1 00:00:00,000 --> 00:00:09,960 Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the 2 00:00:09,960 --> 00:00:15,700 latest trends, tips, and tools in the world of marketing AI, helping you get the best 3 00:00:15,700 --> 00:00:18,540 results from your marketing efforts. 4 00:00:18,540 --> 00:00:22,160 Now let's join our hosts, Paul Avery and Martin Broadhurst. 5 00:00:22,160 --> 00:00:23,760 Hello, everybody. 6 00:00:23,760 --> 00:00:28,680 Welcome to episode 43 of Artificially Intelligent Marketing. 7 00:00:28,680 --> 00:00:30,520 It's also kind of our birthday today. 8 00:00:30,520 --> 00:00:33,040 In fact, our birthday was a week or two ago. 9 00:00:33,040 --> 00:00:37,320 If you've been a listener since the start, you've had to put up with Martin and I talking 10 00:00:37,320 --> 00:00:41,000 nonsense about AI and marketing in your ear for the last year. 11 00:00:41,000 --> 00:00:43,120 We really appreciate your time. 12 00:00:43,120 --> 00:00:47,080 I'm joined by said Martin Broadhurst, as always. 13 00:00:47,080 --> 00:00:48,080 How are you, Martin? 14 00:00:48,080 --> 00:00:49,080 Now the podcast is a year old. 15 00:00:49,080 --> 00:00:52,200 A lot has happened in that time as well, hasn't it? 16 00:00:52,200 --> 00:00:55,400 It's developed at quite a pace, this AI space that we cover. 17 00:00:55,400 --> 00:00:59,560 I don't think I would have thought some of the innovations that we've seen would have 18 00:00:59,560 --> 00:01:02,840 emerged quite the speed that they have done. 19 00:01:02,840 --> 00:01:04,040 Yeah, it is. 20 00:01:04,040 --> 00:01:08,080 We've been riding that exponential curve as we've been trying to cover the news items 21 00:01:08,080 --> 00:01:09,080 as they've come out. 22 00:01:09,080 --> 00:01:14,560 There's elements now to how fast it all jumps forward that we're almost taking for granted 23 00:01:14,560 --> 00:01:15,560 a bit. 24 00:01:15,560 --> 00:01:17,640 It's like Sora and its amazing video. 25 00:01:17,640 --> 00:01:20,480 This is the one that sticks in my mind most recently. 26 00:01:20,480 --> 00:01:26,840 How quickly as humans we came to fixate on the very minor errors that it makes. 27 00:01:26,840 --> 00:01:31,200 Maybe they're not so minor, but compared to what it can actually do, why aren't we like, 28 00:01:31,200 --> 00:01:32,200 what? 29 00:01:32,200 --> 00:01:33,200 How can it even do that? 30 00:01:33,200 --> 00:01:34,200 But there you go. 31 00:01:34,200 --> 00:01:37,200 We were living this this week, weren't we, Martin? 32 00:01:37,200 --> 00:01:38,720 Because we were on tour. 33 00:01:38,720 --> 00:01:45,820 We were delivering an AI workshop for a global marketing team, getting them up to speed on 34 00:01:45,820 --> 00:01:52,040 what the latest technology can do, what it can't do, and helping them get stuck into 35 00:01:52,040 --> 00:01:55,880 the tools with a real practical hands-on workshop. 36 00:01:55,880 --> 00:01:57,400 I must say that was brilliant. 37 00:01:57,400 --> 00:02:05,480 Some of the creative uses of AI for a campaign creation within, what was it, 50 minutes or 38 00:02:05,480 --> 00:02:06,480 so? 39 00:02:06,480 --> 00:02:07,480 It was amazing. 40 00:02:07,480 --> 00:02:11,840 Yeah, I always come away from the workshops we run, Martin, hoping that the people we're 41 00:02:11,840 --> 00:02:17,080 working with are inspired to more practically grab hold of the AI tools and use them in 42 00:02:17,080 --> 00:02:18,080 their everyday. 43 00:02:18,080 --> 00:02:22,840 But those creative ideas they're coming up with, they inspire me when I get back to the 44 00:02:22,840 --> 00:02:25,600 office for the ideas that I'm hoping to implement. 45 00:02:25,600 --> 00:02:33,320 As you said, really practically exploring how you can go from usually a fictitious product 46 00:02:33,320 --> 00:02:39,600 launch idea to a campaign theme, imagery, a launch plan, and all the other materials, 47 00:02:39,600 --> 00:02:42,400 of course, that we saw produced by the team this week was amazing. 48 00:02:42,400 --> 00:02:46,280 We had songs produced by Copilot with Suno. 49 00:02:46,280 --> 00:02:49,440 It is amazing what you can do in a short time. 50 00:02:49,440 --> 00:02:51,360 Yeah, it was great fun. 51 00:02:51,360 --> 00:02:56,240 If you're listening to this podcast and thinking, oh, I really love for Martin and Paul to come 52 00:02:56,240 --> 00:03:01,800 in and help my team figure out how we can actually make AI a useful part of our tech 53 00:03:01,800 --> 00:03:04,800 stack, then get in touch with us and we will gladly help. 54 00:03:04,800 --> 00:03:09,640 Because I think one of the things that we really saw this week was how just spending 55 00:03:09,640 --> 00:03:14,000 a little bit of time as a team to actually play with the tools can really help you make 56 00:03:14,000 --> 00:03:16,760 quite a big leap forward. 57 00:03:16,760 --> 00:03:20,520 Speaking of leap forwards, or leaps forward, Martin, should we leap forward into the first 58 00:03:20,520 --> 00:03:22,560 story of the week? 59 00:03:22,560 --> 00:03:23,560 Beautiful segue, Paul. 60 00:03:23,560 --> 00:03:24,560 Expertly done. 61 00:03:24,560 --> 00:03:30,080 You wouldn't think I was jet lagged after we spent 30 hours in the States this week, 62 00:03:30,080 --> 00:03:31,080 would you? 63 00:03:31,080 --> 00:03:32,080 But no, I'm bang on it. 64 00:03:32,080 --> 00:03:33,960 Let's get into that first story. 65 00:03:33,960 --> 00:03:40,600 First up is 11 Labs, one of the tools that people were using in their creative workshop 66 00:03:40,600 --> 00:03:41,600 outputs. 67 00:03:41,600 --> 00:03:44,120 Neatly done as well, I must say. 68 00:03:44,120 --> 00:03:53,680 So 11 Labs have moved from AI voice synthesis into the realms of AI created sound effects. 69 00:03:53,680 --> 00:04:01,440 So it's a whole new way for creators to use AI generated audio within their podcasts, 70 00:04:01,440 --> 00:04:06,080 within their video production, within their music production scripts, you name it. 71 00:04:06,080 --> 00:04:09,660 You can now do it within 11 Labs. 72 00:04:09,660 --> 00:04:15,940 How it works is the model was trained on a vast data set of real world sounds, which 73 00:04:15,940 --> 00:04:22,160 were meticulously labeled and described by some data labelers. 74 00:04:22,160 --> 00:04:29,240 And all of this training has gone in really to help the model replicate these noises. 75 00:04:29,240 --> 00:04:37,600 And it's available in an early stage access account, which I've got access to when you 76 00:04:37,600 --> 00:04:42,280 had a play with it alongside me while we were traveling. 77 00:04:42,280 --> 00:04:51,600 Ostensibly it should be able to create everything from atmospheric sounds to weather phenomena, 78 00:04:51,600 --> 00:04:57,360 noises like hammers hitting nails and things like that. 79 00:04:57,360 --> 00:05:06,000 But I think early uses, certainly my experience with it straight off the bat alongside the 80 00:05:06,000 --> 00:05:11,040 sound effects that you asked me to put into it, where that it is limited in some areas 81 00:05:11,040 --> 00:05:17,840 and I've read some reviews and online discussion saying that, yeah, it's okay. 82 00:05:17,840 --> 00:05:19,440 It's a good first stab. 83 00:05:19,440 --> 00:05:24,560 This is the first generation of this model, so it's only going to get better from here. 84 00:05:24,560 --> 00:05:26,860 But there are definitely some things that it struggles with. 85 00:05:26,860 --> 00:05:32,880 In my initial experiments, I found that it is quite good with things like thunderstorm 86 00:05:32,880 --> 00:05:36,800 noise effects or birdsong in the countryside. 87 00:05:36,800 --> 00:05:41,820 But where it really struggles is electronic sounds. 88 00:05:41,820 --> 00:05:43,480 It didn't seem to get any of those right. 89 00:05:43,480 --> 00:05:47,840 For instance, I tried to get it to recreate certain sound effects that one might hear 90 00:05:47,840 --> 00:05:53,160 in a video game, but it made kind of weird screeching noises. 91 00:05:53,160 --> 00:06:00,040 And if you try and do something a bit abstract, then it really goes awry, doesn't it Paul? 92 00:06:00,040 --> 00:06:04,280 You know me, I love to try and break these tools instant that we get hold of them. 93 00:06:04,280 --> 00:06:11,880 And certainly a dog playing a harmonica didn't quite work out as I was hoping it to. 94 00:06:11,880 --> 00:06:15,560 The car engine sound that you came up with, a car engine beating like a heartbeat was 95 00:06:15,560 --> 00:06:19,720 kind of okay, but really not very usable. 