1 00:00:00,000 --> 00:00:10,000 Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the 2 00:00:10,000 --> 00:00:15,740 latest trends, tips, and tools in the world of marketing AI, helping you get the best 3 00:00:15,740 --> 00:00:18,560 results from your marketing efforts. 4 00:00:18,560 --> 00:00:23,060 Now let's join our hosts, Paul Avery and Martin Broadhurst. 5 00:00:23,060 --> 00:00:27,600 Welcome to episode 10 of Artificially Intelligent Marketing. 6 00:00:27,600 --> 00:00:31,440 We're glad to have you here where myself and my good friend Martin Broadhurst. 7 00:00:31,440 --> 00:00:32,440 Hi there, Martin. 8 00:00:32,440 --> 00:00:33,440 Hello, Paul. 9 00:00:33,440 --> 00:00:34,440 10 episodes. 10 00:00:34,440 --> 00:00:35,440 Wow, we've done well. 11 00:00:35,440 --> 00:00:36,440 We stuck at it. 12 00:00:36,440 --> 00:00:37,440 Woop, woop. 13 00:00:37,440 --> 00:00:39,720 The audience are like, oh God, they stuck at it. 14 00:00:39,720 --> 00:00:40,960 But hey, you're still here too. 15 00:00:40,960 --> 00:00:41,960 We love you for it. 16 00:00:41,960 --> 00:00:48,440 We're going to be diving into this week's very cool AI news and interesting topics with 17 00:00:48,440 --> 00:00:52,680 all of the lowdown on what it means for marketers and what they need to know. 18 00:00:52,680 --> 00:00:56,120 This is a chunky one this week, isn't it, Martin? 19 00:00:56,120 --> 00:01:00,360 We have a bunch of really interesting stories and a ton of short snippets. 20 00:01:00,360 --> 00:01:05,000 So this week for the bigger stories, we're going to focus on Google's I.O. conference, 21 00:01:05,000 --> 00:01:11,200 which was full of a ton of updates, many of them related to AI. 22 00:01:11,200 --> 00:01:17,160 And I read a few things online implying that Google had gone from massively behind the 23 00:01:17,160 --> 00:01:19,200 curve to suddenly in front of the curve. 24 00:01:19,200 --> 00:01:24,400 So maybe we can get into detail about whether that's hyperbole or if it's a fair reflection, 25 00:01:24,400 --> 00:01:25,480 Martin. 26 00:01:25,480 --> 00:01:30,520 We're also going to talk a little bit about Anthropix, LLM Claude, which they've expanded 27 00:01:30,520 --> 00:01:32,840 the capabilities of and Martin's been playing with that. 28 00:01:32,840 --> 00:01:34,760 So I think it'll be interesting to go into that. 29 00:01:34,760 --> 00:01:41,080 We're going to talk about Meta's announcement of their new generative AI tools for advertisers. 30 00:01:41,080 --> 00:01:45,440 We're going to talk about Stability AI's new text animation tool. 31 00:01:45,440 --> 00:01:51,480 And we're going to talk about some news from Microsoft about the chatbot ads are coming. 32 00:01:51,480 --> 00:01:54,560 So they're going to be like the sort of main stories of the week. 33 00:01:54,560 --> 00:01:59,000 And then in terms of the tools of the week, I've been lucky enough to get onto the Firefly 34 00:01:59,000 --> 00:02:00,440 how to be Firefly beta. 35 00:02:00,440 --> 00:02:04,960 And I've been having a little play with that and I'll share my experiences in terms of 36 00:02:04,960 --> 00:02:08,600 the short snippets, the things we're going to very briefly mention, but we won't have 37 00:02:08,600 --> 00:02:10,760 time to get into detail on. 38 00:02:10,760 --> 00:02:13,080 Take a deep breath because there's a lot here. 39 00:02:13,080 --> 00:02:19,600 There was this week an article in Bloomberg reporting that BuzzFeed readers spend 40 percent 40 00:02:19,600 --> 00:02:25,400 more time on AI generated quizzes compared to the traditional ones that have been created 41 00:02:25,400 --> 00:02:26,400 by humans. 42 00:02:26,400 --> 00:02:30,520 I haven't been able to go into this story in depth, but it'd be really interesting to 43 00:02:30,520 --> 00:02:33,360 know what has driven that. 44 00:02:33,360 --> 00:02:34,360 So I think that's one we'll probably follow up on. 45 00:02:34,360 --> 00:02:37,320 I'd actually forgotten BuzzFeed was still a thing. 46 00:02:37,320 --> 00:02:38,360 Well there you go. 47 00:02:38,360 --> 00:02:39,360 And here we are. 48 00:02:39,360 --> 00:02:40,360 There we are. 49 00:02:40,360 --> 00:02:41,360 There you are. 50 00:02:41,360 --> 00:02:44,320 So I think that's that could be interesting. 51 00:02:44,320 --> 00:02:49,000 And there's a number of ways I could imagine AI can use insights about the successes of 52 00:02:49,000 --> 00:02:54,400 previous quizzes to be able to optimize quizzes for humans in a way that us poor humans have 53 00:02:54,400 --> 00:02:59,960 to somewhat follow our experience and gut as opposed to data gleaned from a million 54 00:02:59,960 --> 00:03:04,120 other quizzes, which they would perhaps be able to do. 55 00:03:04,120 --> 00:03:10,680 There was a research report where OpenAI used GPT-4 to try and understand how GPT works. 56 00:03:10,680 --> 00:03:13,880 If that's making your brain hurt a bit, you're not the only one. 57 00:03:13,880 --> 00:03:20,520 In essence, there's a lot of discussion about how large language models, especially the 58 00:03:20,520 --> 00:03:25,600 most new ones, the largest ones built on a trillion parameters, that we don't really 59 00:03:25,600 --> 00:03:27,320 know how they work. 60 00:03:27,320 --> 00:03:32,400 We know the fundamentals, but we can't explain after the fact how they generated a particular 61 00:03:32,400 --> 00:03:33,480 output. 62 00:03:33,480 --> 00:03:39,280 And so what OpenAI tried to do is try to use GPT-4 to try and make some predictions about 63 00:03:39,280 --> 00:03:45,280 what a particular model would output at scale and got particularly close in a number of 64 00:03:45,280 --> 00:03:46,280 areas. 65 00:03:46,280 --> 00:03:51,480 So I think there was limitations in the research, but it could help us to identify how the models 66 00:03:51,480 --> 00:03:57,360 behave, which could become important when we're thinking about things like safety and 67 00:03:57,360 --> 00:03:59,440 alignment issues, which we've talked about before. 68 00:03:59,440 --> 00:04:03,720 Yeah, explainability is going to be so important in this area going forward. 69 00:04:03,720 --> 00:04:10,040 And I think a lot of that you'll see only becoming more important with the expectations 70 00:04:10,040 --> 00:04:11,040 of regulators. 71 00:04:11,040 --> 00:04:12,040 Absolutely. 72 00:04:12,040 --> 00:04:14,680 So you can see why they commissioned this research. 73 00:04:14,680 --> 00:04:19,880 Obviously explaining how GPT-2 works compared to GPT-4 is going to be a lot easier given 74 00:04:19,880 --> 00:04:23,600 the size of the training data set and how the model works and all that other stuff. 75 00:04:23,600 --> 00:04:25,480 But there you go. 76 00:04:25,480 --> 00:04:29,800 Wendy's in the US is working on an AI chat bot that's being developed in collaboration 77 00:04:29,800 --> 00:04:35,520 with Google to service customers at its drive-throughs, which is quite interesting. 78 00:04:35,520 --> 00:04:39,560 I mean, we've seen a lot of automation in the fast food space over the last couple of 79 00:04:39,560 --> 00:04:40,560 years as it is. 80 00:04:40,560 --> 00:04:45,480 And here's another potential job within the fast food industry that we might see supplanted 81 00:04:45,480 --> 00:04:50,880 very human no longer manages that interaction, but a chat bot does instead. 82 00:04:50,880 --> 00:04:52,920 Interesting. 83 00:04:52,920 --> 00:04:57,080 We've got mid journey 5.1 is already here and producing much higher quality images, 84 00:04:57,080 --> 00:05:00,200 especially when you're working with shorter prompts. 85 00:05:00,200 --> 00:05:03,840 And I don't know about you, Mime, but I'm playing with lots of different image generation 86 00:05:03,840 --> 00:05:08,820 tools at the moment, but it's hard to look beyond mid journey in terms of just quality 87 00:05:08,820 --> 00:05:09,820 of output. 88 00:05:09,820 --> 00:05:10,820 Yeah. 89 00:05:10,820 --> 00:05:19,040 I was looking at a comparison of mid journey 3 to mid journey 5.1, same prompt and just 90 00:05:19,040 --> 00:05:20,040 looking at the outputs. 91 00:05:20,040 --> 00:05:25,720 And it's, I mean, you would have thought there was five, six, seven years worth of research 92 00:05:25,720 --> 00:05:27,000 difference between them. 93 00:05:27,000 --> 00:05:29,200 Add it barely 12 months. 94 00:05:29,200 --> 00:05:31,520 Yeah, that's kind of insane. 95 00:05:31,520 --> 00:05:37,240 So I think mid journey is proving at least for us, the tool of choice that we are playing 96 00:05:37,240 --> 00:05:43,000 with at the moment, because it seems to be the most capable when it comes to producing 97 00:05:43,000 --> 00:05:46,120 really quality polished outputs. 98 00:05:46,120 --> 00:05:48,160 There was a leaked memo. 99 00:05:48,160 --> 00:05:52,960 In fact, this is after we launched the podcast about an hour after we went live last week. 100 00:05:52,960 --> 00:05:57,880 There's old news now, seven days behind, old news about a leaked memo from Google where 101 00:05:57,880 --> 00:06:03,680 an internal employee, supposedly an internal employee, we should say, questions the ability 102 00:06:03,680 --> 00:06:08,640 of large companies like Google and OpenAI to be able to innovate and differentiate their 103 00:06:08,640 --> 00:06:14,080 products faster than those working on open source alternatives. 104 00:06:14,080 --> 00:06:19,960 For a huge amount of reasons, not least just innovative ideas and bandwidth and the number 105 00:06:19,960 --> 00:06:24,240 of people that can tinker with open source tools and take them forward. 106 00:06:24,240 --> 00:06:29,360 They gave some specific examples in there, didn't they, about being able to run the open 107 00:06:29,360 --> 00:06:35,840 source models on smartphones already, albeit very slowly at five tokens per minute. 108 00:06:35,840 --> 00:06:40,280 But just talking about the level of optimization that you can get when you take an open source 109 00:06:40,280 --> 00:06:44,680 model and fine tune it and optimize it to these specific use cases. 110 00:06:44,680 --> 00:06:49,160 And the language that they used in the memo as well, it's quite interesting how they said, 111 00:06:49,160 --> 00:06:54,000 we do not have a moat when it comes to AI and nor does OpenAI. 112 00:06:54,000 --> 00:06:59,680 However, I have thoughts which will come onto in the Google I.O. story. 