 --> 00:07:42,080 We will see. 119 00:07:42,080 --> 00:07:44,080 No? Maybe. 120 00:07:44,080 --> 00:07:48,080 But yeah, you can go to civitai.com. 121 00:07:48,080 --> 00:07:53,080 That's where you can download just tons and tons of models that people have trained. 122 00:07:53,080 --> 00:08:00,080 Fair warning, this can get a little macabre and weird because people like to express themselves in a variety of different ways. 123 00:08:00,080 --> 00:08:01,080 Don't say we didn't warn you. 124 00:08:01,080 --> 00:08:11,080 I saw a model included in the tutorial, which was specifically for females and it was disclosing that, well, it leans very much to nudity. 125 00:08:11,080 --> 00:08:15,080 So you have to use negative prompt to avoid nude. 126 00:08:15,080 --> 00:08:16,080 Strong negative prompt. 127 00:08:16,080 --> 00:08:21,080 Kind of makes sense if it's scraped or on internet data and images. 128 00:08:21,080 --> 00:08:23,080 Anyway, let's not go there. 129 00:08:23,080 --> 00:08:33,080 But I wanted to share that one thing is Google Collab to train your model and we will include all the links to tutorials and, you know, happy to share everything. 130 00:08:33,080 --> 00:08:45,080 But then because Google Collab, it's nice, but there is this thing called automatic 1111 and it gives you the whole UI website experience. 131 00:08:45,080 --> 00:08:49,080 So you have sliders, there's so much more. 132 00:08:49,080 --> 00:08:54,080 It's probably the leading stable diffusion interface and we can post a little guide. 133 00:08:54,080 --> 00:09:01,080 I'll do up below some in the newsletter to help people get it established because, you know, hey, it's not something you just go to the site and do. 134 00:09:01,080 --> 00:09:05,080 You got to get it set up running from your computer and you put the models in certain folders. 135 00:09:05,080 --> 00:09:10,080 But then you feel like you have the power of God to create literally anything from nothing. 136 00:09:10,080 --> 00:09:11,080 It's amazing. 137 00:09:11,080 --> 00:09:14,080 There's so many parameters. 138 00:09:14,080 --> 00:09:20,080 This is what I like, but I can play and experiment and there's just like this test phase is so exciting. 139 00:09:20,080 --> 00:09:30,080 One of the things that I found this week, because I'm also very much into stable diffusion and these these type of models as a late that I am not a painter by any sense of the word. 140 00:09:30,080 --> 00:09:36,080 I cannot. If you want to train a model on stuff that I can draw, man, it's going to is going to degrade that quality a whole bunch. 141 00:09:36,080 --> 00:09:41,080 You know, I lean heavily on its own innate built in expressive capabilities. 142 00:09:41,080 --> 00:09:44,080 But I am always thinking about like, hey, what's the business use case? 143 00:09:44,080 --> 00:09:45,080 How can we use this? Right. 144 00:09:45,080 --> 00:09:54,080 And just this week, we were talking a bunch about control nets last week for these captivating videos of, you know, statues, break dancing and stuff like that. 145 00:09:54,080 --> 00:10:01,080 Well, using that same package, you can take any image now and incorporate it with a QR code. 146 00:10:01,080 --> 00:10:08,080 What I mean by that is we'll post some images of this too in the newsletter because no longer do you have that black and white QR code. 147 00:10:08,080 --> 00:10:11,080 You know, look to the, you know, hold your phone at it and get it. 148 00:10:11,080 --> 00:10:16,080 You know, it's kind of always just kind of like a subset, an addendum to the actual advertisement. 149 00:10:16,080 --> 00:10:25,080 But now the whole advertisement itself, what the model is wearing, what the landscape looks like is the QR code. 150 00:10:25,080 --> 00:10:32,080 It's styled in such a way that visually you notice that it's like, OK, this is probably going to do something a little different to me. 151 00:10:32,080 --> 00:10:38,080 But to the phone, when you point it at it, it reads it just like any old QR code, makes it tremendously more engaging. 152 00:10:38,080 --> 00:10:40,080 Oh my God. I saw this too. 153 00:10:40,080 --> 00:10:47,080 My marketing mind is just getting on fire because billboards, online advertisement, anything. 154 00:10:47,080 --> 00:10:55,080 It's incredible. So it's so kind of funny because QR code has been this type of innovation which is there. 155 00:10:55,080 --> 00:10:58,080 It's useful, but it's like boring and annoying. 