96 00:06:19,720 --> 00:06:24,760 Yeah, if you want to get a bit crazy and do some interesting things, it fell apart quite 97 00:06:24,760 --> 00:06:25,760 quickly didn't it? 98 00:06:25,760 --> 00:06:32,600 But it sounds like if you're asking for real use case sounds that you would genuinely want 99 00:06:32,600 --> 00:06:37,760 to use like a cowbell ringing or something, then it performs a little bit better, but 100 00:06:37,760 --> 00:06:41,720 still with space to grow into a real production capable tool. 101 00:06:41,720 --> 00:06:43,280 Is that a fair description Martin? 102 00:06:43,280 --> 00:06:44,480 Yeah, very much so. 103 00:06:44,480 --> 00:06:53,560 I asked it for Blackbirds singing in the English countryside and it gives you five outputs when 104 00:06:53,560 --> 00:06:54,560 it creates it. 105 00:06:54,560 --> 00:06:56,920 So it doesn't just give you one version. 106 00:06:56,920 --> 00:07:00,560 And some of them sounded a little bit electronic. 107 00:07:00,560 --> 00:07:02,080 They sounded very synthetic. 108 00:07:02,080 --> 00:07:04,700 They had an element to that. 109 00:07:04,700 --> 00:07:10,720 On another example, I asked it to create the sound of a Ferrari sports car engine starting 110 00:07:10,720 --> 00:07:12,240 up. 111 00:07:12,240 --> 00:07:14,280 And that was interesting. 112 00:07:14,280 --> 00:07:16,760 They did sound good when they started up. 113 00:07:16,760 --> 00:07:18,040 They're all quite different. 114 00:07:18,040 --> 00:07:26,720 And I think if you were a petrol head, you'd be going, it's not a Ferrari Enzo or whatever. 115 00:07:26,720 --> 00:07:34,520 I don't know the models, but if you were, you would pick it apart quite comfortably, 116 00:07:34,520 --> 00:07:35,520 I think. 117 00:07:35,520 --> 00:07:37,480 But they did sound pretty good. 118 00:07:37,480 --> 00:07:43,400 I asked it to do thunderstorms recorded from inside a tin roof shelter. 119 00:07:43,400 --> 00:07:49,120 That was odd because the first two that it created were just like a screech. 120 00:07:49,120 --> 00:07:50,120 It just had failed. 121 00:07:50,120 --> 00:07:54,320 The generation had just not completed the task at all. 122 00:07:54,320 --> 00:07:59,320 But then the subsequent sounds that it created were pretty good. 123 00:07:59,320 --> 00:08:04,840 So I think if you're listening to this and you're a podcast creator, agency side or 124 00:08:04,840 --> 00:08:10,840 client side, or you're looking to create sound effects for videos, it's worth probably 125 00:08:10,840 --> 00:08:14,600 having a play with this, but just manage your expectations in terms of what you're going 126 00:08:14,600 --> 00:08:18,800 to get out of it and keep an eye on the technology as it emerges. 127 00:08:18,800 --> 00:08:22,600 This is one to probably keep an eye on rather than to wrap straight into your production 128 00:08:22,600 --> 00:08:23,600 stack. 129 00:08:23,600 --> 00:08:24,600 Yeah. 130 00:08:24,600 --> 00:08:27,640 And I think 11 Labs would probably echo that caution, right? 131 00:08:27,640 --> 00:08:33,320 Play around with it, test it, give them feedback, but don't go using this as your default sound 132 00:08:33,320 --> 00:08:35,640 effect library on day one. 133 00:08:35,640 --> 00:08:36,640 Makes sense. 134 00:08:36,640 --> 00:08:37,840 What's good for us to know these things are emerging. 135 00:08:37,840 --> 00:08:42,040 And I think again, when you're in marketing, understanding the tools that are emerging 136 00:08:42,040 --> 00:08:47,040 and evolving, because this is, as we say, a thousand times per day per workshop, this 137 00:08:47,040 --> 00:08:49,600 is as bad as these tools will ever be. 138 00:08:49,600 --> 00:08:53,320 And now this gives you an indication of something you probably could wrap into your production 139 00:08:53,320 --> 00:08:58,520 stack in three or maybe six months time as these tools improve. 140 00:08:58,520 --> 00:08:59,520 Good stuff. 141 00:08:59,520 --> 00:09:04,280 Speaking of tools improving, Mid Journey had a really interesting announcement this week. 142 00:09:04,280 --> 00:09:11,520 So Mid Journey is the AI driven image generator, probably best in class for the quality of 143 00:09:11,520 --> 00:09:15,040 images produced, especially when it comes to photo realism. 144 00:09:15,040 --> 00:09:18,800 And this week they launched a new feature that allows users to maintain consistency 145 00:09:18,800 --> 00:09:21,320 in character appearance across multiple images. 146 00:09:21,320 --> 00:09:26,000 So for most people playing with these tools, this has been a really big ask. 147 00:09:26,000 --> 00:09:31,920 If you want to create storyboards or you want to create like a graphic novel, the inability 148 00:09:31,920 --> 00:09:39,920 to control character consistency, there are ways around that using things like seed values, 149 00:09:39,920 --> 00:09:42,640 but it doesn't work that well on average. 150 00:09:42,640 --> 00:09:47,040 So this is a big leap forward, I think, in terms of being able to give us as creators 151 00:09:47,040 --> 00:09:51,780 across a wide range of fields, marketing being one of the control over character consistency 152 00:09:51,780 --> 00:09:53,620 between images. 153 00:09:53,620 --> 00:09:58,160 The way it works is you're going to use a specific type of tag in Mid Journey. 154 00:09:58,160 --> 00:10:04,320 So people who are familiar with Mid Journey will be used to using dash dash and then some 155 00:10:04,320 --> 00:10:06,560 information to control different aspects of the image. 156 00:10:06,560 --> 00:10:12,800 In this case, it's dash dash CREF and then followed by a URL to an image of that character. 157 00:10:12,800 --> 00:10:17,880 So in other words, you basically say pay attention to this image, which contains this character 158 00:10:17,880 --> 00:10:20,200 to drive the consistency. 159 00:10:20,200 --> 00:10:26,080 Then you can use a separate tag called the CW tag, so dash dash CW, which is a numerical 160 00:10:26,080 --> 00:10:33,520 value from 0 to 100, which basically is like a weighting variable in terms of how closely 161 00:10:33,520 --> 00:10:38,760 should your new image resemble the character that you're trying to create. 162 00:10:38,760 --> 00:10:44,080 So it works best when those characters have been originally created in Mid Journey. 163 00:10:44,080 --> 00:10:48,520 So you go in Mid Journey, create a character and then that's the image that you use as 164 00:10:48,520 --> 00:10:51,000 the source image for maintaining the character. 165 00:10:51,000 --> 00:10:54,720 It can work with images from outside Mid Journey. 166 00:10:54,720 --> 00:10:58,560 It just doesn't tend to work as well. 167 00:10:58,560 --> 00:11:02,960 There's obviously going to be some ethical and copyright concerns here, Martin, in terms 168 00:11:02,960 --> 00:11:08,080 of using characters based images to feed into these tools. 169 00:11:08,080 --> 00:11:09,080 What do you think? 170 00:11:09,080 --> 00:11:11,400 Do you think this is going to be a big issue? 171 00:11:11,400 --> 00:11:19,920 Well, we've said before that Mid Journey is one of the tools where you can create images 172 00:11:19,920 --> 00:11:25,560 from copyrighted characters, which OpenAI might say, absolutely not, we're not doing 173 00:11:25,560 --> 00:11:27,000 it. 174 00:11:27,000 --> 00:11:32,860 Now that we can do that with consistency as well, I can see it being open to abuse. 175 00:11:32,860 --> 00:11:38,440 But again, with any of these things, I like to think that we're speaking to an audience 176 00:11:38,440 --> 00:11:45,320 of reputable marketers that will be paying attention to these kinds of details and therefore 177 00:11:45,320 --> 00:11:48,680 you probably shouldn't come unstuck. 178 00:11:48,680 --> 00:11:54,880 Don't go taking photos of Joe Biden and sticking them in Mid Journey and getting them to create 179 00:11:54,880 --> 00:11:59,200 photos endorsing your company. 180 00:11:59,200 --> 00:12:00,200 Try not to do that. 181 00:12:00,200 --> 00:12:03,520 Otherwise, I think for the most part, you'd probably be fine. 182 00:12:03,520 --> 00:12:04,520 I think you're right. 183 00:12:04,520 --> 00:12:08,800 I think for marketers, hopefully we know the frameworks we need to operate within, but 184 00:12:08,800 --> 00:12:12,000 I suspect there will be bad actors out there that misuse this. 185 00:12:12,000 --> 00:12:16,760 So I think as consumers, we have to be ready for what the outcomes of that might be. 186 00:12:16,760 --> 00:12:21,620 Maybe it's much, much better at using consistent characters that Mid Journey itself has created, 187 00:12:21,620 --> 00:12:24,360 but Mid Journey will also give you a half decent image of Joe Biden. 188 00:12:24,360 --> 00:12:28,120 So I don't know where that leaves us. 189 00:12:28,120 --> 00:12:32,320 If you want to have a play with it, it is available via Mid Journey's version six, which 190 00:12:32,320 --> 00:12:33,500 is in beta. 191 00:12:33,500 --> 00:12:38,060 So if you are on the Discord, you can actually use this today. 