113 00:06:59,680 --> 00:07:00,680 Yeah, yeah. 114 00:07:00,680 --> 00:07:07,000 I don't quite agree with that moat statement because I think having an installed user base 115 00:07:07,000 --> 00:07:11,800 using Google apps or Microsoft based apps and hundreds of millions is a fairly significant 116 00:07:11,800 --> 00:07:12,800 moat. 117 00:07:12,800 --> 00:07:14,040 Yeah, that moat seems quite wide. 118 00:07:14,040 --> 00:07:15,040 Yeah. 119 00:07:15,040 --> 00:07:20,080 I think there's a commercial moat there that just may not be an innovation moat and product 120 00:07:20,080 --> 00:07:22,280 development moat. 121 00:07:22,280 --> 00:07:26,840 And the point you made is a really good one, Martin, in terms of open source tinkerers finding 122 00:07:26,840 --> 00:07:34,160 ways to train models at a vastly reduced cost and much faster than some of the models that 123 00:07:34,160 --> 00:07:37,520 are coming out of the bigger players. 124 00:07:37,520 --> 00:07:42,120 But actually without any real dropping quality in the output of the large language models, 125 00:07:42,120 --> 00:07:48,080 like a chat bot that's been trained on far less data for $500 over the space of a week 126 00:07:48,080 --> 00:07:50,880 performs not that far off GPT-4. 127 00:07:50,880 --> 00:07:54,040 And I think that's a major thing that came up in that. 128 00:07:54,040 --> 00:07:56,840 If you're interested, find that memo online. 129 00:07:56,840 --> 00:08:01,360 It's absolutely full of zingers in terms of, I think, insights into what it's like to be 130 00:08:01,360 --> 00:08:08,400 on the inside of developing these models and how the Ecosystem is looking. 131 00:08:08,400 --> 00:08:11,240 So I think, yeah, definitely worth looking into that. 132 00:08:11,240 --> 00:08:13,960 And then the last one is actually just a silly thing. 133 00:08:13,960 --> 00:08:16,960 What I say is a silly thing, it could well be real. 134 00:08:16,960 --> 00:08:21,520 I learned this week that someone's building an app where they're trying to train the model, 135 00:08:21,520 --> 00:08:27,040 a text to bark model, so that you can talk to your dog. 136 00:08:27,040 --> 00:08:28,680 I know that sounds insane. 137 00:08:28,680 --> 00:08:34,000 The website looks like really professional, like it's not an April Fools joke, which is 138 00:08:34,000 --> 00:08:37,640 what I first thought it must be, something like that. 139 00:08:37,640 --> 00:08:38,640 I've signed up. 140 00:08:38,640 --> 00:08:43,280 I have a little puppy, a little Whippet, so I'll be attempting to hold deep conversations 141 00:08:43,280 --> 00:08:46,880 about large language models with my Whippet. 142 00:08:46,880 --> 00:08:48,680 They must be barking mad, aren't they? 143 00:08:48,680 --> 00:08:49,680 Fine. 144 00:08:49,680 --> 00:08:50,680 Boom, boom. 145 00:08:50,680 --> 00:08:55,000 Yeah, you've got to think they're barking up the wrong tree with that one. 146 00:08:55,000 --> 00:08:57,800 It certainly gives one pause for thought, I would say. 147 00:08:57,800 --> 00:09:00,480 Oh, we could go all week, but we won't. 148 00:09:00,480 --> 00:09:01,480 We absolutely should not. 149 00:09:01,480 --> 00:09:03,480 There is too much news there. 150 00:09:03,480 --> 00:09:04,480 Unsubscribed. 151 00:09:04,480 --> 00:09:05,480 Sorry about that. 152 00:09:05,480 --> 00:09:09,040 No more puns for the rest of the session, we promise. 153 00:09:09,040 --> 00:09:11,440 Right, so that's your short snippets. 154 00:09:11,440 --> 00:09:13,200 Let's get into these big old chunky stories. 155 00:09:13,200 --> 00:09:16,480 Martin, tell us about the Google I.O. conference. 156 00:09:16,480 --> 00:09:24,360 Google I.O., the annual conference where Google announces all things new for the next 12 months 157 00:09:24,360 --> 00:09:26,280 in Google land, really. 158 00:09:26,280 --> 00:09:31,320 It covers everything from search to hardware to Chrome. 159 00:09:31,320 --> 00:09:33,760 You name it, they are talking about it. 160 00:09:33,760 --> 00:09:38,200 This year, the theme was very squarely AI. 161 00:09:38,200 --> 00:09:47,560 In fact, I saw one AI word count that said in the keynote, they used the term AI 140 162 00:09:47,560 --> 00:09:49,000 times throughout the keynote. 163 00:09:49,000 --> 00:09:52,920 So it featured front and center. 164 00:09:52,920 --> 00:09:56,080 A few headline pieces. 165 00:09:56,080 --> 00:10:03,360 One was that Palm II, their large language model, their foundational large language model, 166 00:10:03,360 --> 00:10:11,280 has been basically baked in to the, or is going to be baked into all the Google products 167 00:10:11,280 --> 00:10:13,240 coming soon. 168 00:10:13,240 --> 00:10:18,040 If you're using Bard, that had a major upgrade on the day of the event. 169 00:10:18,040 --> 00:10:22,240 So they took it from Palm to Palm II. 170 00:10:22,240 --> 00:10:28,300 And in the coming weeks and months, we can expect Bard to become multimodal with image 171 00:10:28,300 --> 00:10:30,920 generation capabilities. 172 00:10:30,920 --> 00:10:35,200 They also upgraded the code abilities of Bard as well. 173 00:10:35,200 --> 00:10:38,760 So a couple of details about the Palm II model. 174 00:10:38,760 --> 00:10:44,520 It's a 540 billion parameter model. 175 00:10:44,520 --> 00:10:50,880 It uses a compute optimal scaling, which means it scales the model size and the training 176 00:10:50,880 --> 00:10:57,620 data set size in proportion to each other, which makes Palm II smaller and more efficient 177 00:10:57,620 --> 00:11:00,880 than Palm I. 178 00:11:00,880 --> 00:11:07,640 It's also used in some of their state of the art domain specific large language models, 179 00:11:07,640 --> 00:11:12,740 such as MedPalm II and SecPalm II. 180 00:11:12,740 --> 00:11:16,360 And that just a few, they mentioned a few interesting things on MedPalm. 181 00:11:16,360 --> 00:11:24,760 So MedPalm is the medical LLM, the kind of domain specific, in the medical domain, should 182 00:11:24,760 --> 00:11:26,480 I say. 183 00:11:26,480 --> 00:11:34,560 And they said that MedPalm II is the first large language model to perform at an expert 184 00:11:34,560 --> 00:11:42,800 test taker level on the MedQA data set of the US medical licensing examination, reaching 185 00:11:42,800 --> 00:11:46,080 85% plus accuracy. 186 00:11:46,080 --> 00:11:53,880 And it's the first ARA system to reach a passing score on a couple of other exams. 187 00:11:53,880 --> 00:12:03,880 So the NED MCQA data set and the NEAT medical examination question scoring 72.3%. 188 00:12:03,880 --> 00:12:14,240 In fact, the BBC had an article covering this story saying, meet MedPalm II, the AI to replace 189 00:12:14,240 --> 00:12:17,640 your GP, which I thought was an interesting headline. 190 00:12:17,640 --> 00:12:28,920 But it goes to show that it's creeping into that mainstream worldview that this is coming. 191 00:12:28,920 --> 00:12:31,480 This is potentially down the track. 192 00:12:31,480 --> 00:12:38,480 Clearly, it's not going to replace the GP, but these tools are going to be used in diagnostics 193 00:12:38,480 --> 00:12:39,760 at some point soon. 194 00:12:39,760 --> 00:12:41,240 We've spoke about it before. 195 00:12:41,240 --> 00:12:46,640 But yeah, the main takeaway here was that Palm II is being baked into all of the products. 196 00:12:46,640 --> 00:12:53,000 So you're going to see this being rolled out into search, into Google Workspace, Gmail, 197 00:12:53,000 --> 00:12:54,000 and more. 198 00:12:54,000 --> 00:12:55,760 There are a couple of other cool demos as well. 199 00:12:55,760 --> 00:13:00,840 I don't know if you managed to catch any of this, but the Google Maps fly through the 200 00:13:00,840 --> 00:13:02,120 city. 201 00:13:02,120 --> 00:13:04,800 That was a very cool demo. 202 00:13:04,800 --> 00:13:07,960 Basically, they had... 203 00:13:07,960 --> 00:13:12,280 When you do directions on Google Maps, and normally we get this top down view with a 204 00:13:12,280 --> 00:13:24,640 blue line, well now you will get a 3D rendered map view flying through the city streets with 205 00:13:24,640 --> 00:13:32,240 real time weather updates or weather simulations built into that model as well, all generated 206 00:13:32,240 --> 00:13:35,640 through AI imaging of the big cities. 207 00:13:35,640 --> 00:13:40,560 Now that's going to be rolled out across New York, LA, London, Tokyo, etc., all the big 208 00:13:40,560 --> 00:13:43,200 cities in the coming months. 209 00:13:43,200 --> 00:13:44,200 That looked very cool. 210 00:13:44,200 --> 00:13:48,440 They've also got something about video translation, lip-sync. 211 00:13:48,440 --> 00:13:55,080 So it dubs videos for international markets and it does the lip-syncing in time with it 212 00:13:55,080 --> 00:13:57,340 as well. 213 00:13:57,340 --> 00:14:03,320 That one obviously causes people to prickle a little bit because of the potential for 214 00:14:03,320 --> 00:14:07,120 deep fake issues with that one. 215 00:14:07,120 --> 00:14:10,720 They did say that they're only releasing that technology to select partners. 216 00:14:10,720 --> 00:14:17,400 So you can imagine that being rolled out to the likes of Khan Academy, Coursera, those 217 00:14:17,400 --> 00:14:18,400 kinds of institutions. 218 00:14:18,400 --> 00:14:19,840 Do you know the example I saw? 219 00:14:19,840 --> 00:14:21,360 I saw this a while ago. 220 00:14:21,360 --> 00:14:25,040 I don't know if it was Google's technology, a different one, and I don't know what the 221 00:14:25,040 --> 00:14:26,040 movie was. 222 00:14:26,040 --> 00:14:32,080 It was the one where two characters get stuck at the top of a very tall pole and they're 223 00:14:32,080 --> 00:14:33,080 going to fall off. 224 00:14:33,080 --> 00:14:34,320 It's like a drama over 90 minutes. 225 00:14:34,320 --> 00:14:37,680 It was in the cinemas, so it's probably a famous movie and I'm probably doing a terrible 226 00:14:37,680 --> 00:14:39,640 job of explaining it. 227 00:14:39,640 --> 00:14:44,840 But the example I saw, they overdubbed it with a Spanish voice and lip-synced it. 228 00:14:44,840 --> 00:14:51,660 It's a US film with US actors in it and it looked like it had originally been recorded 229 00:14:51,660 --> 00:14:53,840 in Spanish. 230 00:14:53,840 --> 00:14:58,280 Same actors, just their lips moving in perfect time to perfect Spanish. 231 00:14:58,280 --> 00:15:03,680 I assume the Spanish was delivered by voice actors, but the lip-syncing was done by AI 232 00:15:03,680 --> 00:15:04,960 and it looked incredible. 233 00:15:04,960 --> 00:15:06,600 Yeah, yeah. 234 00:15:06,600 --> 00:15:10,720 The example that they showed in the demo the other day was very solid as well. 235 00:15:10,720 --> 00:15:15,000 So this is going to be an interesting one to see how. 236 00:15:15,000 --> 00:15:16,720 I want to see who they give it to. 237 00:15:16,720 --> 00:15:18,720 Who is this technology being deployed to? 238 00:15:18,720 --> 00:15:21,300 Because they're clearly not going to make this available to everyone. 