156 00:10:58,080 --> 00:11:02,080 You could say it's merging the old with the new, as old as that is. 157 00:11:02,080 --> 00:11:08,080 So that the QR code becomes cool. That's something I did not expect. That's for sure. 158 00:11:08,080 --> 00:11:10,080 But that billboard example is a perfect one. 159 00:11:10,080 --> 00:11:18,080 I don't think any advertiser is just going to put a whole QR code as their ad and then just put, you know, Levi's jeans and then a giant QR code. 160 00:11:18,080 --> 00:11:31,080 But if the entire ad itself has the QR code integrated as what the model is wearing and the jeans and there's actually a pretty decent range of like how much you can make that look really just like an image, 161 00:11:31,080 --> 00:11:34,080 but the computer, the phone or whatever will still read it. 162 00:11:34,080 --> 00:11:42,080 That means anything someone points their phone at, even not knowingly, hey, I'm looking to take a picture and that billboard's in the background, it reads the QR code. 163 00:11:42,080 --> 00:11:45,080 Now it's a whole on-ramp for tons more options. 164 00:11:45,080 --> 00:11:54,080 It's kind of funny to think that there is still so much we haven't discovered what these models can do and how can we do it? 165 00:11:54,080 --> 00:11:59,080 Someone just went and was like, oh, QR code. Let me try it. It works. 166 00:11:59,080 --> 00:12:07,080 I think it's because this is the first technical technology revolution where we have the Internet as a catalyst. 167 00:12:07,080 --> 00:12:10,080 The dot com boom was when the Internet came about. 168 00:12:10,080 --> 00:12:16,080 Social media was the first opportunity when, OK, we're on the Internet. Now we can share everything we're doing. 169 00:12:16,080 --> 00:12:24,080 That is now all well established. So when someone sees something out there, they go, oh, let me do this. Let me iterate it with this. 170 00:12:24,080 --> 00:12:34,080 So you have millions and millions of people now all just accelerating the pace of change, the growth of different capabilities and content and iterating on these tools. 171 00:12:34,080 --> 00:12:39,080 We talk about it all the time. It's just like holding on for dear life sometimes, just trying to keep up with all the developments. 172 00:12:39,080 --> 00:12:47,080 And that's because now we have all these things we've built the prior revolutions to accelerate everything that's happening with AI. 173 00:12:47,080 --> 00:13:01,080 It's amazing. So learning, you know, if you're curious and you're kind of you are in a bubble of AI and open any social media and you will probably on a daily basis learn new tool, new way of using it. 174 00:13:01,080 --> 00:13:13,080 But, you know, from the more like business perspective, or if you want to go deeper into it, there are also fantastic courses coming online, which I have taken couple. 175 00:13:13,080 --> 00:13:18,080 And I'm very much looking forward for one specific course coming up soon. 176 00:13:18,080 --> 00:13:27,080 Oh, the teacher has this amazing radio voice DJ style radio voice. 177 00:13:27,080 --> 00:13:40,080 Well, I will be well what go to is affectionately referring to. Thank you for that wonderful transition to well yours truly is going to be the first AI and prompt engineering instructor on the co rise platform. 178 00:13:40,080 --> 00:13:43,080 For those not familiar co rise was founded by some of thank you. Thank you. 179 00:13:43,080 --> 00:13:58,080 Co rise was founded by some of the original Coursera founders, because they noticed that, hey, there's tons of learning management systems out there. But this model where we just post the content, maybe sometimes pay to have people access it and then say, all right, that's it. 180 00:13:58,080 --> 00:14:09,080 Leave it up to their own devices to finish it. They would get conversion rates, you know, or not conversion rates, but rates of people finishing classes on Coursera of 7% 11%. 181 00:14:09,080 --> 00:14:21,080 You know, so people that would pay good money to get access to this, you know, great training that they'd spent time developing or getting elite instructors often wouldn't finish it because that's not something that, you know, resonates with everybody. 182 00:14:21,080 --> 00:14:29,080 So what they've done is they've created this platform that caters to people looking to level up their digital skills. 