192 00:12:38,060 --> 00:12:42,360 At some point in the near future, Mid Journey will be moving out of Discord and into a web-based 193 00:12:42,360 --> 00:12:46,920 interface because for most of us marketers, we don't have Discord accounts, we're not 194 00:12:46,920 --> 00:12:48,220 in Discord day in, day out. 195 00:12:48,220 --> 00:12:50,920 It's a bit of a clunky user experience. 196 00:12:50,920 --> 00:12:55,040 And as I understand it, at this point, it's still power users that are allowed to use 197 00:12:55,040 --> 00:12:56,120 the website version. 198 00:12:56,120 --> 00:13:02,120 People have generated 10, 20,000 images in Mid Journey, but hopefully that testing period 199 00:13:02,120 --> 00:13:08,840 will be over soon and more marketers can get their hands on the tool and then have a play 200 00:13:08,840 --> 00:13:12,960 with maintaining consistent characters. 201 00:13:12,960 --> 00:13:16,760 Onto our next story and it's a big regulation story. 202 00:13:16,760 --> 00:13:20,360 Martin, tell us about the EU's AI Act this week. 203 00:13:20,360 --> 00:13:24,200 We have, we've followed this one from the start, haven't we? 204 00:13:24,200 --> 00:13:30,560 We've been giving updates as it's progressed through each of the legislative stages. 205 00:13:30,560 --> 00:13:36,660 And this week on March 13th, so a couple of days ago, European Parliament has passed the 206 00:13:36,660 --> 00:13:40,760 world's first comprehensive AI regulation. 207 00:13:40,760 --> 00:13:46,400 So the legislation aims to address the risks associated with AI technologies and promotes 208 00:13:46,400 --> 00:13:51,120 what they're calling a more human-centric approach to AI development and use. 209 00:13:51,120 --> 00:13:57,120 The core of this is around that risk level that it presents to the users. 210 00:13:57,120 --> 00:14:02,920 So it introduces a framework that categorizes AI products based on their risk level, applying 211 00:14:02,920 --> 00:14:09,640 varying degrees of scrutiny to ensure that the deployment of the technology is not going 212 00:14:09,640 --> 00:14:15,680 to be used for something that could inadvertently cause harm or be used maliciously. 213 00:14:15,680 --> 00:14:19,280 So you can think of this as being like an insurance policy. 214 00:14:19,280 --> 00:14:25,180 If we think about a specific use case, if you were using artificial intelligence to 215 00:14:25,180 --> 00:14:31,640 say whether somebody was eligible or not for a type of insurance, health insurance, life 216 00:14:31,640 --> 00:14:33,940 insurance, what have you. 217 00:14:33,940 --> 00:14:42,760 If the AI was automatically saying approved or denied, that's actually a high risk categorization 218 00:14:42,760 --> 00:14:45,440 because it could have a real detrimental impact. 219 00:14:45,440 --> 00:14:50,240 All of the biases in the model and the model will have some bias in it. 220 00:14:50,240 --> 00:14:56,640 This just inherent within AI systems could end up having detrimental outcomes for the 221 00:14:56,640 --> 00:14:57,640 human. 222 00:14:57,640 --> 00:15:04,960 What organizations have to do is assess the level of harm. 223 00:15:04,960 --> 00:15:11,040 If for instance, you're an AI company helping people write better subject lines for their 224 00:15:11,040 --> 00:15:18,960 email tools, it seems likely that this is a high risk application for AI. 225 00:15:18,960 --> 00:15:22,400 So you've probably got a low risk categorization, you need to be less concerned. 226 00:15:22,400 --> 00:15:27,600 But if you're deploying this in some sort of process where there is potential to cause 227 00:15:27,600 --> 00:15:33,200 harm, yeah, definitely one that you want to keep an eye on. 228 00:15:33,200 --> 00:15:43,480 In terms of on the global front, this is a world first and much like the EU's GDPR set 229 00:15:43,480 --> 00:15:51,260 the tone for other regulators around the world to think about how we use data, I think this 230 00:15:51,260 --> 00:16:00,240 is going to be a kind of scaffold upon which other regulators hang their own legislation 231 00:16:00,240 --> 00:16:01,240 from. 232 00:16:01,240 --> 00:16:07,040 They'll use that as the basis for their own legislation around the world. 233 00:16:07,040 --> 00:16:10,800 There was some discussion around copyright provisions. 234 00:16:10,800 --> 00:16:15,680 Obviously, copyright has been a major talking point when it comes to AI because generative 235 00:16:15,680 --> 00:16:26,760 AI in particular has caused lots of copyright holders to say, hey, you're scraping our content. 236 00:16:26,760 --> 00:16:33,440 I'm not entirely sure that they've fully resolved that other than saying that you have to make 237 00:16:33,440 --> 00:16:39,840 available the information that is in your training data. 238 00:16:39,840 --> 00:16:46,240 And that is going to be a bone of contention going forward, I dare say. 239 00:16:46,240 --> 00:16:53,520 Yeah, definitely for those commercial companies where OpenAI became, as Elon Musk would have 240 00:16:53,520 --> 00:16:57,440 them rebranded as closed AI and not really giving too much detail about how the models 241 00:16:57,440 --> 00:17:00,240 are trained and how the models work. 242 00:17:00,240 --> 00:17:06,860 Yeah, I guess if I think through what the wider spread ramifications could be, is those 243 00:17:06,860 --> 00:17:13,720 in Europe could see that their access to some tools like ChatGPT, Copilot, might be paused 244 00:17:13,720 --> 00:17:23,040 for a short or perhaps longer period while OpenAI and whoever gets in line with those 245 00:17:23,040 --> 00:17:24,040 regulations. 246 00:17:24,040 --> 00:17:26,800 I think that would be quite interesting to see how that plays out. 247 00:17:26,800 --> 00:17:27,800 Absolutely. 248 00:17:27,800 --> 00:17:34,560 We saw when ChatGPT was launched, there was pushback in Italy and Italian users couldn't 249 00:17:34,560 --> 00:17:38,000 use ChatGPT for a short period of time. 250 00:17:38,000 --> 00:17:40,720 So there is precedent for this. 251 00:17:40,720 --> 00:17:47,420 I would expect that there's been enough dialogue between the leaders of the tech companies 252 00:17:47,420 --> 00:17:55,760 and the regulators to find some compromise, but we know that the EU doesn't back down 253 00:17:55,760 --> 00:18:03,280 when it gets the smell of blood or I don't know what the right phrase would be, but they 254 00:18:03,280 --> 00:18:06,200 don't give an inch really. 255 00:18:06,200 --> 00:18:11,560 Tech companies have had to pay significant fines for GDPR breaches and things like that. 256 00:18:11,560 --> 00:18:16,200 So this is going to be the next battleground, I think. 257 00:18:16,200 --> 00:18:19,920 So something important that I think we'll all have to keep an eye on there, Martin. 258 00:18:19,920 --> 00:18:22,160 Thank you for taking us through that. 259 00:18:22,160 --> 00:18:27,520 In the next story, we're going to talk very briefly about Clawed 3 and the new models 260 00:18:27,520 --> 00:18:28,520 from Clawed 3. 261 00:18:28,520 --> 00:18:32,600 We touched on this in the last episode, but until this week, in fact, I think it was like 262 00:18:32,600 --> 00:18:36,480 yesterday or the day before, you couldn't access the smallest model of Clawed 3, which 263 00:18:36,480 --> 00:18:38,980 is called Haiku. 264 00:18:38,980 --> 00:18:43,600 And it's quite an interesting one because it's like super fast, super low latency, very 265 00:18:43,600 --> 00:18:47,360 cost effective, especially compared to other models. 266 00:18:47,360 --> 00:18:52,060 So to give you an example, it can process approximately 21,000 tokens per second, which 267 00:18:52,060 --> 00:18:58,320 is about 30 pages, which makes it about three times faster than competitors for the majority 268 00:18:58,320 --> 00:18:59,320 of workloads. 269 00:18:59,320 --> 00:19:04,960 And it's also extremely cheap, so to give you an example of this, it can analyze 400 270 00:19:04,960 --> 00:19:08,760 Supreme Court cases for just a dollar. 271 00:19:08,760 --> 00:19:17,360 But of course, it's not at the same level as GBT4 and Clawed 3 Opus. 272 00:19:17,360 --> 00:19:25,040 So we're expecting that this will be mostly leveraged for behind the scenes, high bandwidth 273 00:19:25,040 --> 00:19:31,160 work, like summarizing lots of cool transcripts for sentiment analysis and those types of 274 00:19:31,160 --> 00:19:40,040 things rather than content creation or brainstorming ideation, logical data processing. 275 00:19:40,040 --> 00:19:45,840 It's a bit more of a behind the scenes sort of carry out all of the work. 276 00:19:45,840 --> 00:19:49,960 But what we love about it, Martin, is it's got image and vision capabilities, right, 277 00:19:49,960 --> 00:19:50,960 which is pretty cool. 278 00:19:50,960 --> 00:19:56,240 Yeah, that is the really interesting thing, because you mentioned it can analyze 400 Supreme 279 00:19:56,240 --> 00:19:58,080 Court cases for a dollar. 