239 00:15:21,300 --> 00:15:22,800 They were very big on safety. 240 00:15:22,800 --> 00:15:27,320 That's one of the big messages that came through throughout on the keynote is that Google is 241 00:15:27,320 --> 00:15:30,200 taking AI safety and responsibility very seriously. 242 00:15:30,200 --> 00:15:33,480 They want to reassure you that that is all in hand. 243 00:15:33,480 --> 00:15:37,920 I would think the artificially intelligent marketing podcasts are probably pretty high 244 00:15:37,920 --> 00:15:43,360 on their list of priorities in getting the video version of this lip-synced in another 245 00:15:43,360 --> 00:15:46,760 40 languages and helping us bring this to the world. 246 00:15:46,760 --> 00:15:51,640 Well, if anything, you say them helping us bring it to the world, if anything, we'd be 247 00:15:51,640 --> 00:15:55,640 helping them bring their model to the world. 248 00:15:55,640 --> 00:15:58,080 I think that's how they should see it. 249 00:15:58,080 --> 00:15:59,520 What a watertight argument. 250 00:15:59,520 --> 00:16:00,520 Absolutely. 251 00:16:00,520 --> 00:16:05,720 We'll get on the old emails and we'll get that all over to them. 252 00:16:05,720 --> 00:16:07,400 Sundar, please give us access. 253 00:16:07,400 --> 00:16:09,400 In fact, just give us access, not please. 254 00:16:09,400 --> 00:16:10,400 No please. 255 00:16:10,400 --> 00:16:13,240 For your own good. 256 00:16:13,240 --> 00:16:15,480 That's quite enough nonsense. 257 00:16:15,480 --> 00:16:17,640 There was a few other cool things around this, wasn't there? 258 00:16:17,640 --> 00:16:21,320 Like this generative AI everywhere concept. 259 00:16:21,320 --> 00:16:23,560 Tell us about that, Mark. 260 00:16:23,560 --> 00:16:34,080 So this is basically the palm to being deployed in every aspect of the Google interface. 261 00:16:34,080 --> 00:16:42,720 In the demo, they showed search, Gmail, they talked about... 262 00:16:42,720 --> 00:16:49,480 So they had a Google Sheets example where you were in Google Sheets and you could just 263 00:16:49,480 --> 00:16:59,540 ask it to... if you were a dog walker and you needed to manage your dog walking clients, 264 00:16:59,540 --> 00:17:05,960 you could just say, create a table to help me manage my dog walking clients. 265 00:17:05,960 --> 00:17:14,760 And it would then in Sheets create a table with columns such as client name, dog's name, 266 00:17:14,760 --> 00:17:22,960 speed, frequency of walks, cost per hour, all that just done for you in an instant because 267 00:17:22,960 --> 00:17:28,080 they had trained this on millions of spreadsheets. 268 00:17:28,080 --> 00:17:29,680 So it just knows and it can do that. 269 00:17:29,680 --> 00:17:31,320 So that was pretty cool. 270 00:17:31,320 --> 00:17:35,720 Although I did see some responses of people on Reddit saying, oh no, now they're going... 271 00:17:35,720 --> 00:17:41,080 I think they actually, the term that they used in the demo was we're going to free you 272 00:17:41,080 --> 00:17:46,680 up from having to create all of these spreadsheets and people will say, creating spreadsheets 273 00:17:46,680 --> 00:17:49,440 is the only thing about my job that I enjoy. 274 00:17:49,440 --> 00:17:58,000 Yeah, I think also the downloads, the content marketing offers behind landing pages and 275 00:17:58,000 --> 00:18:03,720 forms of get your template, that kind of puts the sword to those as well. 276 00:18:03,720 --> 00:18:09,080 In terms of the example, I think I'd have full buy into that if there was an extra column 277 00:18:09,080 --> 00:18:16,960 that plugged in via API to the text bark tool so that I could explain the itinerary to the 278 00:18:16,960 --> 00:18:18,480 dogs I was walking. 279 00:18:18,480 --> 00:18:23,200 That tool is called Sarama by the way, S-A-R-A-M-A for those that want to go Google it. 280 00:18:23,200 --> 00:18:25,240 I promise you it looks real. 281 00:18:25,240 --> 00:18:28,800 But yes, it's cool to see that. 282 00:18:28,800 --> 00:18:34,000 And yeah, like we were saying right at the beginning of the podcast, Google going from 283 00:18:34,000 --> 00:18:39,720 being quite far behind and everybody waiting for copilot to announcing all of this stuff. 284 00:18:39,720 --> 00:18:46,920 And I've had BARD through my personal Gmail account for a while, but now I've been able 285 00:18:46,920 --> 00:18:52,360 to get BARD through my Bystrata Google Apps account. 286 00:18:52,360 --> 00:18:59,400 And there was a how to online from Google, how to get early access to a number of these 287 00:18:59,400 --> 00:19:00,400 features. 288 00:19:00,400 --> 00:19:07,480 I'm actually expecting to start to see BARD popping up to help me write emails and summarize 289 00:19:07,480 --> 00:19:10,600 text and rewrite text in documents fairly quickly. 290 00:19:10,600 --> 00:19:14,400 Whereas I haven't seen, I know you've got one bit of information that maybe goes against 291 00:19:14,400 --> 00:19:18,960 this, but I haven't seen a huge amount of movement on the Microsoft side as it relates 292 00:19:18,960 --> 00:19:21,840 to copilot making it into people's hands. 293 00:19:21,840 --> 00:19:26,160 Google Labs is being, they did announce this, that Google Labs is where you need to head 294 00:19:26,160 --> 00:19:29,560 to to sign up for early access to these things. 295 00:19:29,560 --> 00:19:33,320 They're going to have write it for me in Gmail. 296 00:19:33,320 --> 00:19:40,520 So where we've seen autocomplete on sentences in recent years, it's now going to have a 297 00:19:40,520 --> 00:19:46,320 full write it for me and with a text expander and a tone changer and all of that kind of 298 00:19:46,320 --> 00:19:48,680 stuff baked in. 299 00:19:48,680 --> 00:19:53,480 One thing about the generative AI everywhere, if we look at the fundamental model of what 300 00:19:53,480 --> 00:20:00,240 powers the Google enterprises is ads and search. 301 00:20:00,240 --> 00:20:03,440 And that was very revealing. 302 00:20:03,440 --> 00:20:09,080 They showed a, if in fact, if you go onto the Google website, you can see generative 303 00:20:09,080 --> 00:20:17,600 AI in search and there is a GIF that shows generative AI implemented into Google shopping. 304 00:20:17,600 --> 00:20:26,720 And on the example, there is a search query, someone saying that the search query is words 305 00:20:26,720 --> 00:20:35,600 to the effect of an ideal commuter bike for a five mile ride with hills. 306 00:20:35,600 --> 00:20:42,260 Now typically that result would normally, it would have brought up some Google shopping 307 00:20:42,260 --> 00:20:53,280 results and then some blog posts with best bikes, best commuter bikes and all of that 308 00:20:53,280 --> 00:20:54,280 kind of thing. 309 00:20:54,280 --> 00:20:58,320 There might have been a featured snippet at the top of that that had an excerpt from one 310 00:20:58,320 --> 00:21:00,120 of these blogs. 311 00:21:00,120 --> 00:21:05,360 If you watch this GIF show the generative AI, what happens is you do the search and 312 00:21:05,360 --> 00:21:07,200 then the shopping results come up. 313 00:21:07,200 --> 00:21:10,540 So you get your ads, they're still getting the ad money in there. 314 00:21:10,540 --> 00:21:18,200 And then beneath that, a full screen takeover, it pushes the organic completely off screen 315 00:21:18,200 --> 00:21:26,840 and comes up with a generative AI, like, you know, barred style response detailing different 316 00:21:26,840 --> 00:21:34,760 models of bike, bringing in images of the bikes with some ads in the side unit as well 317 00:21:34,760 --> 00:21:37,560 of the response. 318 00:21:37,560 --> 00:21:38,560 Organics gone. 319 00:21:38,560 --> 00:21:46,480 Like there is, because then beneath that, it's encouraging you to ask follow on questions. 320 00:21:46,480 --> 00:21:52,200 It's giving you suggested questions to follow on. 321 00:21:52,200 --> 00:21:58,000 The organic results are quite literally pushed down the screen like, bye, you were ranking 322 00:21:58,000 --> 00:22:02,600 number one for this and now you're not even visible. 323 00:22:02,600 --> 00:22:05,520 Yeah I think it's, that could end up being fascinating. 324 00:22:05,520 --> 00:22:10,680 I was found it interesting where the barred style response was also providing advice on 325 00:22:10,680 --> 00:22:15,520 how to make a decision for the type of bike based on different use cases and what's important 326 00:22:15,520 --> 00:22:18,040 to you and stuff like that. 327 00:22:18,040 --> 00:22:19,640 I saw mixed reaction to this. 328 00:22:19,640 --> 00:22:24,320 I saw some people online who were kind of excited because I think it still includes 329 00:22:24,320 --> 00:22:27,920 references for example, where relevant and that type of thing. 330 00:22:27,920 --> 00:22:32,740 But yeah, in terms of just getting pushed below the fold quite far into an experience 331 00:22:32,740 --> 00:22:39,000 that the searcher can continue on in the chat style interface is obviously a bit scary for 332 00:22:39,000 --> 00:22:40,840 SEOs and content marketers. 333 00:22:40,840 --> 00:22:47,680 I think the other factor there is we don't know how these models produce the outputs 334 00:22:47,680 --> 00:22:49,160 they produce, right? 335 00:22:49,160 --> 00:22:53,200 Nor does Google presumably or any of the other developers, hence the paper we talked about 336 00:22:53,200 --> 00:22:57,280 with GPT-4 trying to figure out how GPT-2 works. 337 00:22:57,280 --> 00:23:03,360 SEO has always been about reverse engineering what you think you needed to do to please 338 00:23:03,360 --> 00:23:06,600 the Google gods to get your content to rank first. 339 00:23:06,600 --> 00:23:09,840 How are you going to run those tests now? 340 00:23:09,840 --> 00:23:15,840 How are you going to figure out how to get your content even if you can get mentioned 341 00:23:15,840 --> 00:23:20,800 in those barred style conversational responses? 342 00:23:20,800 --> 00:23:25,600 Yeah, I mean, it's going to be fascinating to see how that plays out, but you would think 343 00:23:25,600 --> 00:23:29,720 that that's just not going to happen and you have to hope people scroll down to the organics. 344 00:23:29,720 --> 00:23:34,880 They gave one other insight into the scale of Google shopping actually. 345 00:23:34,880 --> 00:23:46,560 They said there are 1.8 billion updates to the Google shopping inventory every hour. 346 00:23:46,560 --> 00:23:49,360 Yikes. 