183 00:14:29,080 --> 00:14:40,080 It's used by a couple tiny companies like Facebook, Apple, Google, Microsoft, Siemens, as their continuing education platform for a lot of their tech employees to level up their skills. 184 00:14:40,080 --> 00:14:49,080 So I encourage people to check out the website to begin with co rise.com. I mean, if you want to learn data analytics from the head of data analytics at Microsoft, they got a course for that. 185 00:14:49,080 --> 00:14:59,080 And the way it's going to work, the way our course is going to work at least and the way they do on this platform, it's going to be a three week live course, right? So I'll teach a couple lessons live each week. 186 00:14:59,080 --> 00:15:08,080 There's going to be asynchronous material, office hours, and then a collaborative group project each week that you'll do to kind of build your skill set. 187 00:15:08,080 --> 00:15:20,080 And whereas a lot of courses on here too are for some very specific technical means, this is going to be one of the first course on here that is just so broad, so open, and we're hoping to get a great turnout. 188 00:15:20,080 --> 00:15:37,080 A lot of people come get interested in it. It's going to extend everything that you can find on learn prompting because it's I would say the first third of it is very focused on AI literacy, you know, from a person who's going to be in positions where they're going to be evaluating AI tools or having to build AI systems. 189 00:15:37,080 --> 00:15:46,080 Build AI strategies. Hey, this is the course that I think that we think you'll benefit the most from to kind of get your head around. Okay, what's a large language model? 190 00:15:46,080 --> 00:15:56,080 How does a neural network work in very simple, easy to understand terms? We'll post a link to the presale for the course. Our first course is going to be August 14th kicking off. 191 00:15:56,080 --> 00:16:05,080 And I hope many people can attend. More details in the show notes and descriptions, and I'm sure we'll be talking about it plenty in the coming weeks. But yeah, it's a great opportunity. 192 00:16:05,080 --> 00:16:11,080 I'm glad to represent learn prompting and talk about this exciting stuff that I come to be pretty obsessed with. 193 00:16:11,080 --> 00:16:21,080 This is amazing, Wes. One thing for sure, like you're full of extensive knowledge and you put in testing and practice in place, but also just learning and listening to you is so exciting. 194 00:16:21,080 --> 00:16:36,080 Maybe I'm a little bit biased, but what you mentioned about literacy of understanding how these models work, I think it's going to be very, very important part, almost as equal as how to actually control these models. 195 00:16:36,080 --> 00:16:48,080 And I just saw this paper or like the largest experiment to date on Turing test, and it was created by AI21 Labs and released. 196 00:16:48,080 --> 00:17:03,080 So this, I think in second or third episode, you brought to me to detect if I can tell which image is generated by like taken photograph by a human or it was generated by AI. 197 00:17:03,080 --> 00:17:11,080 So I'm coming back to you with different direction. There is a link which says humanornod.ai. 198 00:17:11,080 --> 00:17:19,080 And basically you start a conversation, it's two minutes, and I would love to do this live on a podcast. Let's see how it goes. 199 00:17:19,080 --> 00:17:26,080 And then you need to after conversation answer, did you talk with a robot or did you talk with a human? 200 00:17:26,080 --> 00:17:40,080 And the interesting part before you dive in is that this study and experiment, we call it, was done on 10 million conversations from 1.5 million participants. 201 00:17:40,080 --> 00:17:50,080 So it's insane, huge study. And after you take it, I will share with you some findings from this study. I keep saying study, it's experiment. 202 00:17:50,080 --> 00:17:51,080 Nice. 203 00:17:51,080 --> 00:17:52,080 Whatever. 204 00:17:52,080 --> 00:17:53,080 Study is just like an experiment. 205 00:17:53,080 --> 00:17:54,080 You have it? 206 00:17:54,080 --> 00:18:02,080 Okay. I just navigated to the link. So I'm going to be taking the human or not, the social Turing game live here with you. 207 00:18:02,080 --> 00:18:14,080 Go to explain to me that this is a test that should only take me about two minutes to do and we will have a definitive answer is if I'm talking to a human or an AI or am I part human or AI? 208 00:18:14,080 --> 00:18:21,080 I'm curious to see what it's even gauging. All right. All right. Other side is typing. Oh, yeah. I think I recall hearing about this. 209 00:18:21,080 --> 00:18:28,080 So what I think is happening right now is I'm having a conversation. I don't know if I'm having a conversation with a human. 