280 00:19:58,080 --> 00:20:04,280 Well, it can also process 2,500 images for just a dollar. 281 00:20:04,280 --> 00:20:06,640 That's an incredible amount of image processing. 282 00:20:06,640 --> 00:20:08,240 Now, what can you do with that? 283 00:20:08,240 --> 00:20:14,920 Well, you can turn data from a photo into structured text data, because it can do the 284 00:20:14,920 --> 00:20:20,920 extraction and then put that into digitized tabulated data. 285 00:20:20,920 --> 00:20:26,240 So if you've got lots of handwritten forms that have been given to your company, maybe 286 00:20:26,240 --> 00:20:32,200 you've got an archive of data that you need digitizing, now you can scan all of those 287 00:20:32,200 --> 00:20:41,080 sheets, run them through this model and get all of the text extract neatly formulated. 288 00:20:41,080 --> 00:20:47,440 You could pass it out in JSON files, put it into CSV or tabulated data, however you want 289 00:20:47,440 --> 00:20:50,160 it formatted, it could do it for you. 290 00:20:50,160 --> 00:20:51,840 And it's incredibly cheap. 291 00:20:51,840 --> 00:20:54,000 Yeah, I think it's super interesting. 292 00:20:54,000 --> 00:21:00,480 I can definitely imagine taking screen grabs of software routinely while a user uses it 293 00:21:00,480 --> 00:21:04,760 and then trying to use that to create documentation about how that software works for creating 294 00:21:04,760 --> 00:21:06,560 things like software manuals. 295 00:21:06,560 --> 00:21:09,600 Yeah, I think you're right. 296 00:21:09,600 --> 00:21:16,280 Again, creativity and imagination in terms of now I have this capability at low latency 297 00:21:16,280 --> 00:21:21,920 and low cost, what can I do now, especially using those vision capabilities that I couldn't 298 00:21:21,920 --> 00:21:23,120 do before? 299 00:21:23,120 --> 00:21:29,440 Because certainly if you use, for example, like I talked a lot about Magi on the podcast, 300 00:21:29,440 --> 00:21:34,160 that has the vision capabilities of Gemini and ChatGPT built into it. 301 00:21:34,160 --> 00:21:39,920 But the cost of using those tools is more than just using text alone, because it costs 302 00:21:39,920 --> 00:21:41,280 more to use those APIs. 303 00:21:41,280 --> 00:21:45,080 So to have access to a vision tool at low cost, I think that really does open up some 304 00:21:45,080 --> 00:21:47,920 interesting capabilities. 305 00:21:47,920 --> 00:21:53,560 I think the important thing for users when it comes to Haiku is to remember that this 306 00:21:53,560 --> 00:21:58,120 isn't about the creative outputs or the reasoning abilities. 307 00:21:58,120 --> 00:22:01,640 This is the workhorse as you identified. 308 00:22:01,640 --> 00:22:07,680 This is about executing tasks, formatting data, extracting data, summarization. 309 00:22:07,680 --> 00:22:11,280 It's those kinds of applications where this comes into its own. 310 00:22:11,280 --> 00:22:12,280 Completely agree. 311 00:22:12,280 --> 00:22:19,320 So next story, there's an interesting quote from Sam Altman doing the rounds on the socials 312 00:22:19,320 --> 00:22:26,080 at the moment, originally uncovered, or at least to us by Paul Reutzer and Mike Kaput 313 00:22:26,080 --> 00:22:27,080 from Sam Altman. 314 00:22:27,080 --> 00:22:28,200 Tell us about this story, Martin. 315 00:22:28,200 --> 00:22:33,360 In an interview recently, Sam Altman said that artificial general intelligence will 316 00:22:33,360 --> 00:22:41,760 take over 95% of tasks currently done by marketing agencies, strategists, and creative professionals. 317 00:22:41,760 --> 00:22:47,800 They said that the integration of AI into marketing, it's already happening and generative 318 00:22:47,800 --> 00:22:53,520 AI is taking on more and more of these tasks within the industry. 319 00:22:53,520 --> 00:22:55,680 So that is the next frontier. 320 00:22:55,680 --> 00:23:01,560 As soon as we get to AGI, the marketing industry needs to watch out. 321 00:23:01,560 --> 00:23:09,760 Now, in terms of the timeframe for that, Sam was saying approximately five years. 322 00:23:09,760 --> 00:23:14,040 So if you're a marketing agency owner, Paul, I don't know if you know any, you might have 323 00:23:14,040 --> 00:23:15,800 to pay attention. 324 00:23:15,800 --> 00:23:18,480 This is going to be a bit disruptive according to Sam. 325 00:23:18,480 --> 00:23:20,480 What do you think? 326 00:23:20,480 --> 00:23:21,480 Yeah. 327 00:23:21,480 --> 00:23:24,160 I think this is really interesting. 328 00:23:24,160 --> 00:23:30,720 But after I digested it for a while, I really thought about, well, this is probably true 329 00:23:30,720 --> 00:23:32,440 of most of knowledge work. 330 00:23:32,440 --> 00:23:40,640 So you could probably swap out marketing agency for counting, consultancy, a huge amount of 331 00:23:40,640 --> 00:23:41,920 other industries. 332 00:23:41,920 --> 00:23:49,320 I think what makes the marketing part interesting is the creative aspects, right? 333 00:23:49,320 --> 00:23:53,280 Running creative campaigns, analyzing the data, and then improving those, which is somewhere 334 00:23:53,280 --> 00:23:55,240 that I think AI can help a lot. 335 00:23:55,240 --> 00:23:57,560 But creative production is hard. 336 00:23:57,560 --> 00:24:02,480 We've talked about the limitations of the current tool set a lot already. 337 00:24:02,480 --> 00:24:07,960 So I think that's where I almost, for good or real, see marketing as quite a high bar 338 00:24:07,960 --> 00:24:13,560 because it's content creation, it's strategy, it's data analysis, it's creative outputs, 339 00:24:13,560 --> 00:24:16,240 it's empathy and customer understanding. 340 00:24:16,240 --> 00:24:21,400 And I think if AI can do a lot of those things, it can probably do a lot of jobs. 341 00:24:21,400 --> 00:24:27,480 The other thing is the references here in terms of AGI, or by definition, AGI can do 342 00:24:27,480 --> 00:24:29,640 lots of things that humans can do. 343 00:24:29,640 --> 00:24:33,680 So I think we're still on that timeline towards AGI. 344 00:24:33,680 --> 00:24:38,360 When we hit AGI, the world's going to be very, very different place to what it is now. 345 00:24:38,360 --> 00:24:43,600 And I think assuming this is possible, come onto that in a minute, but assuming this is 346 00:24:43,600 --> 00:24:49,640 possible, our economy has to be fundamentally rethought in terms of what do humans do with 347 00:24:49,640 --> 00:24:50,640 their time? 348 00:24:50,640 --> 00:24:51,640 How do we have purpose? 349 00:24:51,640 --> 00:24:53,340 How do we deliver value? 350 00:24:53,340 --> 00:24:57,000 How does the economy work from a monetary perspective? 351 00:24:57,000 --> 00:24:58,840 How do we all have money to live? 352 00:24:58,840 --> 00:25:04,400 Do we live in a world where everybody's got a robot that helps them and we're all doing 353 00:25:04,400 --> 00:25:08,720 fantastic pursuits and it's kind of like Star Trek? 354 00:25:08,720 --> 00:25:11,240 Or is it going to be more like, what's that movie? 355 00:25:11,240 --> 00:25:12,240 Is it Wally? 356 00:25:12,240 --> 00:25:16,000 Is that the character where all of the humans basically become extremely overweight and 357 00:25:16,000 --> 00:25:20,880 live in these portable pods where they don't really do anything for themselves anymore 358 00:25:20,880 --> 00:25:22,400 because the computers can do it? 359 00:25:22,400 --> 00:25:25,920 Or that would be the more dystopian view potentially. 360 00:25:25,920 --> 00:25:31,480 But all of this is dependent on the current tech architectures that we have getting us 361 00:25:31,480 --> 00:25:36,160 to an AI that can basically do most, if not all things that humans can do better than 362 00:25:36,160 --> 00:25:37,160 a human. 363 00:25:37,160 --> 00:25:39,320 And I'm just stuck on that. 364 00:25:39,320 --> 00:25:41,680 I feel the exponential curve were on mine. 365 00:25:41,680 --> 00:25:42,680 I really do. 366 00:25:42,680 --> 00:25:44,760 So it feels plausible. 367 00:25:44,760 --> 00:25:51,220 But I also know that there have been other AI areas of enthusiasm followed by AI winters 368 00:25:51,220 --> 00:25:55,120 where we hit a barrier that we couldn't get through. 369 00:25:55,120 --> 00:25:58,560 And we'll talk about some other awesome stuff that Meta and other companies are doing at 370 00:25:58,560 --> 00:26:03,680 the moment, but this kind of assumes the current paradigms of models that we're creating will 371 00:26:03,680 --> 00:26:06,280 get us there and they may not. 372 00:26:06,280 --> 00:26:07,800 So we just don't know. 373 00:26:07,800 --> 00:26:09,400 Do you have a perspective on this? 374 00:26:09,400 --> 00:26:12,520 I think it's interesting to see where it's currently at. 375 00:26:12,520 --> 00:26:15,560 And we saw this in the workshop this week. 376 00:26:15,560 --> 00:26:26,840 If you get a team of humans and give them the AI tools, the speed at which that creative 377 00:26:26,840 --> 00:26:33,540 development, that iteration, that first draft, a lot of work can be done in a very short 378 00:26:33,540 --> 00:26:34,540 window of time. 379 00:26:34,540 --> 00:26:38,520 I think some of the copywriting that was done in some of the examples that we saw with the 380 00:26:38,520 --> 00:26:43,560 team we were working with this week, it was work that was done in about eight minutes 381 00:26:43,560 --> 00:26:47,900 that was really quite impressive. 