347 00:23:49,360 --> 00:24:04,520 And if you think about this is plugged into the new LLM and the generative results, yeah, 348 00:24:04,520 --> 00:24:08,640 how are you going to game that from organic? 349 00:24:08,640 --> 00:24:15,360 It sounds like if they're showing ads above the chat response and within the chat response, 350 00:24:15,360 --> 00:24:22,540 I think they're thinking, well, we found a way to not destroy our ad based business model. 351 00:24:22,540 --> 00:24:26,840 We just found a way to squeeze ads into as many aspects of the conversational style barred 352 00:24:26,840 --> 00:24:29,120 response as we could. 353 00:24:29,120 --> 00:24:34,720 So if we want to get our brands and our links up into that conversational space, we got 354 00:24:34,720 --> 00:24:37,280 to pay for it. 355 00:24:37,280 --> 00:24:45,720 I think that was certainly the message that I was getting from that. 356 00:24:45,720 --> 00:24:50,000 SEOs everywhere wondering how is this going to change things. 357 00:24:50,000 --> 00:24:51,000 Yeah. 358 00:24:51,000 --> 00:24:54,560 So what's your key takeaway from this, do you think? 359 00:24:54,560 --> 00:25:01,600 My big takeaway was that this is all about UX. 360 00:25:01,600 --> 00:25:07,800 This is about making AI just accessible wherever you are. 361 00:25:07,800 --> 00:25:16,200 And whilst if you've been in this, you know, if you got excited about chat GPT from November 362 00:25:16,200 --> 00:25:18,200 onwards and went, oh, AI, this is exciting. 363 00:25:18,200 --> 00:25:23,280 There was still a huge swathe of people that haven't seen it or played with it or given 364 00:25:23,280 --> 00:25:25,880 it much thought. 365 00:25:25,880 --> 00:25:29,200 This is about making AI accessible to absolutely everyone. 366 00:25:29,200 --> 00:25:37,520 Not everybody wants to or has any kind of real need to be a prompt engineer. 367 00:25:37,520 --> 00:25:43,200 I was working the other day showing my wife how I was doing something with chat GPT and 368 00:25:43,200 --> 00:25:52,000 she just observed the way that I wrote a prompt and how detailed it was and how how thrifted 369 00:25:52,000 --> 00:25:53,000 it was. 370 00:25:53,000 --> 00:25:56,880 And she was like, you've got to understand that you clearly think differently. 371 00:25:56,880 --> 00:26:01,160 And when you've played with these things, you do, you can see a good prompter and that 372 00:26:01,160 --> 00:26:04,720 they understand how to put all of the context in there. 373 00:26:04,720 --> 00:26:07,000 People don't have time for that. 374 00:26:07,000 --> 00:26:14,320 The rest of the society, getting about your jobs and doing things, you don't want to be 375 00:26:14,320 --> 00:26:17,920 building these long, complicated, crafted prompts. 376 00:26:17,920 --> 00:26:22,640 You just want a thing in the Google Sheets now to do a thing for you. 377 00:26:22,640 --> 00:26:28,880 You want whatever tool you're using, you want it to be there to be easy to be seamless. 378 00:26:28,880 --> 00:26:37,660 And I think that is where Google have really pushed the envelope here. 379 00:26:37,660 --> 00:26:42,520 If it lives up to the promise, this is looking like a fairly seamless integration. 380 00:26:42,520 --> 00:26:50,560 Obviously there's a difference between the shiny marketing pitch to the lived reality. 381 00:26:50,560 --> 00:26:56,800 But this is about just making generative AI everywhere. 382 00:26:56,800 --> 00:27:02,840 And it's a UX play and I can see that this could be a very strong one. 383 00:27:02,840 --> 00:27:03,840 I completely agree. 384 00:27:03,840 --> 00:27:11,720 I was asked this week when we can expect to see generative AI and other AI, I guess mostly 385 00:27:11,720 --> 00:27:16,400 generative AI tools in common use in workplaces. 386 00:27:16,400 --> 00:27:24,520 And for me, and this was before the Google IO conference, I said it's when it's baked 387 00:27:24,520 --> 00:27:29,360 into tools that people use every day already and it's just there. 388 00:27:29,360 --> 00:27:34,280 I said that would drive adoption very quickly. 389 00:27:34,280 --> 00:27:39,400 And when CoPilot gets properly launched and we've got it in Word and PowerPoint and Excel 390 00:27:39,400 --> 00:27:44,120 and it will be interesting to see how fast we get this Palm2 driven tools. 391 00:27:44,120 --> 00:27:45,400 They're not calling it BARD, are they? 392 00:27:45,400 --> 00:27:47,120 I think they're calling it something else internally. 393 00:27:47,120 --> 00:27:50,080 I can't remember what they're calling it. 394 00:27:50,080 --> 00:27:55,240 When we see that in Sheets and Slides and Docs and Gmail, that's what will drive, I 395 00:27:55,240 --> 00:27:59,320 think will go very, very quickly from the old adoption curve. 396 00:27:59,320 --> 00:28:03,960 We'll go very quickly from the early adopters, which I think is people, many of which probably 397 00:28:03,960 --> 00:28:08,200 who listen to this podcast playing with Mid Journey and playing with Writer and Jasper 398 00:28:08,200 --> 00:28:13,040 and HyperWrite, these are the tools to get a feel for the landscape and see what they 399 00:28:13,040 --> 00:28:14,040 like. 400 00:28:14,040 --> 00:28:19,520 I think you go racing into the early majority once you put those tools in things that people 401 00:28:19,520 --> 00:28:26,120 use day in, day out already, which I think for most people is Microsoft and Google Apps. 402 00:28:26,120 --> 00:28:27,440 Yeah, absolutely. 403 00:28:27,440 --> 00:28:32,800 And it will be the point when you don't even notice that you're using AI. 404 00:28:32,800 --> 00:28:35,320 That'll be when you've got true takeoff. 405 00:28:35,320 --> 00:28:36,320 Agreed. 406 00:28:36,320 --> 00:28:41,040 So I think the next couple of months on that front are going to be fascinating because 407 00:28:41,040 --> 00:28:46,560 I think that point is probably going to come, I want to say faster than I thought. 408 00:28:46,560 --> 00:28:50,000 I was so excited about the copilot video and they made it sound like it was polished enough 409 00:28:50,000 --> 00:28:54,280 to launch pretty much straight away and then we've not really seen much, have we? 410 00:28:54,280 --> 00:28:57,800 There are interesting developments with Bing and the edge sidebar, which does give me hope 411 00:28:57,800 --> 00:28:59,600 that they are making a lot of progress. 412 00:28:59,600 --> 00:29:04,480 But yes, I think in the next couple of months, copilot in Microsoft and Google's equivalent 413 00:29:04,480 --> 00:29:09,080 in Slides and Docs and all that stuff and Gmail is going to be the accelerator for mass 414 00:29:09,080 --> 00:29:11,360 adoption, I think. 415 00:29:11,360 --> 00:29:12,360 Rightio. 416 00:29:12,360 --> 00:29:18,320 That's a long story, but it required it because there was so much there to go through. 417 00:29:18,320 --> 00:29:21,040 Let's get through the others fairly quickly if we can. 418 00:29:21,040 --> 00:29:25,160 We spent the lovely people's time who are listening to our podcast. 419 00:29:25,160 --> 00:29:26,160 Let's chat. 420 00:29:26,160 --> 00:29:27,880 Anthropic expanding Claude. 421 00:29:27,880 --> 00:29:34,280 So for those that are not aware, Anthropic is somewhat of a competitor, I guess, to OpenAI 422 00:29:34,280 --> 00:29:36,600 and other large language model developers. 423 00:29:36,600 --> 00:29:41,000 They have created a tool called Claude, which can be interacted with in different ways, 424 00:29:41,000 --> 00:29:44,320 but in essence has a chatbot just like chat GPT. 425 00:29:44,320 --> 00:29:49,720 And what's really interesting here is that they have upgraded the chatbot to allow it 426 00:29:49,720 --> 00:29:58,200 to process up to, I think it's 100,000 tokens, which translate to about 75,000 words of context, 427 00:29:58,200 --> 00:30:05,960 which is three times more than OpenAI's max, which is also only in beta, that maximum context. 428 00:30:05,960 --> 00:30:09,760 And only through the API. 429 00:30:09,760 --> 00:30:10,760 And only through the API. 430 00:30:10,760 --> 00:30:13,760 So you can't actually easily access it. 431 00:30:13,760 --> 00:30:19,040 This is a game changer for a lot of applications because it means that the AI can quickly digest 432 00:30:19,040 --> 00:30:23,760 and summarize, reference and analyze large volumes of information. 433 00:30:23,760 --> 00:30:29,400 I think I read somewhere online, was it the great Gatsby book that they pushed into the 434 00:30:29,400 --> 00:30:31,860 tool, changed one sentence and asked it to find it. 435 00:30:31,860 --> 00:30:34,320 And it was able to figure out which sentence had been changed. 436 00:30:34,320 --> 00:30:36,360 So that was one example. 437 00:30:36,360 --> 00:30:43,000 I saw someone else had asked it to analyze an 85 page Netflix corporate, Netflix corporate 438 00:30:43,000 --> 00:30:47,960 filing highlighting all the important stuff for investors and creating tables from balance 439 00:30:47,960 --> 00:30:52,160 sheets and providing analysis of marketing positioning and stuff like that, which I think 440 00:30:52,160 --> 00:30:54,440 is, is really interesting. 441 00:30:54,440 --> 00:31:00,600 This is another big upgrade in terms of competing with OpenAI and honestly offering something 442 00:31:00,600 --> 00:31:03,920 better than OpenAI. 443 00:31:03,920 --> 00:31:07,920 I was reading in one of the newsletters I subscribed to the, at this point is probably 444 00:31:07,920 --> 00:31:12,400 going to help perform several chat with your PDF and chat with your Word document tools 445 00:31:12,400 --> 00:31:17,200 that we'd previously mentioned on the podcast, just because of the sheer amount of information 446 00:31:17,200 --> 00:31:21,720 you can give it in one go. 447 00:31:21,720 --> 00:31:24,880 So yeah, I think you've been playing with this this week, haven't you Martin, since 448 00:31:24,880 --> 00:31:25,880 it was released? 449 00:31:25,880 --> 00:31:27,420 What was your thoughts on it? 450 00:31:27,420 --> 00:31:29,920 Very impressed. 451 00:31:29,920 --> 00:31:31,700 And it is a game changer. 452 00:31:31,700 --> 00:31:36,440 The ability to just throw huge volumes of text, ultimately it's about a novel. 453 00:31:36,440 --> 00:31:39,840 So a typical novel is about 70 to 100,000 words. 454 00:31:39,840 --> 00:31:44,160 So you can stick an entire novel in there and then just start having a chat with it. 455 00:31:44,160 --> 00:31:46,280 It's fantastic. 