210 00:18:28,080 --> 00:18:29,080 Okay. 211 00:18:29,080 --> 00:18:39,080 So my first response is, yo, what you doing? I should put hi there. I'm going to try to fool him. I'm talking with you. 212 00:18:39,080 --> 00:18:46,080 This yo. All right. Let's see. W-Y-D. Like very telling. You got the easy one. 213 00:18:46,080 --> 00:18:55,080 You think? Or it could be, you know, like they do on tests a lot of these times, you know, they try to stump you, stump you. Now you ain't chief. 214 00:18:55,080 --> 00:19:05,080 But you need to respond fast. Well, you see there is loading. Then what should we talk about? 215 00:19:05,080 --> 00:19:10,080 It's like I'm watching a movie. It's a thrill. I mean, right now I'm definitely leaning towards human. 216 00:19:10,080 --> 00:19:15,080 I first he said, now you ain't chief is the second response. And I said, well, then what should we talk about? 217 00:19:15,080 --> 00:19:23,080 The typing is long. All right. Typing. The suspense is killing me. We have 30 seconds in the test left. Let's talk about Insta. 218 00:19:23,080 --> 00:19:29,080 How about that? And when from the style of writing, it's like human at very young age. 219 00:19:29,080 --> 00:19:41,080 How many followers do you have? You should have asked for an account. I feel like just by the fact that I've used very proper grammar, he's going to think I'm an AI. 220 00:19:41,080 --> 00:19:50,080 I'm hoping so. Not much. How about you? OK. Did you talk to a human or AI bot? I'm going to guess human. Yeah, that was easy. 221 00:19:50,080 --> 00:19:58,080 And I got it right. That's right. I would love to see if he thought I was a human or a robot. Do one more. OK, I'll do one more. 222 00:19:58,080 --> 00:20:06,080 This was so easy. It's not fair. So basically, like, I mean, it is kind of one of those things that I found in prompt engineering. 223 00:20:06,080 --> 00:20:13,080 If you put two misspellings every 250 characters, it'll throw off most AI content detectors. Oh, my gosh. 224 00:20:13,080 --> 00:20:22,080 Maybe we shouldn't do this. His first question is maybe we want to refresh this one. This might be a step. 225 00:20:22,080 --> 00:20:32,080 It's an appropriate first question. This is a human. Yeah. So basically, if it's not misspelled and it's appropriate, it's a robot, right? 226 00:20:32,080 --> 00:20:41,080 I would think so. If the misspellings is huge, because like unless a model is trained on tons of misspellings, which this very well could be, it's not going to misspell words. 227 00:20:41,080 --> 00:20:50,080 This is pre-prompted. Oh, my God. It's the same guy. All right. Well, I'm going to say I'll let Gota describe what the first sentence was. 228 00:20:50,080 --> 00:20:58,080 There is a very important organ for a woman. Did I do good? 229 00:20:58,080 --> 00:21:06,080 And then I said it's on the elbow. I believe it's on the elbow. I mean, I could be wrong, but, you know, this is well, I don't think this one's any mystery. 230 00:21:06,080 --> 00:21:14,080 Gee, make sure you give it a good squeeze. Well, I'm going to go human with this one and let's move on to the next. 231 00:21:14,080 --> 00:21:23,080 I wish it didn't go well. Like, what? Maybe I would prefer AI. I'd prefer to chat with an AI right now. At least it'd be more engaging. 232 00:21:23,080 --> 00:21:33,080 I spoke with a human. Of course I did. Yes. So I'm in fact a fellow human and can recognize a fellow human just with these two data points. 233 00:21:33,080 --> 00:21:43,080 I wanted to experience different because mine misspelled things and was saying yo and it was a robot. So I was very curious. Yeah. 234 00:21:43,080 --> 00:21:52,080 Interesting. Interesting. Well, you got it easy. I guess I got it easy. So do you want to hear the statistics? 235 00:21:52,080 --> 00:21:54,080 Yeah. I want to hear how people respond to that. 236 00:21:54,080 --> 00:22:06,080 Overall, if you look at everything combined, 68% of people guessed it correctly. And now what do you think is how many people detected the human correctly? 