382 00:26:47,900 --> 00:26:53,000 Not the finished article, but really impressive. 383 00:26:53,000 --> 00:26:57,980 And that was many humans working with AI as a tool to assist them. 384 00:26:57,980 --> 00:27:06,420 I think the notion of completely, you know, 95% of the tasks, that seems high to me. 385 00:27:06,420 --> 00:27:14,760 In five years time, 95% of tasks being done by AI does seem high, but then pace of change. 386 00:27:14,760 --> 00:27:20,880 And when you break down what some of the tasks that we do, whether it's project management, 387 00:27:20,880 --> 00:27:24,280 I'm thinking, yeah, well, okay, the AI will get better at copywriting. 388 00:27:24,280 --> 00:27:30,160 It will get better at creative development and concepts for design or from video. 389 00:27:30,160 --> 00:27:36,000 And then thinking in five years time, crikey, a lot has changed in one year, nevermind five 390 00:27:36,000 --> 00:27:37,000 years time. 391 00:27:37,000 --> 00:27:45,040 So we will definitely be further along than which bits of the marketing agency tasks and 392 00:27:45,040 --> 00:27:49,280 task workload isn't done by AI. 393 00:27:49,280 --> 00:27:55,800 I find myself thinking, well, actually maybe, maybe he's right. 394 00:27:55,800 --> 00:27:56,800 I swing back and forth. 395 00:27:56,800 --> 00:28:04,520 AGI as a term I struggle with because exactly what you've said, we've had these AI predictions 396 00:28:04,520 --> 00:28:09,280 in the past followed by AI winters where you're looking for the next breakthrough that's going 397 00:28:09,280 --> 00:28:14,280 to take us into the next epoch of AI activity. 398 00:28:14,280 --> 00:28:22,120 Will we get at the current level of AI with LLMs and existing model architectures? 399 00:28:22,120 --> 00:28:24,320 Will that replace marketing agencies? 400 00:28:24,320 --> 00:28:26,400 It's a big ask. 401 00:28:26,400 --> 00:28:28,240 I'm not going to profess to have the answer. 402 00:28:28,240 --> 00:28:29,240 Yeah. 403 00:28:29,240 --> 00:28:33,680 And I think the other thing that these predictions sometimes lack is we may have the technical 404 00:28:33,680 --> 00:28:39,360 capabilities to do so, but there are a lot of things that have to happen to make this 405 00:28:39,360 --> 00:28:40,360 possible. 406 00:28:40,360 --> 00:28:44,960 So even what we're seeing in the work that we're doing, there's the technical capabilities 407 00:28:44,960 --> 00:28:48,640 of the AI, but then there's also the training of the people to use it. 408 00:28:48,640 --> 00:28:54,280 There's also the willingness to change, which some humans in general are pretty good at 409 00:28:54,280 --> 00:29:00,100 change in terms of we're adaptable, but most humans on the micro level resist change. 410 00:29:00,100 --> 00:29:03,920 So how do you manage that change management aspect? 411 00:29:03,920 --> 00:29:09,960 Then from a how do the governments of the world, regulators, how do they manage the 412 00:29:09,960 --> 00:29:14,640 emergence of these tools because of the disruption that will bring to our way of life and the 413 00:29:14,640 --> 00:29:15,640 economy? 414 00:29:15,640 --> 00:29:23,280 So again, we may be able, technically able to do some of this stuff, but to my understanding, 415 00:29:23,280 --> 00:29:27,560 humans will have to choose to implement it and will they? 416 00:29:27,560 --> 00:29:32,920 How will they implement it and how easy will that be and how long will it take? 417 00:29:32,920 --> 00:29:36,840 So then I just look at all those different streams of things that have to happen. 418 00:29:36,840 --> 00:29:38,840 And I'm not saying that this isn't where we're going to be. 419 00:29:38,840 --> 00:29:43,960 I just think it's probably more complex than can we technically do it with the computing 420 00:29:43,960 --> 00:29:45,760 power and architectures we have. 421 00:29:45,760 --> 00:29:53,840 Yeah, I've no doubt that in five years time, there will be kind of marketing agency as 422 00:29:53,840 --> 00:30:00,480 a service subscription style AI powered tools available where people have cobbled together 423 00:30:00,480 --> 00:30:05,240 a bunch of different AI agents and they can execute very specific things. 424 00:30:05,240 --> 00:30:12,200 You might want it to come up with a campaign concept and then give you a kind of channel 425 00:30:12,200 --> 00:30:14,600 strategy for some advertising. 426 00:30:14,600 --> 00:30:19,680 It will design you a landing page and you can press deploy and you will get a campaign 427 00:30:19,680 --> 00:30:27,240 designed, copywritten, pushed out with some data capture forms and landing pages and integrations 428 00:30:27,240 --> 00:30:31,440 into Meta and all of the other advertising platforms. 429 00:30:31,440 --> 00:30:37,080 And that's a good chunk of what a digital marketing agency might do for somebody. 430 00:30:37,080 --> 00:30:42,960 But it's not everything, but if you're a smaller business, that might be perfect for what you 431 00:30:42,960 --> 00:30:52,040 need if you're looking for a global omni-channel product launch and brand activation. 432 00:30:52,040 --> 00:30:59,800 Do I think an AI is capable of doing all of that with all of the human logistics and coordination 433 00:30:59,800 --> 00:31:01,800 and that's required? 434 00:31:01,800 --> 00:31:02,800 No. 435 00:31:02,800 --> 00:31:07,320 I think it's, I love the fact you just brought that up because if just thinking about and 436 00:31:07,320 --> 00:31:14,120 taking it through to a more extreme logical conclusion, if everybody has access to this 437 00:31:14,120 --> 00:31:20,480 tool then everybody can use that marketing in a box tool that you described. 438 00:31:20,480 --> 00:31:23,640 That's the new bar, that's new table stakes. 439 00:31:23,640 --> 00:31:28,740 And in the battle for human attention, and we've talked about this in previous workshops 440 00:31:28,740 --> 00:31:36,400 and podcasts about the deluge of rubbish content that ChatGPT and other large language models 441 00:31:36,400 --> 00:31:37,400 could enable. 442 00:31:37,400 --> 00:31:42,400 And being able to churn out a load of rubbish content is useful for a short period, but 443 00:31:42,400 --> 00:31:46,920 we're already seeing that Google is now starting to penalize those sites that it feels have 444 00:31:46,920 --> 00:31:52,480 added lots of content quickly and have a good feeling that that content is AI generated for 445 00:31:52,480 --> 00:31:57,720 SEO and to appease search engines not to provide value to humans. 446 00:31:57,720 --> 00:32:01,680 And so we're going to see if everybody's got access to these tools, that's now the lowest 447 00:32:01,680 --> 00:32:02,680 common denominator. 448 00:32:02,680 --> 00:32:04,000 And how do you go above that? 449 00:32:04,000 --> 00:32:07,680 Well, maybe there's other tools that you pay more money for, but I think again, that's 450 00:32:07,680 --> 00:32:10,640 where humans will give you that creative competitive edge. 451 00:32:10,640 --> 00:32:16,720 So these tools in hands of absolute domain experts are going to get better results than 452 00:32:16,720 --> 00:32:19,120 these tools in the hands of non-domain experts. 453 00:32:19,120 --> 00:32:22,840 And so all it does is just push the bar up for everyone. 454 00:32:22,840 --> 00:32:28,400 And we get a good insight into this in the life sciences and we worked in the life sciences 455 00:32:28,400 --> 00:32:29,400 this week. 456 00:32:29,400 --> 00:32:35,080 I know it's not an area you haven't worked as much in, but hopefully it was quite interesting 457 00:32:35,080 --> 00:32:40,560 to see just how incapable some of the tools are in some niche domains. 