456 00:31:46,280 --> 00:31:51,080 A simple workflow that we have on this podcast, I think we might have even mentioned it when 457 00:31:51,080 --> 00:31:58,040 we were talking about the co-hair summarization tool that they announced a few weeks back, 458 00:31:58,040 --> 00:32:04,200 taking a transcript from a podcast, a podcast that sometimes runs over an hour, because 459 00:32:04,200 --> 00:32:07,800 some people just don't shut up. 460 00:32:07,800 --> 00:32:11,680 So taking a transcript for a full show like that, there's a lot of words. 461 00:32:11,680 --> 00:32:15,680 And if you try and put that into chat GPT, you're not getting close. 462 00:32:15,680 --> 00:32:20,000 Like it's just, you have to chop it up into so many small bits. 463 00:32:20,000 --> 00:32:27,520 Whereas what I did was gave Claude an example of the way that we format our show notes from 464 00:32:27,520 --> 00:32:28,520 previous episodes. 465 00:32:28,520 --> 00:32:29,880 I said, this is how we format them. 466 00:32:29,880 --> 00:32:35,160 I want to give you a transcript and then can you turn that into show notes? 467 00:32:35,160 --> 00:32:40,000 And I did that for the last episode that we recorded, chucked in the whole transcript 468 00:32:40,000 --> 00:32:42,440 and hit go. 469 00:32:42,440 --> 00:32:49,840 And it produced show notes that were surprisingly similar to the actual show notes from last 470 00:32:49,840 --> 00:32:51,360 week's episode. 471 00:32:51,360 --> 00:32:59,520 The headlines were actually identical, which I found really quite surprising to have got 472 00:32:59,520 --> 00:33:03,220 that much accuracy, but yeah, absolutely totally usable. 473 00:33:03,220 --> 00:33:05,800 And then I stuck in another transcript and it did it again. 474 00:33:05,800 --> 00:33:14,880 And yeah, that much context is, in terms of if you're looking for a differentiator against 475 00:33:14,880 --> 00:33:21,640 something like GPT-4, massive amounts of context will do it. 476 00:33:21,640 --> 00:33:22,640 Absolutely. 477 00:33:22,640 --> 00:33:29,040 I think that's probably where the next battle in the war of the generative AI companies 478 00:33:29,040 --> 00:33:34,280 may be fought is actually in contextual understanding and the ability to deal with large amounts 479 00:33:34,280 --> 00:33:37,520 of information. 480 00:33:37,520 --> 00:33:42,520 We developed a process here at Biostrat of some of the podcasts and audio and webinars 481 00:33:42,520 --> 00:33:47,960 and things that we do where we generate the transcript, which is often a good 10, 15, 482 00:33:47,960 --> 00:33:49,600 20,000 words for an hour. 483 00:33:49,600 --> 00:33:54,800 You'd be surprised how much nonsense we can talk on this podcast especially, but when 484 00:33:54,800 --> 00:33:58,120 people are speaking quickly, any transcript is usually a lot of content. 485 00:33:58,120 --> 00:34:03,000 And we got to the point where we had to split it and chunk it and you had to give Chachie 486 00:34:03,000 --> 00:34:07,880 PT maybe five, six, eight, 10 sections. 487 00:34:07,880 --> 00:34:08,880 You had to split that. 488 00:34:08,880 --> 00:34:10,700 And that's kind of manual. 489 00:34:10,700 --> 00:34:14,480 And there's always the danger that Chachie PT forgets what one of the first sections 490 00:34:14,480 --> 00:34:16,280 was about and all this other stuff. 491 00:34:16,280 --> 00:34:20,280 So the ability, not any of the speed and efficiency, but probably the quality of the output to 492 00:34:20,280 --> 00:34:24,720 just be able to dump that whole transcript in, it really will... 493 00:34:24,720 --> 00:34:26,440 I can't wait to get a hold of it. 494 00:34:26,440 --> 00:34:28,280 I've now put myself on the wait list. 495 00:34:28,280 --> 00:34:30,480 I know you've had it for a while, Martin. 496 00:34:30,480 --> 00:34:32,920 Maybe you can have a word for me. 497 00:34:32,920 --> 00:34:33,920 You know, ask Claus. 498 00:34:33,920 --> 00:34:37,600 Hi, friend Paul, please. 499 00:34:37,600 --> 00:34:44,560 But yeah, if you're developing content like webinars, podcasts, the ability to now turn 500 00:34:44,560 --> 00:34:48,720 that into text-based content like blog posts or eBooks has to be so much faster and easier 501 00:34:48,720 --> 00:34:50,000 now, I would have thought. 502 00:34:50,000 --> 00:34:51,000 Very much so. 503 00:34:51,000 --> 00:34:56,200 And just on the point about taking contextual content, we spoke the other week about embeddings 504 00:34:56,200 --> 00:35:03,300 and fine tuning and putting huge amounts of data and then training your own model on top 505 00:35:03,300 --> 00:35:08,520 of any of the existing, you know, GPT-4s or whatever. 506 00:35:08,520 --> 00:35:10,880 And these are complicated processes. 507 00:35:10,880 --> 00:35:16,480 And for many SMEs in particular, you don't have huge, huge, huge enterprise levels of 508 00:35:16,480 --> 00:35:23,160 knowledge base and documentation, but you do have maybe a novels of content. 509 00:35:23,160 --> 00:35:24,680 And now you can just stick that in a prompt. 510 00:35:24,680 --> 00:35:29,800 You don't need to create vector embeddings and databases and integrations. 511 00:35:29,800 --> 00:35:32,160 Just stick it all in a prompt and be like, there you go. 512 00:35:32,160 --> 00:35:38,080 That's our business handbook or our, all of our blog posts from the past 12 months, use 513 00:35:38,080 --> 00:35:41,280 that as context and help me produce more. 514 00:35:41,280 --> 00:35:42,280 Could you imagine that? 515 00:35:42,280 --> 00:35:47,400 Give it sort of a hundred or blog posts, might end up being slightly too many, but, and then 516 00:35:47,400 --> 00:35:52,360 say, which of these could be pulled together to make a summarized, insightful eBook on 517 00:35:52,360 --> 00:35:54,180 that particular topic, right? 518 00:35:54,180 --> 00:35:57,640 Give me infographic ideas based on the data that you find in these different blog posts. 519 00:35:57,640 --> 00:35:58,640 Yeah. 520 00:35:58,640 --> 00:35:59,640 I mean, you could do so much cool stuff. 521 00:35:59,640 --> 00:36:02,880 Right. In the interest of time, because I think we could go into so much detail on that 522 00:36:02,880 --> 00:36:04,720 one, we will crack on. 523 00:36:04,720 --> 00:36:08,840 Let's talk Stability AI and the text to animation tool it launched, right? 524 00:36:08,840 --> 00:36:13,800 Stability AI has launched Stable Animation SDK. 525 00:36:13,800 --> 00:36:18,160 So this is a tool designed for artists and developers to implement the most advanced 526 00:36:18,160 --> 00:36:22,080 stable diffusion models to create animations. 527 00:36:22,080 --> 00:36:27,340 And there's a number of ways that users can create animations so they can do through prompts 528 00:36:27,340 --> 00:36:33,760 without images, so text based prompts, or they can add a source image and use it as 529 00:36:33,760 --> 00:36:40,560 the kind of seed input, or they can use a source video. 530 00:36:40,560 --> 00:36:44,220 So you can also have text input plus an initial image input. 531 00:36:44,220 --> 00:36:48,680 So the user will provide the initial image that acts as a starting point for the animation 532 00:36:48,680 --> 00:36:55,680 and the text prompt is then used to basically guide the end result. 533 00:36:55,680 --> 00:37:01,280 And all of this has been made available for developers to integrate into their products, 534 00:37:01,280 --> 00:37:06,400 their platforms, what have you, via this new SDK. 535 00:37:06,400 --> 00:37:11,400 So that's just, again, making it really accessible, stable diffusion. 536 00:37:11,400 --> 00:37:18,000 So my understanding is that this is actually now, is this open source? 537 00:37:18,000 --> 00:37:19,000 I'm not sure. 538 00:37:19,000 --> 00:37:20,840 Is the underlying model open source? 539 00:37:20,840 --> 00:37:23,320 I don't know if you can access it. 540 00:37:23,320 --> 00:37:26,520 So I did dive a bit deep into this. 541 00:37:26,520 --> 00:37:28,840 Certainly you can't access it via a web app yet. 542 00:37:28,840 --> 00:37:34,800 So it's not going to be easy for your average marketer, if I take myself as an average marketer 543 00:37:34,800 --> 00:37:36,600 to get involved with playing with it. 544 00:37:36,600 --> 00:37:40,880 As I understand it at the moment, you have to download the necessary files and install 545 00:37:40,880 --> 00:37:45,760 the SDK using Python and you have to have a bit of coding experience to be able to use 546 00:37:45,760 --> 00:37:47,240 it at the moment, I think. 547 00:37:47,240 --> 00:37:57,320 Yes, I'm just looking at the model so it's all API driven pricing based on the size of 548 00:37:57,320 --> 00:37:58,440 the image. 549 00:37:58,440 --> 00:38:04,040 So dimension and image, that's 512 pixels by 512 pixels. 550 00:38:04,040 --> 00:38:15,680 The credit cost per operation will be 0.058 cents and the credit cost per resample operation, 551 00:38:15,680 --> 00:38:22,000 so 3D rendering mode, which sounds pretty cool, is 0.17 cents. 552 00:38:22,000 --> 00:38:28,240 My take home on the pricing just looking at it was we've had a chat off air about how 553 00:38:28,240 --> 00:38:32,640 these models work to generate video versus text and it's certainly more complicated than 554 00:38:32,640 --> 00:38:34,000 what I'm about to say. 555 00:38:34,000 --> 00:38:41,440 But in essence, if a static image is a frame and an animation is operating at 20, 24, 30 556 00:38:41,440 --> 00:38:46,560 frames a second, in essence these tools are just creating many more images and stringing 557 00:38:46,560 --> 00:38:52,200 them together to create moving images, i.e. video and animation. 558 00:38:52,200 --> 00:38:57,880 So it's probably going to eat up many more credits and cost much more to create these 559 00:38:57,880 --> 00:39:04,520 short animations than it would when you are creating static images. 560 00:39:04,520 --> 00:39:10,080 The cost of those generations for the different tools you might use like Dream Studio or whatever 561 00:39:10,080 --> 00:39:13,880 has been dropping all the time, so I don't think it's going to be prohibitively expensive 562 00:39:13,880 --> 00:39:20,040 for anyone if I'm honest, but it probably will cost 25 times as much, give or take, 563 00:39:20,040 --> 00:39:24,640 one would imagine if you're creating a couple of seconds worth of video and actually doing 564 00:39:24,640 --> 00:39:27,680 it at 24 frames a second or whatever it may be. 565 00:39:27,680 --> 00:39:33,960 The actual outputs themselves, they look pretty, you know, at the moment pretty raw, aren't 566 00:39:33,960 --> 00:39:34,960 they? 567 00:39:34,960 --> 00:39:35,960 Yeah. 568 00:39:35,960 --> 00:39:36,960 It's probably fair to say. 569 00:39:36,960 --> 00:39:37,960 Bit glitchy. 