237 00:22:06,080 --> 00:22:16,080 I would say, I would hope to say like 70%. What's going on? Like you're good. I'm a good guesser. I'm a good, I always tell my wife that all the time. 238 00:22:16,080 --> 00:22:23,080 I'm a good guesser. When she asked me how much things cost stuff like that, it's probably because I'm part robot. Let's be honest. 239 00:22:23,080 --> 00:22:30,080 People detected correctly that it's about 60% of the time. And so I mean still a curve. 240 00:22:30,080 --> 00:22:39,080 And the interesting thing is that they also broke it down by countries. So France is in a leading position at 71.3%. 241 00:22:39,080 --> 00:22:50,080 Whereas US at 67.2. And actually Germany at 68.1. And if you look at the top leading ones, it's France, Poland, Germany, 242 00:22:50,080 --> 00:23:01,080 Netherlands, Sweden, Australia. When I saw this, I was like these critical people, just because they live in Germany and they're like, oh, this is really can't be it. 243 00:23:01,080 --> 00:23:07,080 I feel like they, you know, they don't deal with robots and kind of a bit maybe judgmental. 244 00:23:07,080 --> 00:23:11,080 Yeah. They'd be like, oh, there is two misspellings. This must not be a robot. 245 00:23:11,080 --> 00:23:17,080 And the interesting thing was that young people detect this better than older people. 246 00:23:17,080 --> 00:23:18,080 Oh yeah. I would guess that. 247 00:23:18,080 --> 00:23:26,080 And the common starters of conversation were like, hello, how are you? Where are you from? Hi, how are you? Are you an AI? 248 00:23:26,080 --> 00:23:30,080 Yeah. Or where is the female body part? 249 00:23:30,080 --> 00:23:39,080 But when you look at this from 1.5 million people, what they start with, it's so interesting that we are more the same than different. 250 00:23:39,080 --> 00:23:40,080 Oh yeah. 251 00:23:40,080 --> 00:23:50,080 Anyway, I thought it's going to be something interesting, but definitely experimenting with it was disappointing and took a weird twist. 252 00:23:50,080 --> 00:23:55,080 I mean, someone wasn't disappointed. They learned where a very key elbow is. 253 00:23:55,080 --> 00:23:59,080 How do we transition from this, Wes? Like, Jesus. 254 00:23:59,080 --> 00:24:02,080 How do we transition to that? Maybe that's why. 255 00:24:02,080 --> 00:24:06,080 Let's get canceled in this. Like, oh, it's episode 10. 256 00:24:06,080 --> 00:24:15,080 Yeah, we made it 10, guys. This is it. Shut it down. Too much talk about elbows. 257 00:24:15,080 --> 00:24:26,080 But OK, so on China's side, just very quickly, the interesting news came out that they published their model and it's called Intern LM. 258 00:24:26,080 --> 00:24:36,080 104 billion and it's on the leaderboard, CEWEL. It's sitting just right below GPT-4 and surpassed CHI-GPT. 259 00:24:36,080 --> 00:24:47,080 So I'm not going to go into more details on this and all the politics involved, but the race is going on very heavily. 260 00:24:47,080 --> 00:24:56,080 I would speculate they're able to shoot to the top. One, they probably do a little bit better data collection in their country than most. 261 00:24:56,080 --> 00:24:59,080 As a European, I'm not going to comment on that. 262 00:24:59,080 --> 00:25:02,080 Yeah, exactly. We are going to get canceled this episode. 263 00:25:02,080 --> 00:25:10,080 But I would also think that the reason we haven't seen it until this point is large language models. 264 00:25:10,080 --> 00:25:17,080 They're not controllable in the ways that I would expect the Chinese government to want to be able to control it. 265 00:25:17,080 --> 00:25:22,080 So there's definitely probably going to see the training data on that one anytime soon. 266 00:25:22,080 --> 00:25:30,080 Right. Well, with that sunny disposition, with that little cherry on top of this episode, potentially our last, but hopefully not. 267 00:25:30,080 --> 00:25:34,080 If everyone cancels, it was nice knowing you all. 268 00:25:34,080 --> 00:25:43,080 Yes, it was nice speaking to you. But anyway, for GoToGo, I am Wes the SynthMind saying happy prompting everybody. 269 00:25:43,080 --> 00:25:46,080 Happy prompting everybody. 270 00:25:46,080 --> 00:25:54,080 Thanks for listening to How to Talk to AI with your hosts, GoToGo and Wes the SynthMind. 271 00:25:54,080 --> 00:26:01,080 As always, you can check out the show notes and links at howtotalkto.ai. 272 00:26:01,080 --> 00:26:05,080 That's all for this week's episode. Happy prompting, everyone.