458 00:32:40,560 --> 00:32:45,000 We did some image generation work, trying to produce scientifically accurate images 459 00:32:45,000 --> 00:32:47,960 and all of the tools completely suck at it. 460 00:32:47,960 --> 00:32:54,320 You can't use Mid Journey or any other tool for any type of image production work in the 461 00:32:54,320 --> 00:33:00,080 life sciences outside of stock photos of people in laboratories or maybe some abstract images 462 00:33:00,080 --> 00:33:03,800 that you think are like interesting and beautiful, but that you're not claiming are scientifically 463 00:33:03,800 --> 00:33:05,400 accurate in any way. 464 00:33:05,400 --> 00:33:09,520 And to be quite honest, Chatch, E.P.T. and other tools, they struggle at written content 465 00:33:09,520 --> 00:33:15,480 creation as well because they just don't have the deep domain expertise that a subject matter 466 00:33:15,480 --> 00:33:20,400 expert has either within a team like mine, we've got a bunch of PhD level science writers 467 00:33:20,400 --> 00:33:25,360 and certainly within life science businesses as a whole in terms of all of the subject 468 00:33:25,360 --> 00:33:30,520 matter experts that they have and the domain knowledge on the cutting edge that these tools 469 00:33:30,520 --> 00:33:31,520 don't have. 470 00:33:31,520 --> 00:33:35,600 And I see that as a bit of a prelude to this, which is, like you say, it pulls the average 471 00:33:35,600 --> 00:33:40,200 up but there's going to be a load of use cases where it's not fit for purpose and it will 472 00:33:40,200 --> 00:33:45,200 get better over time, but I think that will take some time. 473 00:33:45,200 --> 00:33:51,080 Let's move into our next story then, which is about IBM announcing layoffs in its communications 474 00:33:51,080 --> 00:33:54,680 and marketing divisions, probably somewhat related to the conversation we were just having 475 00:33:54,680 --> 00:33:55,680 Martin. 476 00:33:55,680 --> 00:34:00,720 And this comes hot off of the press of the Klarna examples that we talked about in the 477 00:34:00,720 --> 00:34:06,040 previous podcast where they were able to replace human customer service reps with AI driven 478 00:34:06,040 --> 00:34:10,560 chat bots to boost profit margins. 479 00:34:10,560 --> 00:34:15,160 So IBM has initiated layoffs within its communications and marketing staff, affecting a lot of the 480 00:34:15,160 --> 00:34:19,280 low single digit performance of a percentage of its global workforce. 481 00:34:19,280 --> 00:34:23,080 The move is part of a broader workforce rebalancing effort announced earlier in the year during 482 00:34:23,080 --> 00:34:27,920 the company's fourth quarter earnings, where IBM aims to maintain its employment levels 483 00:34:27,920 --> 00:34:33,260 through 2024 by focusing on areas of high demand among its clients, such as AI and hybrid 484 00:34:33,260 --> 00:34:35,900 cloud technologies. 485 00:34:35,900 --> 00:34:41,380 So this is quite interesting because these are Marcom's jobs that are going and they 486 00:34:41,380 --> 00:34:46,100 follow a recent announcement from IBM about the impact of AI on the workforce. 487 00:34:46,100 --> 00:34:51,280 So this makes this a particularly pertinent story for us to be covering, doesn't it Martin? 488 00:34:51,280 --> 00:34:52,280 Yeah. 489 00:34:52,280 --> 00:35:02,400 And it comes hot off the tail of them announcing a hiring pause for approximately 7,800 positions, 490 00:35:02,400 --> 00:35:07,360 which they say are potentially replaceable by AI technology. 491 00:35:07,360 --> 00:35:12,880 So these are predominantly back office jobs and this was announced late last year. 492 00:35:12,880 --> 00:35:18,720 And at the time we discussed this on the podcast and now we're starting to see this come to 493 00:35:18,720 --> 00:35:19,720 fruition. 494 00:35:19,720 --> 00:35:26,720 And I think considering Sam Altman's story and the reporting of 95% of Marcom's agency 495 00:35:26,720 --> 00:35:32,920 jobs or tasks being effectively done by AI, well, IBM are trying to turn that into a reality 496 00:35:32,920 --> 00:35:35,120 with their internal team. 497 00:35:35,120 --> 00:35:40,360 They are still investing, but they're looking to invest in their teams that are working 498 00:35:40,360 --> 00:35:43,480 on AI technologies, on their cloud technologies. 499 00:35:43,480 --> 00:35:50,480 So Watson X Development Studio is aiming to train companies in building and releasing 500 00:35:50,480 --> 00:35:52,480 machine learning models. 501 00:35:52,480 --> 00:35:54,960 This is where they're putting their focus. 502 00:35:54,960 --> 00:36:00,180 I suspect that huge amounts of the work that was previously done by their global Marcom's 503 00:36:00,180 --> 00:36:06,080 team will still get done in-house, but it will be done by Watson. 504 00:36:06,080 --> 00:36:07,080 Yeah. 505 00:36:07,080 --> 00:36:08,080 It's an interesting one. 506 00:36:08,080 --> 00:36:13,880 I mean, we have the story from last week, right, about Klarna's customer service delivery, 507 00:36:13,880 --> 00:36:20,240 where we don't have any data outside of a small amount of information from the company. 508 00:36:20,240 --> 00:36:23,560 And they have obviously been questions that have been levied at the company. 509 00:36:23,560 --> 00:36:27,280 And at least as time of this recording, I don't know if you know different Martin, but 510 00:36:27,280 --> 00:36:31,760 from what I've seen so far, they haven't really been able to respond to a lot of those questions. 511 00:36:31,760 --> 00:36:39,300 And it's not that we're not saying that that use case wasn't real, but we haven't had the 512 00:36:39,300 --> 00:36:41,200 ability to really scrutinize it. 513 00:36:41,200 --> 00:36:43,280 And we were reading a report on the plane, weren't we? 514 00:36:43,280 --> 00:36:50,520 We were flying over to the US, questioning the real world impact of AI and where are 515 00:36:50,520 --> 00:36:55,400 all the applications that massively boost efficiency and effectiveness and drive shareholder 516 00:36:55,400 --> 00:36:58,520 value and increase profits and go to the bottom line and all that. 517 00:36:58,520 --> 00:37:03,080 And they are kind of few and far between in terms of written case studies, but I hold 518 00:37:03,080 --> 00:37:09,100 that in tension with the businesses that I know that are getting the 10, 20, 30% productivity 519 00:37:09,100 --> 00:37:12,120 and efficiency gains using the tools. 520 00:37:12,120 --> 00:37:18,520 So I still think we're in the middle of trying to see the productivity gains and efficiency 521 00:37:18,520 --> 00:37:20,080 gains that these tools really allow. 522 00:37:20,080 --> 00:37:24,720 But when we see moves like this by IBM, we have to assume that just because they're not 523 00:37:24,720 --> 00:37:28,440 reporting what they are, they are seeing them and that gives them the confidence to make 524 00:37:28,440 --> 00:37:30,440 these infrastructure changes. 525 00:37:30,440 --> 00:37:33,920 Yeah, it's interesting you mentioned the report from the information. 526 00:37:33,920 --> 00:37:39,040 It was on the information that they said big cloud providers, so sales teams at the likes 527 00:37:39,040 --> 00:37:47,920 of Google Cloud, Amazon AWS are all saying to dampen expectations of sales of generative 528 00:37:47,920 --> 00:37:51,520 AI integrations and services. 529 00:37:51,520 --> 00:37:58,360 And that's predominantly through clients' hesitance to implement them in customer facing 530 00:37:58,360 --> 00:38:03,500 applications, driven by inaccuracy and hallucinations. 531 00:38:03,500 --> 00:38:06,680 That seems to be the big issue for people at the moment. 532 00:38:06,680 --> 00:38:08,280 And that's completely understandable. 533 00:38:08,280 --> 00:38:14,800 Even with rag-based systems, I've seen reports saying that rag-based integrations are 70 534 00:38:14,800 --> 00:38:22,600 or 80% accurate, but 70 or 80% accurate still leaves a lot of room for improvement. 535 00:38:22,600 --> 00:38:31,160 If you're getting a 20 or 30% error rate or just lacking in the right bit of detail when 536 00:38:31,160 --> 00:38:35,240 responding to customers, that's not going to be production ready. 537 00:38:35,240 --> 00:38:40,440 So yeah, I think we're going to see more and more applications for generative AI in the 538 00:38:40,440 --> 00:38:46,600 back office, for internal processes, for alleviating all of those jobs in an organization where 539 00:38:46,600 --> 00:38:52,040 it's somebody taking data from this place to another place and things like that. 540 00:38:52,040 --> 00:38:58,640 And this takes us back to the reason why the Haiku API deployment is so interesting. 541 00:38:58,640 --> 00:38:59,840 I agree. 542 00:38:59,840 --> 00:39:05,840 So to add further complexity to this discussion in terms of how powerful is AI, what can it 543 00:39:05,840 --> 00:39:10,400 do, is it really adding productivity gains, probably because of some of these changes 544 00:39:10,400 --> 00:39:14,600 we're seeing within businesses, but how fast will we get to a position where it's really 545 00:39:14,600 --> 00:39:18,360 taking a huge amount of work away from humans? 546 00:39:18,360 --> 00:39:24,480 Meta's had a bit of a leap supposedly in their development of their AI tools, haven't they 547 00:39:24,480 --> 00:39:25,480 Martin? 548 00:39:25,480 --> 00:39:28,240 This is very much on the engineering side of things. 