570 00:39:37,960 --> 00:39:41,680 Bit, yeah, just a bit, a little bit janky. 571 00:39:41,680 --> 00:39:46,320 The transitions aren't super smooth at the moment, much like we mentioned in last week's 572 00:39:46,320 --> 00:39:51,720 episode with the anime trailer and the vehicle that's driving along the road and then suddenly 573 00:39:51,720 --> 00:39:53,080 becomes the road. 574 00:39:53,080 --> 00:39:56,400 It's kind of got that sly vibe to it. 575 00:39:56,400 --> 00:39:57,720 But again, this is all day one. 576 00:39:57,720 --> 00:40:01,300 If you look at, you know, what we said about mid-journey earlier on, just fast forward 577 00:40:01,300 --> 00:40:07,200 12 months of where this is going to be, yeah, and the acceleration in this area is going 578 00:40:07,200 --> 00:40:08,200 to be huge. 579 00:40:08,200 --> 00:40:11,320 That's important to remember because we were like, oh, look at their hands. 580 00:40:11,320 --> 00:40:14,720 The hands look terrible and they've got 52 fingers and oh, look at the faces. 581 00:40:14,720 --> 00:40:15,720 They look like ghouls. 582 00:40:15,720 --> 00:40:21,080 And now of course we've got the Pope in his pimps. 583 00:40:21,080 --> 00:40:22,080 Balenciaga coat. 584 00:40:22,080 --> 00:40:26,000 Yeah, in his jacket looking like the boss. 585 00:40:26,000 --> 00:40:28,360 So yeah, I think you're right. 586 00:40:28,360 --> 00:40:30,680 I think we can expect to see that improve quite quickly. 587 00:40:30,680 --> 00:40:37,960 I think for marketers using it now, providing they can get it set up, you could probably 588 00:40:37,960 --> 00:40:45,120 create some interesting glitchy style things to play with and get attention with online. 589 00:40:45,120 --> 00:40:46,600 But yeah, we're not a production ready yet. 590 00:40:46,600 --> 00:40:50,160 But as Martin says, probably won't be that long until we see the develops we need to 591 00:40:50,160 --> 00:40:51,600 actually get that quality. 592 00:40:51,600 --> 00:40:55,880 So yes, worth digging that one out and having a look. 593 00:40:55,880 --> 00:41:01,200 Right, next news is Meta announces generative AI for advertisers. 594 00:41:01,200 --> 00:41:05,760 So Meta has unveiled a new product called AI Sandbox, which is aimed at streamlining 595 00:41:05,760 --> 00:41:09,360 the process of creating ad creatives for advertisers. 596 00:41:09,360 --> 00:41:11,280 The Sandbox provides three key functions. 597 00:41:11,280 --> 00:41:13,760 It can generate several versions of ad text. 598 00:41:13,760 --> 00:41:17,280 It can create a diverse range of creative assets for campaigns. 599 00:41:17,280 --> 00:41:22,640 And it includes a cropping tool for tailoring visuals for different sizes and formats, depending 600 00:41:22,640 --> 00:41:29,640 on where you're using it within the meta sphere, if you like, of across Facebook and all these 601 00:41:29,640 --> 00:41:32,560 other tools. 602 00:41:32,560 --> 00:41:36,920 We are not able to fully explore it yet and tell you what it can do because it's not launched 603 00:41:36,920 --> 00:41:37,920 yet. 604 00:41:37,920 --> 00:41:39,200 It's not how you can actually access it. 605 00:41:39,200 --> 00:41:43,080 But it certainly sounds like an interesting development for marketers. 606 00:41:43,080 --> 00:41:46,040 You could imagine a number of benefits that are going to come from this. 607 00:41:46,040 --> 00:41:50,720 I mean, off the top of our heads, you know, efficiency, because if you can use the AI 608 00:41:50,720 --> 00:41:56,440 Sandbox, you can significantly streamline the ad creation process and come up with multiple 609 00:41:56,440 --> 00:41:59,560 ad copy variations, diverse campaign assets and stuff. 610 00:41:59,560 --> 00:42:06,520 And it'll probably do that all quite quickly or within the Meta ads interface itself. 611 00:42:06,520 --> 00:42:07,520 Right. 612 00:42:07,520 --> 00:42:12,120 You can imagine improvements in creativity if you've got a bit of writers or creative 613 00:42:12,120 --> 00:42:17,160 block a little bit like how people are already using Mid Journey and Chat GPT to just give 614 00:42:17,160 --> 00:42:20,200 some ideas to get your own ideas going. 615 00:42:20,200 --> 00:42:25,600 So you might not use the outputs that the AI Sandbox produces, but it might just get 616 00:42:25,600 --> 00:42:29,400 your thought process going. 617 00:42:29,400 --> 00:42:34,240 I think you can definitely imagine some real time adaptation and performance improvements 618 00:42:34,240 --> 00:42:35,680 of your ads over time, right? 619 00:42:35,680 --> 00:42:40,160 Because if the AI driven system is gathering data on what text and creative is working 620 00:42:40,160 --> 00:42:44,120 best for your brand or for different product times, different audiences, different messages, 621 00:42:44,120 --> 00:42:51,360 all that stuff, it could auto adjust the creative and copy running ongoing experiments to improve 622 00:42:51,360 --> 00:42:53,320 results on an ongoing basis. 623 00:42:53,320 --> 00:42:54,320 Right. 624 00:42:54,320 --> 00:43:00,040 Which I think there are tools, fairly sophisticated, expensive tools that are offering to do that 625 00:43:00,040 --> 00:43:01,120 across different platforms. 626 00:43:01,120 --> 00:43:05,660 But this would be all within the Meta ads interface itself, which would bring it to 627 00:43:05,660 --> 00:43:06,660 many more people. 628 00:43:06,660 --> 00:43:07,660 Right. 629 00:43:07,660 --> 00:43:08,660 Yeah. 630 00:43:08,660 --> 00:43:09,660 I have thoughts on that. 631 00:43:09,660 --> 00:43:16,160 So, when you read the press release for this, the technology and the implementation, again, 632 00:43:16,160 --> 00:43:17,720 it's a kind of UI player. 633 00:43:17,720 --> 00:43:23,280 We've got some cool tech and we've made it really easy for advertisers to do background 634 00:43:23,280 --> 00:43:28,900 generation and testing of different creatives and what have you. 635 00:43:28,900 --> 00:43:34,000 Just reading some of the detail, they talk a bit about this Advantage Plus, which is 636 00:43:34,000 --> 00:43:41,600 basically AI ML powered optimization of your campaigns. 637 00:43:41,600 --> 00:43:48,720 I don't know if you ever speak to any Google AdWords specialists who talk about Performance 638 00:43:48,720 --> 00:43:49,720 Max. 639 00:43:49,720 --> 00:43:56,840 So, Performance Max is Google's answer to, well, it's not answers, Google's product that 640 00:43:56,840 --> 00:44:01,460 says basically let us deal with your campaigns. 641 00:44:01,460 --> 00:44:09,120 Give us budget and we will deploy ads across all of our ad units and they will be brilliant. 642 00:44:09,120 --> 00:44:12,880 Trust us, our machine learning algorithms know what they're doing. 643 00:44:12,880 --> 00:44:18,840 If you speak to any AdWords specialist who has been doing this in the industry for many 644 00:44:18,840 --> 00:44:22,940 years, they will tell you, do not turn on Performance Max. 645 00:44:22,940 --> 00:44:32,740 This is Google putting your ads in sub-optimized ad units, focus on tightly targeted keywords 646 00:44:32,740 --> 00:44:36,240 with high search intent, all of that kind of stuff. 647 00:44:36,240 --> 00:44:46,000 Be careful of turning on Performance Max because Google will spend your search intent money 648 00:44:46,000 --> 00:44:50,720 on keywords that you have no interest in bidding on. 649 00:44:50,720 --> 00:44:55,840 My concern with something like this Advantage Plus, just reading the press release makes 650 00:44:55,840 --> 00:45:05,320 me think, okay, this is Meta trying to sell ad units that if you go in and optimize your 651 00:45:05,320 --> 00:45:06,400 campaign, you know what you're doing. 652 00:45:06,400 --> 00:45:10,440 I'm a reasonably confident advertiser on Meta for instance. 653 00:45:10,440 --> 00:45:13,080 I'm not brilliant, but I kind of know my way around the system. 654 00:45:13,080 --> 00:45:18,520 I know what ad units I will and won't buy for instance. 655 00:45:18,520 --> 00:45:23,040 Advantage Plus will just start putting your ads anywhere that they want to sell an ad 656 00:45:23,040 --> 00:45:30,760 based on and they will use this AI creative editor to make the ads work in that ad space. 657 00:45:30,760 --> 00:45:33,640 It makes it sound like, oh yeah, we're really doing you a favor. 658 00:45:33,640 --> 00:45:37,400 Whereas really what they're saying is, oh, we'll make your creative fit to that ad unit 659 00:45:37,400 --> 00:45:42,480 that you had no intention of ever buying so they can sell more inventory. 660 00:45:42,480 --> 00:45:45,520 Maybe I'm being cynical, Paul, but... 661 00:45:45,520 --> 00:45:51,200 I think we're going to try and make this one of our shorter episodes and it's not, unfortunately. 662 00:45:51,200 --> 00:45:54,560 There's a lot of good discussion points to be had. 663 00:45:54,560 --> 00:46:00,880 I am not a Google Ads expert, but I know enough to be dangerous. 664 00:46:00,880 --> 00:46:03,000 As you said, I do have those conversations. 665 00:46:03,000 --> 00:46:08,120 I am extremely cool as many people in the world and those listening to the podcast will 666 00:46:08,120 --> 00:46:10,880 know and a lot of my friends are Google Ads specialists. 667 00:46:10,880 --> 00:46:17,280 I've had this discussion a number of times and I think my experience has been in the 668 00:46:17,280 --> 00:46:21,360 early days, avoid switching on the automation. 669 00:46:21,360 --> 00:46:24,480 Google just wants to spend more of your money. 670 00:46:24,480 --> 00:46:28,560 But my impression from a lot of the people I've spoken to is as the algorithms have got 671 00:46:28,560 --> 00:46:34,800 better, as long as you've manually run the program for enough time to give it the algorithm 672 00:46:34,800 --> 00:46:41,200 an opportunity to train on your particular ad set and your keywords that you're bidding 673 00:46:41,200 --> 00:46:46,200 on, so it can actually learn a bit about who the best people are to target. 674 00:46:46,200 --> 00:46:51,320 Then you adjust it for things like cost per conversion based bidding. 675 00:46:51,320 --> 00:46:59,320 You can actually do better than micromanaging the cost per click bidding and using single 676 00:46:59,320 --> 00:47:02,920 keyword ad groups and all this other stuff that Google Ads specialists used to do in 677 00:47:02,920 --> 00:47:03,920 the past. 678 00:47:03,920 --> 00:47:07,800 There's actually a book on this, it's very interesting, by a guy called Patrick Gilbert 679 00:47:07,800 --> 00:47:14,200 called Join or Die, Digital Advertising in the Age of Automation, where he actually makes, 680 00:47:14,200 --> 00:47:20,240 he's a Google Ads expert, an argument for why now is the time to actually turn over 681 00:47:20,240 --> 00:47:26,920 the reins to the algorithms to optimize your campaigns for you. 682 00:47:26,920 --> 00:47:28,120 Super controversial. 683 00:47:28,120 --> 00:47:34,720 I can certainly imagine how your Google Ads specialists wouldn't want that to be common 684 00:47:34,720 --> 00:47:39,320 knowledge or to be the common belief because I guess in theory it might mean that they 685 00:47:39,320 --> 00:47:42,760 feel or that their clients feel that they're no longer needed. 686 00:47:42,760 --> 00:47:44,120 I don't think that's true. 687 00:47:44,120 --> 00:47:50,960 I think it's exactly like when you were using ChatGPT, you need an expert monitoring things 688 00:47:50,960 --> 00:47:53,520 to make sure that nothing untoward happens. 689 00:47:53,520 --> 00:47:58,800 When you're using Google Ads, untoward means accidentally spending thousands or tens of 690 00:47:58,800 --> 00:48:02,760 thousands of pounds by accident on something that produces absolutely nothing. 691 00:48:02,760 --> 00:48:06,440 So expert in the loop is actually critical. 692 00:48:06,440 --> 00:48:10,980 But yeah, my impression had been that we're starting to reach a tipping point in terms 693 00:48:10,980 --> 00:48:12,860 of people's attitudes on that. 694 00:48:12,860 --> 00:48:17,000 If you're listening to this and you're a Google Ads specialist and you've got an opinion, 695 00:48:17,000 --> 00:48:19,480 hit us up on the Twitter's. 696 00:48:19,480 --> 00:48:23,920 Our Twitter had the line, never remember E is Martin. 697 00:48:23,920 --> 00:48:29,400 Aim podcal, no God. 698 00:48:29,400 --> 00:48:32,480 Now normally people who run podcasts would cut this, but we're not going. 699 00:48:32,480 --> 00:48:33,480 No, no. 700 00:48:33,480 --> 00:48:34,480 Cool. 701 00:48:34,480 --> 00:48:35,480 That's how we like it. 702 00:48:35,480 --> 00:48:43,080 I can tell you with certainty actually, never in doubt it's Aim or AI Marketing Pod. 703 00:48:43,080 --> 00:48:44,720 It's at AI Marketing Pod. 704 00:48:44,720 --> 00:48:45,720 There we go. 705 00:48:45,720 --> 00:48:46,720 There you go. 706 00:48:46,720 --> 00:48:47,720 AI Marketing Pod. 707 00:48:47,720 --> 00:48:50,120 We'd love to hear what you've got to say. 708 00:48:50,120 --> 00:48:53,680 And if you want to come on as an interviewer on the podcast, we'd love to get you on messages 709 00:48:53,680 --> 00:48:56,080 on the Twitter's, on the LinkedIn's. 710 00:48:56,080 --> 00:48:58,840 We'd love to hear your opinion. 711 00:48:58,840 --> 00:49:04,000 So yeah, I think that's very interesting, Martin, because for me, we've heard a lot 712 00:49:04,000 --> 00:49:14,660 this phrase, AI won't replace people, person X, but people who use AI will replace people. 713 00:49:14,660 --> 00:49:20,280 But this is one of those examples where it could almost become quite quantifiable because 714 00:49:20,280 --> 00:49:28,400 you could imagine that if this works as intended, our previous discussion completely accepted, 715 00:49:28,400 --> 00:49:33,120 the companies that are able to rapidly iterate automatically to improve their ad copy and 716 00:49:33,120 --> 00:49:38,400 ad created to drive conversions and sales are going to very quickly pour more money 717 00:49:38,400 --> 00:49:39,560 into that system. 718 00:49:39,560 --> 00:49:44,760 A, they're going to out compete the people who don't use the tools. 719 00:49:44,760 --> 00:49:48,560 And then when everybody's using the tools and everybody's pouring money into the system, 720 00:49:48,560 --> 00:49:53,420 all the cost per clicks and all the other costs are going to go up and putting more 721 00:49:53,420 --> 00:49:54,800 pressure on the system. 722 00:49:54,800 --> 00:49:58,460 Meta's very happy that everybody's pouring money in and happy to spend more. 723 00:49:58,460 --> 00:50:03,260 But at some point, one assumes they're going to saturate the audiences through this. 724 00:50:03,260 --> 00:50:08,960 And we're going to end up seeing so many ad variations and ads in our feeds that it could 725 00:50:08,960 --> 00:50:10,160 really pee us all off. 726 00:50:10,160 --> 00:50:14,240 So it'd be very interesting to see what the second order consequences of this would be. 727 00:50:14,240 --> 00:50:15,240 Yeah. 728 00:50:15,240 --> 00:50:16,240 Cool. 729 00:50:16,240 --> 00:50:17,240 Right. 730 00:50:17,240 --> 00:50:18,760 In the interest of speed, we'll rattle on. 731 00:50:18,760 --> 00:50:20,760 We've got a couple of extra things to chat through. 732 00:50:20,760 --> 00:50:24,560 While we're talking about ads, let's talk about chat bot ads are coming. 733 00:50:24,560 --> 00:50:29,560 So Microsoft this week has indicated that advertisements delivered by AI chat bots are 734 00:50:29,560 --> 00:50:34,360 on the horizon, potentially changing the landscape of search advertising. 735 00:50:34,360 --> 00:50:39,320 This is not dissimilar to what we touched on briefly in terms of what Martin told us 736 00:50:39,320 --> 00:50:44,280 in terms of Google search and ads showing up in the actual conversational area. 737 00:50:44,280 --> 00:50:49,100 I think this is what Microsoft is planning to do on its own platforms. 738 00:50:49,100 --> 00:50:55,720 But also interestingly, with partners who have AI chat bots on their own websites, a 739 00:50:55,720 --> 00:51:00,640 little bit like Google Display Network or, you know, Display Network advertising, but 740 00:51:00,640 --> 00:51:03,000 for chat bots. 741 00:51:03,000 --> 00:51:08,640 I would imagine that for most marketers, they're going to be wanting to use this to expand 742 00:51:08,640 --> 00:51:13,160 their reach once it sort of makes its way into different chat bots across the web. 743 00:51:13,160 --> 00:51:18,840 But it also got me thinking that as a publisher, I could imagine that if I had a big repository 744 00:51:18,840 --> 00:51:25,040 of content on my website, you know, what springs to mind for me is the natures and the sciences 745 00:51:25,040 --> 00:51:32,000 and the cells of the life science world, which have all this incredible peer review research 746 00:51:32,000 --> 00:51:36,760 articles that you could query and ask questions of via a chat bot. 747 00:51:36,760 --> 00:51:43,840 But as a publisher, monetizing that by introducing ads, relevant ads in the chat bot experience 748 00:51:43,840 --> 00:51:45,360 would be quite relevant. 749 00:51:45,360 --> 00:51:50,480 And then, at least in our market, I can imagine that being of interest to life science marketers, 750 00:51:50,480 --> 00:51:55,140 marketing instruments, reagents, tools that perhaps were used in those research papers 751 00:51:55,140 --> 00:52:03,280 to ensure that their products or their content is also shown in amongst that chat bot interface. 752 00:52:03,280 --> 00:52:07,080 So this is quite early days on this. 753 00:52:07,080 --> 00:52:11,320 If I've understood it rightly, there's no screen grabs or anything out there that shows 754 00:52:11,320 --> 00:52:13,040 you what this would look like. 755 00:52:13,040 --> 00:52:14,880 But it is a Microsoft press release. 756 00:52:14,880 --> 00:52:16,240 It is real. 757 00:52:16,240 --> 00:52:18,080 This is what they're planning to do. 758 00:52:18,080 --> 00:52:22,120 And as they say in a quote, we're ready to engage to discuss chat experiences, whether 759 00:52:22,120 --> 00:52:27,440 they need both algorithmic organic results and ads monetization or ads monetization only. 760 00:52:27,440 --> 00:52:32,360 We're listening to feedback from our partners as we continue to learn and evolve our offerings. 761 00:52:32,360 --> 00:52:35,480 So let's see how this plays out. 762 00:52:35,480 --> 00:52:39,960 Ads in chat bots and how different companies leverage that on their own chat bots and then 763 00:52:39,960 --> 00:52:43,560 how people bid and buy space in those chat bots. 764 00:52:43,560 --> 00:52:49,120 Yeah, got to find a good use case to have a chat bot on your website in the first place 765 00:52:49,120 --> 00:52:51,680 and then say time to monetize. 766 00:52:51,680 --> 00:52:57,720 Yeah, yeah, I think you've got to be a publisher with the reams of insightful information or 767 00:52:57,720 --> 00:53:01,600 information and content that that chat bot can draw upon, which is why I think it will 768 00:53:01,600 --> 00:53:03,440 end up being a publisher play. 769 00:53:03,440 --> 00:53:09,240 But obviously if there are publishers, niche publishers or broad spectrum publishers who 770 00:53:09,240 --> 00:53:13,280 reach your audience and they introduce these chat bots, there's an arbitrage opportunity 771 00:53:13,280 --> 00:53:18,360 there if you get in early and you can negotiate those deals with those publishers and be the 772 00:53:18,360 --> 00:53:20,280 first that's shown in their chat bot experience. 773 00:53:20,280 --> 00:53:24,040 Because I can imagine that people are going to want to play with it and see how useful 774 00:53:24,040 --> 00:53:29,440 it is compared to trawling through archives of articles or just leading into Google search 775 00:53:29,440 --> 00:53:30,440 or whatever. 776 00:53:30,440 --> 00:53:31,440 So it could be interesting. 777 00:53:31,440 --> 00:53:32,840 We'll see how that one plays out. 778 00:53:32,840 --> 00:53:36,560 Rightio, last bit is tool of the week. 779 00:53:36,560 --> 00:53:38,280 I mentioned this at the beginning. 780 00:53:38,280 --> 00:53:41,720 I was able to get access to Adobe Firefly. 781 00:53:41,720 --> 00:53:47,600 Now you imagine, you could imagine, but hopefully you'll remember, mine and I were super excited 782 00:53:47,600 --> 00:53:52,680 about this after watching the demo video, which was really, really impressive. 783 00:53:52,680 --> 00:53:56,800 And I have now got access within Adobe Firefly. 784 00:53:56,800 --> 00:54:03,300 There are a bunch of tools that you can get access to from being part of the beta. 785 00:54:03,300 --> 00:54:07,360 You can get their text to image, which is honestly not really very different from any 786 00:54:07,360 --> 00:54:11,300 of the other mid journeys and Duallys and other things that are out there. 