549 00:39:28,240 --> 00:39:35,600 On the Meta engineering team blog, they announced some massive investment basically, and the 550 00:39:35,600 --> 00:39:44,400 future of their technology stack for developing and rolling out AI models that we can all 551 00:39:44,400 --> 00:39:46,600 use in the future. 552 00:39:46,600 --> 00:39:55,560 I'm going to try to avoid getting too into the detail on this, but at its core, what 553 00:39:55,560 --> 00:40:02,880 they're saying is that they are going to deploy a new range of AI plusters within their data 554 00:40:02,880 --> 00:40:13,120 centers, so the infrastructure is powered by 24,576 Nvidia Tensor Core H100 GPUs. 555 00:40:13,120 --> 00:40:20,880 Now they are the most cutting edge chips for AI processing available on the market, and 556 00:40:20,880 --> 00:40:25,240 they're a significant upgrade on what they've been using to date. 557 00:40:25,240 --> 00:40:30,400 I think when you start to look at the numbers around this, you start to see how seriously 558 00:40:30,400 --> 00:40:38,680 these organizations, Big Tech, are taking AI development, research and deployment. 559 00:40:38,680 --> 00:40:47,160 So they're looking to expand the H100 GPU count by an additional 350,000 units. 560 00:40:47,160 --> 00:40:48,160 Okay? 561 00:40:48,160 --> 00:40:50,920 So that's kind of abstract. 562 00:40:50,920 --> 00:40:55,960 You hear that and you go, okay, they're investing in a lot of chipsets, 350,000 of them. 563 00:40:55,960 --> 00:40:56,960 Great. 564 00:40:56,960 --> 00:41:05,040 They're aiming to have a total compute power equivalent to 600,000 of these units by the 565 00:41:05,040 --> 00:41:07,600 end of 2024. 566 00:41:07,600 --> 00:41:10,640 So that's massive expansion of compute power. 567 00:41:10,640 --> 00:41:21,640 Just to put this into an investment perspective, the book price for a H100 GPU is between 30 568 00:41:21,640 --> 00:41:31,880 and 35,000 pounds per GPU, which ballparks, you know, plus or minus a billion. 569 00:41:31,880 --> 00:41:38,880 That's a 10 billion pound, yeah, 10 billion pound investment in these chipsets. 570 00:41:38,880 --> 00:41:44,480 And I'm sure the commercial agreement is very different from me walking into PC World and 571 00:41:44,480 --> 00:41:47,880 asking for a H100 chip. 572 00:41:47,880 --> 00:41:51,520 Sure they've got a slightly different setup with Nvidia, but it goes to show that they 573 00:41:51,520 --> 00:41:59,440 are really backing up their research team with some real GPU horsepower. 574 00:41:59,440 --> 00:42:00,440 It's pretty cool. 575 00:42:00,440 --> 00:42:06,760 The Yan LeCun, who's head of all of this over at Meta, has been doing the rounds on a few 576 00:42:06,760 --> 00:42:14,600 podcasts recently and re-referencing the iJEPA models that we've covered previously on the 577 00:42:14,600 --> 00:42:15,600 podcast. 578 00:42:15,600 --> 00:42:21,100 So JEPA stands for Joint Embedding Predictive Architecture, where again, this probably stretches 579 00:42:21,100 --> 00:42:27,080 into the machine learning details and probably beyond the expertise, certainly of me and potentially 580 00:42:27,080 --> 00:42:28,420 both of us. 581 00:42:28,420 --> 00:42:34,880 But it's about providing AI with the opportunity to learn in different ways from how large 582 00:42:34,880 --> 00:42:40,060 language models have learned from the data they've been fed so far in a sort of passive 583 00:42:40,060 --> 00:42:45,040 mechanism of having an AI learn about the world around it a little bit like a two-year-old 584 00:42:45,040 --> 00:42:47,680 would just by observing what's going on. 585 00:42:47,680 --> 00:42:50,720 One assumes that's going to take massive computing power. 586 00:42:50,720 --> 00:42:56,240 And so when you look at these investments from Meta, this isn't just a power that, 587 00:42:56,240 --> 00:42:58,040 you know, Lama 3. 588 00:42:58,040 --> 00:43:03,040 This is because they have these other architectures and these other mechanisms of getting us to 589 00:43:03,040 --> 00:43:08,040 artificial general intelligence that they want to explore and need massive computing 590 00:43:08,040 --> 00:43:09,240 power to enable. 591 00:43:09,240 --> 00:43:14,600 So when we're talking about AGI in five years and all the things that, all the human jobs 592 00:43:14,600 --> 00:43:18,720 that it could potentially replace, as Sab Altman said, and whether or not we've got 593 00:43:18,720 --> 00:43:22,840 the architectures to do that, Yanlokun is one of the people who doesn't believe we have 594 00:43:22,840 --> 00:43:24,120 the architectures to do that. 595 00:43:24,120 --> 00:43:28,520 But at the same time, it's part of a meta team that's spending 10 billion pounds on 596 00:43:28,520 --> 00:43:34,760 all of this compute power to discover those architectures and test them and train them. 597 00:43:34,760 --> 00:43:39,320 And of course, already having some interesting ideas on what those architectures should be. 598 00:43:39,320 --> 00:43:41,540 And we've talked about them on the podcast as well. 599 00:43:41,540 --> 00:43:47,720 We've got their vision segmentation model that's able to detect things in images, which 600 00:43:47,720 --> 00:43:51,920 is probably either that technology or similar technology that's underpinning a lot of the 601 00:43:51,920 --> 00:43:54,640 vision capabilities that we're seeing at the moment. 602 00:43:54,640 --> 00:43:59,160 Their audio translation capabilities are more powerful than any of the tools that I've seen 603 00:43:59,160 --> 00:44:02,720 commercially available so far when you look at the demo videos. 604 00:44:02,720 --> 00:44:07,520 And this investment will just go to further fuel all of that stuff. 605 00:44:07,520 --> 00:44:14,340 You mentioned Yanlokun there and him doing the media rounds recently, and he's really 606 00:44:14,340 --> 00:44:19,680 not keen on the term AGI, which we've talked about on the podcast today. 607 00:44:19,680 --> 00:44:25,040 But if you hear him talk about it, it's not a specific phrase that he likes. 608 00:44:25,040 --> 00:44:28,200 He prefers to use a different terminology. 609 00:44:28,200 --> 00:44:33,800 But if you read the engineering blog that Meta have put out, they actually have a graphic 610 00:44:33,800 --> 00:44:40,260 that talks explicitly about this new cluster model and this major new investment driving 611 00:44:40,260 --> 00:44:42,320 directly through to AGI. 612 00:44:42,320 --> 00:44:49,560 They have a diagram that features AGI at the end of it, which fully enough takes a pathway 613 00:44:49,560 --> 00:44:50,560 through Lama 3. 614 00:44:50,560 --> 00:44:52,560 I read it does. 615 00:44:52,560 --> 00:44:57,640 Well, if they want to get there, they may need to be supported by tools like Devin, 616 00:44:57,640 --> 00:45:02,240 which was released this week, which is, if the hype is to be, believe Martin, the world's 617 00:45:02,240 --> 00:45:05,760 first fully autonomous AI software engineer. 618 00:45:05,760 --> 00:45:10,480 So it's been developed by a company called Cognition in the US, and it's designed to 619 00:45:10,480 --> 00:45:14,320 work autonomously with its own code editor, command line, and browser. 620 00:45:14,320 --> 00:45:19,200 It's supposedly being able to handle complex software engineering tasks. 621 00:45:19,200 --> 00:45:27,280 So this kind of forms an early indication of what a developer type AI agent would be 622 00:45:27,280 --> 00:45:30,880 in terms of taking multiple steps in software development. 623 00:45:30,880 --> 00:45:35,280 So it can plan and execute intricate engineering projects. 624 00:45:35,280 --> 00:45:38,800 It can make thousands of decisions, and it's equipped with different developer tools. 625 00:45:38,800 --> 00:45:42,940 It can collaborate with users, and it can also provide real-time progress reports and 626 00:45:42,940 --> 00:45:45,060 accept feedback from the users. 627 00:45:45,060 --> 00:45:49,640 So it's kind of interesting, but there's a but. 628 00:45:49,640 --> 00:45:55,540 And I think the but is it's kind of a cool leap forward, and best way that we can kind 629 00:45:55,540 --> 00:45:58,220 of put this in some context. 630 00:45:58,220 --> 00:46:04,040 So on a particular coding benchmark, so the SWE Bench Coding benchmark, Devin resolved 631 00:46:04,040 --> 00:46:10,640 nearly 14% of issues end to end, which is better than the previous state of the art, 632 00:46:10,640 --> 00:46:14,680 which I think is probably GPT-4, which was about 2%. 633 00:46:14,680 --> 00:46:23,700 So the move from 2% to 14% is significant, but 14% is still not that large. 634 00:46:23,700 --> 00:46:28,180 The other thing that's worth noting is that the previous state of the art was a models 635 00:46:28,180 --> 00:46:30,420 that were being assisted by humans. 636 00:46:30,420 --> 00:46:34,880 So for example, being told exactly which files to edit, whereas Devin was able to do this 637 00:46:34,880 --> 00:46:36,380 unassisted. 