787 00:54:11,300 --> 00:54:15,320 There is the recolor vector tool, which I haven't played with much, but I can talk to, 788 00:54:15,320 --> 00:54:19,120 is fairly novel and not offered by most of the other tools, certainly tools I've played 789 00:54:19,120 --> 00:54:20,120 with. 790 00:54:20,120 --> 00:54:25,240 And then there's their text effects where you give it a word and then it stylizes the 791 00:54:25,240 --> 00:54:26,240 text. 792 00:54:26,240 --> 00:54:30,400 And I think the things we were most excited in the demo video was things like having the 793 00:54:30,400 --> 00:54:36,560 biostrata logo made out of dripping chocolate or fire and this type of stuff. 794 00:54:36,560 --> 00:54:39,040 So text to image, pretty powerful. 795 00:54:39,040 --> 00:54:41,400 Make some really interesting images. 796 00:54:41,400 --> 00:54:47,120 I think where it lacks finesse potentially, or not finesse actually, because the images 797 00:54:47,120 --> 00:54:53,200 look good, it's more the scope of what you can do with it is it's trained on images owned 798 00:54:53,200 --> 00:54:54,200 by Adobe. 799 00:54:54,200 --> 00:54:58,240 So that's great if you're a large company and you're concerned there's going to be copyright 800 00:54:58,240 --> 00:55:03,340 issues down the line if you're using a stability AI or if you're using mid journey. 801 00:55:03,340 --> 00:55:04,800 But I think it also does limit it. 802 00:55:04,800 --> 00:55:10,100 And I saw an example online, which I also ran myself, of trying to create images of 803 00:55:10,100 --> 00:55:12,680 popular characters or famous people. 804 00:55:12,680 --> 00:55:17,680 You just can't do it because Adobe's image set doesn't know who they are. 805 00:55:17,680 --> 00:55:19,520 So I tried to create an image of Deadpool. 806 00:55:19,520 --> 00:55:24,440 I saw this example online of how mid journey will create an image of Deadpool sat on a 807 00:55:24,440 --> 00:55:32,200 car, but of course Deadpool being a Fox Disney character at this point, Marvel Disney character. 808 00:55:32,200 --> 00:55:36,080 Adobe creates something that looks a bit like Deadpool, but it's clearly not. 809 00:55:36,080 --> 00:55:40,640 So I think that it's the text to image cool, it has some real nice finesse to it produces 810 00:55:40,640 --> 00:55:45,960 really high quality images, but there's stuff that it can't do that other open source tools 811 00:55:45,960 --> 00:55:48,480 can for good or ill. 812 00:55:48,480 --> 00:55:55,000 On the text image front, this was the one I was the most excited about because it looks 813 00:55:55,000 --> 00:55:56,000 so cool. 814 00:55:56,000 --> 00:55:57,000 The text effects. 815 00:55:57,000 --> 00:55:58,600 Yeah, yeah, text effects. 816 00:55:58,600 --> 00:56:00,840 Sorry, I've not described that correctly. 817 00:56:00,840 --> 00:56:06,640 You give it a word and you say, make this word out of dripping paint or balloons or 818 00:56:06,640 --> 00:56:08,320 something like that. 819 00:56:08,320 --> 00:56:09,440 It is powerful. 820 00:56:09,440 --> 00:56:16,360 It creates some really interesting effects, but it runs into limitations extremely quickly. 821 00:56:16,360 --> 00:56:20,080 And what I mean by that is in the example that I was using it for, I gave it the word 822 00:56:20,080 --> 00:56:23,240 biostrata, which is the name of our marketing agency. 823 00:56:23,240 --> 00:56:29,560 And because we work in the life sciences, I said biostrata made out of cell culture 824 00:56:29,560 --> 00:56:34,600 cells and it kind of gave me something that looked kind of biological, but there was no 825 00:56:34,600 --> 00:56:35,600 obvious cells. 826 00:56:35,600 --> 00:56:41,080 It looked a little bit like it had tried to make it out of a graphic artist representation 827 00:56:41,080 --> 00:56:47,960 of the organelles inside cells like mitochondria and nuclei and Golgi apparatus and that type 828 00:56:47,960 --> 00:56:48,960 of stuff. 829 00:56:48,960 --> 00:56:51,600 But none of them really looked accurate. 830 00:56:51,600 --> 00:56:56,360 So it looked like a designer who didn't understand science had created it. 831 00:56:56,360 --> 00:56:57,360 Right. 832 00:56:57,360 --> 00:56:59,720 And I couldn't use that then for obvious reasons. 833 00:56:59,720 --> 00:57:02,320 So then I was like, oh, okay, well, let's try DNA. 834 00:57:02,320 --> 00:57:05,520 I pretty much got the same output as when I said cells. 835 00:57:05,520 --> 00:57:13,880 Oh, it's almost like the backend is like grouping all of this into biology and then doing what 836 00:57:13,880 --> 00:57:15,760 it thinks biology looks like. 837 00:57:15,760 --> 00:57:18,720 I tried chemistry and it looked a bit different. 838 00:57:18,720 --> 00:57:22,760 And in a few places, there were things that looked a little bit like conical flasks and 839 00:57:22,760 --> 00:57:23,760 stuff. 840 00:57:23,760 --> 00:57:28,080 There were a little bit of like biology type structures in there that look a bit like neurons. 841 00:57:28,080 --> 00:57:30,840 I tried neurons, didn't really look like neurons. 842 00:57:30,840 --> 00:57:32,640 It just looked like biology. 843 00:57:32,640 --> 00:57:35,020 So very cool. 844 00:57:35,020 --> 00:57:39,880 You can get some really cool effects, but your ability to really explore a creative 845 00:57:39,880 --> 00:57:46,040 space, at least in the tool as it is, in my experience is limited. 846 00:57:46,040 --> 00:57:51,320 I would imagine that if you keep it to kind of fairly, like the example that they gave 847 00:57:51,320 --> 00:57:55,120 was that they've done like food styling, didn't they? 848 00:57:55,120 --> 00:58:02,640 And like skyscraper, was it landscape, cityscape kind of images. 849 00:58:02,640 --> 00:58:08,680 But yeah, where you start getting into quite technical things there, it is just falling 850 00:58:08,680 --> 00:58:18,640 flat and clearly not many cell culture images in the older stock library of Adobe. 851 00:58:18,640 --> 00:58:24,720 Well, and so herein lies the rub, because there are. 852 00:58:24,720 --> 00:58:29,360 And it could have used, I think I tried at one point, use microscopy images, basically 853 00:58:29,360 --> 00:58:30,840 got the same thing again. 854 00:58:30,840 --> 00:58:36,360 It's like in an attempt to create this usable tool, it's almost like the large language 855 00:58:36,360 --> 00:58:43,240 model that underpins understanding the prompt, groups, related terms to generate an output 856 00:58:43,240 --> 00:58:47,520 that looks a bit biological, but isn't actually the thing you asked for. 857 00:58:47,520 --> 00:58:51,520 And that could be just because it's a demo, really, because it's a beta, and maybe that 858 00:58:51,520 --> 00:58:53,160 will improve. 859 00:58:53,160 --> 00:58:56,040 But I would say you can do some cool stuff with it. 860 00:58:56,040 --> 00:59:02,080 You can probably get some inspiration from it, but you can't just imagine to your heart's 861 00:59:02,080 --> 00:59:04,640 desire and get what you want. 862 00:59:04,640 --> 00:59:08,640 And if we look at the recolor vectors tool, there's similar limitations there. 863 00:59:08,640 --> 00:59:14,640 So for example, when you try and recolor the vectors, you can't give it a hex code and 864 00:59:14,640 --> 00:59:18,760 tell it what color you want, as far as I can tell. 865 00:59:18,760 --> 00:59:25,240 There are 18 different color palettes, colors across the palette, white, black, a neon green, 866 00:59:25,240 --> 00:59:27,520 blue, but you can't give it your color. 867 00:59:27,520 --> 00:59:33,160 So if I wanted to recolor a particular vector in Biosstratus brand colors, that option is 868 00:59:33,160 --> 00:59:34,600 not available right now. 869 00:59:34,600 --> 00:59:38,080 I have to recolor it in the colors that are made available here. 870 00:59:38,080 --> 00:59:46,320 So again, implications of power with limitations that fundamentally probably wouldn't make 871 00:59:46,320 --> 00:59:52,920 it usable for a graphic designer delivering on a client brief, for example. 872 00:59:52,920 --> 00:59:57,080 So cool power, wonderful demo. 873 00:59:57,080 --> 00:59:59,920 Probably if you're a professional designer, it's not going to be able to do what you need 874 00:59:59,920 --> 01:00:02,920 it to do just yet, is what I would say. 875 01:00:02,920 --> 01:00:08,880 In terms of actually using the tool, if people wanted to sign up for it, how do you interface 876 01:00:08,880 --> 01:00:09,880 with it? 877 01:00:09,880 --> 01:00:11,000 Is it through a web application? 878 01:00:11,000 --> 01:00:13,400 Is it a plugin in Photoshop? 879 01:00:13,400 --> 01:00:14,400 Pass on the Photoshop. 880 01:00:14,400 --> 01:00:19,160 I haven't actually opened up Photoshop to see if I can now access it through the fact that 881 01:00:19,160 --> 01:00:20,160 I'm on the beta. 882 01:00:20,160 --> 01:00:24,120 I'm using the web app, which is firefly.adaby.com. 883 01:00:24,120 --> 01:00:27,680 You'll have to sign up for the beta waiting list. 884 01:00:27,680 --> 01:00:36,000 And then you access it all through a web app similar to the dream studio and clip drop. 885 01:00:36,000 --> 01:00:38,040 And those types of tools. 886 01:00:38,040 --> 01:00:39,280 Cool. 887 01:00:39,280 --> 01:00:41,360 So that was your tool of the week. 888 01:00:41,360 --> 01:00:44,240 Thank you so much for joining us here. 889 01:00:44,240 --> 01:00:47,700 Please subscribe if you enjoyed this, share it with your marketing friends if you found 890 01:00:47,700 --> 01:00:48,700 it useful. 891 01:00:48,700 --> 01:00:50,040 Maybe they will too. 892 01:00:50,040 --> 01:00:51,640 Martin, love hanging out with you. 893 01:00:51,640 --> 01:00:52,640 Thanks for your time today. 894 01:00:52,640 --> 01:00:53,640 I'll catch you next week. 895 01:00:53,640 --> 01:00:54,640 Yeah. 896 01:00:54,640 --> 01:00:55,640 See you next week. 897 01:00:55,640 --> 01:01:00,760 If you'd like to go over to artificiallyintelligentmarketing.com, sign up to the newsletter. 898 01:01:00,760 --> 01:01:06,240 You will get this email to you in your inbox as soon as the episode is released. 899 01:01:06,240 --> 01:01:07,240 Never miss an episode. 900 01:01:07,240 --> 01:01:08,240 Oh, what a top tip. 901 01:01:08,240 --> 01:01:09,240 Nice one, Martin. 902 01:01:09,240 --> 01:01:10,240 Have a good weekend. 903 01:01:10,240 --> 01:01:11,240 Speak to you next week. 904 01:01:11,240 --> 01:01:12,240 Cheers. 905 01:01:12,240 --> 01:01:13,240 See you later. 906 01:01:13,240 --> 01:01:14,240 Bye. 907 01:01:14,240 --> 01:01:18,360 Thank you for listening to Artificially Intelligent Marketing. 908 01:01:18,360 --> 01:01:24,400 To stay on top of the latest trends, tips and tools in the world of marketing AI, be 909 01:01:24,400 --> 01:01:26,160 sure to subscribe. 910 01:01:26,160 --> 01:01:54,480 We look forward to seeing you again next week.