638 00:46:36,380 --> 00:46:40,920 So it's quite a big leap forward, but obviously there's quite a lot of space to improve. 639 00:46:40,920 --> 00:46:44,960 I saw Ethan Molykos got access to this already, Martin, and he's posted on LinkedIn about 640 00:46:44,960 --> 00:46:48,200 some of his experience with it. 641 00:46:48,200 --> 00:46:52,840 And that whilst it still has quite a long way to go, it's a good early indication of 642 00:46:52,840 --> 00:46:59,360 what some, what's the word we were using last week, agentic capabilities might be in these 643 00:46:59,360 --> 00:47:00,880 types of tools. 644 00:47:00,880 --> 00:47:03,840 Did you get, did you see this Devin use? 645 00:47:03,840 --> 00:47:06,420 The videos for it are really good. 646 00:47:06,420 --> 00:47:10,600 If you look on the blog from the company, they've got a series of one or two minute 647 00:47:10,600 --> 00:47:16,540 videos showcasing some specific examples of the tool in action. 648 00:47:16,540 --> 00:47:23,340 And while this isn't going to build you a really complex piece of software end to end 649 00:47:23,340 --> 00:47:33,240 and deploy it, it will do small tasks more completely than anything I've seen from any 650 00:47:33,240 --> 00:47:34,600 other tool. 651 00:47:34,600 --> 00:47:44,480 One example was somebody shared a link to a blog post about how to create the stable 652 00:47:44,480 --> 00:47:49,320 diffusion model that can do the text hidden within the image. 653 00:47:49,320 --> 00:47:52,600 I can't remember what that tool is. 654 00:47:52,600 --> 00:48:01,280 But the developer, she shared a link to that blog post and said, this article has some 655 00:48:01,280 --> 00:48:04,040 code in it to get this thing working. 656 00:48:04,040 --> 00:48:06,720 Can you create a tool that uses it? 657 00:48:06,720 --> 00:48:07,960 And that's what, that's it. 658 00:48:07,960 --> 00:48:08,960 That's it. 659 00:48:08,960 --> 00:48:14,480 So shares URL, inputs the blog, and then it goes off, creates a plan of the steps that 660 00:48:14,480 --> 00:48:20,720 it needs to execute and then execute each one of those steps successfully and then delivers 661 00:48:20,720 --> 00:48:25,400 back to her some sample images that it has just created with the software that it's just 662 00:48:25,400 --> 00:48:26,400 written. 663 00:48:26,400 --> 00:48:28,640 Now that's very cool. 664 00:48:28,640 --> 00:48:31,240 It isn't going to be the whole front end at this stage. 665 00:48:31,240 --> 00:48:37,320 It's like it said in the report, just shy of 14% of issues were resolved end to end. 666 00:48:37,320 --> 00:48:39,000 So there's still a long way to go. 667 00:48:39,000 --> 00:48:42,160 But it's significantly from where we were. 668 00:48:42,160 --> 00:48:46,760 Yeah, I remember when we were playing with AutoGPT, Martin, and we just, us and other 669 00:48:46,760 --> 00:48:50,520 people in our network, we're just struggling to get it to work and it's kind of disappeared. 670 00:48:50,520 --> 00:48:51,720 Nobody really talks about it anymore. 671 00:48:51,720 --> 00:48:54,440 And that's because I think mostly it didn't work very well. 672 00:48:54,440 --> 00:48:58,000 This is kind of an improvement on AutoGPT. 673 00:48:58,000 --> 00:48:59,000 It is. 674 00:48:59,000 --> 00:49:04,320 And there is a certain hype around this at the moment, which feels somewhat similar to 675 00:49:04,320 --> 00:49:07,080 the hype around AutoGPT. 676 00:49:07,080 --> 00:49:11,520 And I remember when AutoGPT came out, we were both somewhat skeptical. 677 00:49:11,520 --> 00:49:17,800 We couldn't get it to do very many interesting things while at the same time, AI influences 678 00:49:17,800 --> 00:49:21,360 on the Twittersphere, as on the Linkedins. 679 00:49:21,360 --> 00:49:27,640 We're saying that it can do all of these amazing things and jobs are no longer required and 680 00:49:27,640 --> 00:49:31,280 transpired to be inaccurate. 681 00:49:31,280 --> 00:49:33,960 Whereas this does feel like a genuine leap. 682 00:49:33,960 --> 00:49:39,360 And for a day one model, I'm looking at it going, okay, if I'm a software developer, 683 00:49:39,360 --> 00:49:43,560 I can see this having some very interesting use cases. 684 00:49:43,560 --> 00:49:44,560 Yeah. 685 00:49:44,560 --> 00:49:46,760 I think it's something for us to keep an eye on. 686 00:49:46,760 --> 00:49:53,320 I think it will be overhyped, but I do still also think it's an incremental move forward. 687 00:49:53,320 --> 00:49:56,560 Another thing that's likely to get overhyped is our last story this week, Martin. 688 00:49:56,560 --> 00:49:58,360 We'll talk about this one briefly. 689 00:49:58,360 --> 00:50:01,160 Is the collaboration between Figur and OpenAI. 690 00:50:01,160 --> 00:50:02,160 Yeah. 691 00:50:02,160 --> 00:50:08,120 The Figur company, which creates humanoid robots, has released a video where they've 692 00:50:08,120 --> 00:50:14,880 got their humanoid robot integrated with, it must be GPT for vision. 693 00:50:14,880 --> 00:50:19,160 And they've got a person communicating with the robot, asking it questions and asking 694 00:50:19,160 --> 00:50:20,800 it about what it sees. 695 00:50:20,800 --> 00:50:24,360 And the robot responds and then executes some tasks. 696 00:50:24,360 --> 00:50:33,800 And it's just a more fully rounded version of what a robot, human robot interaction might 697 00:50:33,800 --> 00:50:34,800 look like. 698 00:50:34,800 --> 00:50:36,920 However, there were some issues with it, weren't there? 699 00:50:36,920 --> 00:50:37,920 Yeah. 700 00:50:37,920 --> 00:50:40,040 I mean, the lag is noticeable. 701 00:50:40,040 --> 00:50:46,120 So if you're a listener, find this video on YouTube, it's probably very easy to be found 702 00:50:46,120 --> 00:50:48,560 because I think it's been well shared as well on all the socials. 703 00:50:48,560 --> 00:50:55,000 But in essence, person walks up and asks the, it says that he's hungry and the robot hands 704 00:50:55,000 --> 00:50:58,000 him an apple and all of the movements are very well handled. 705 00:50:58,000 --> 00:51:01,080 The robot also does some cleaning up and some other things. 706 00:51:01,080 --> 00:51:05,600 But there is a noticeable lag between asking the question and then the robot responding 707 00:51:05,600 --> 00:51:10,240 that you can imagine if that was a real life interaction would soon drive you absolutely 708 00:51:10,240 --> 00:51:11,240 crazy. 709 00:51:11,240 --> 00:51:15,960 It's like having a human who's really not that bothered about what you've got to say, 710 00:51:15,960 --> 00:51:20,760 not really paying attention to you and only gets back to you after thinking for 15 to 711 00:51:20,760 --> 00:51:23,200 20 seconds, which is annoying. 712 00:51:23,200 --> 00:51:26,080 Obviously, the latency of these things will drop down. 713 00:51:26,080 --> 00:51:30,720 And when I saw this, Martin, I couldn't help but think, I think when you're on an exponential 714 00:51:30,720 --> 00:51:36,280 curve, we're like, it's easy to say, oh, what's the day AGI will come out? 715 00:51:36,280 --> 00:51:37,840 Like, it just won't be like that. 716 00:51:37,840 --> 00:51:42,680 Like one day we'll go, oh, crumbs, has anybody taken a step back from this and gone two years 717 00:51:42,680 --> 00:51:44,720 ago we didn't have anything like this. 718 00:51:44,720 --> 00:51:48,080 But I do think this will be another one of those moments that froze of us that end up 719 00:51:48,080 --> 00:51:53,080 old enough to remember will be like, you know, all those robots that are working in the factories 720 00:51:53,080 --> 00:51:54,760 or helping cleaning up in the house. 721 00:51:54,760 --> 00:52:00,520 I remember when I saw the video of the first realistic conversation between a human and 722 00:52:00,520 --> 00:52:03,600 a robot, which worked quite well, but had some problems. 723 00:52:03,600 --> 00:52:06,200 And I think that's probably where this will end up. 724 00:52:06,200 --> 00:52:09,640 It will be one of those first steps along the road. 725 00:52:09,640 --> 00:52:16,320 So clearly has some problems, but I think it's, it's a measure of really rapid progress 726 00:52:16,320 --> 00:52:22,080 that happens when large language models are combined with significant improvements in 727 00:52:22,080 --> 00:52:23,080 robotics. 728 00:52:23,080 --> 00:52:24,080 Yeah. 729 00:52:24,080 --> 00:52:27,520 I'm looking forward to seeing where we head with this tech because I like the idea of 730 00:52:27,520 --> 00:52:28,520 a robot cleaner. 731 00:52:28,520 --> 00:52:29,520 Absolutely. 732 00:52:29,520 --> 00:52:33,400 I think my office could definitely do with a robot cleaner, but right. 733 00:52:33,400 --> 00:52:34,720 I think we'll call it there. 734 00:52:34,720 --> 00:52:37,680 Martin, thank you so much for your time as always. 735 00:52:37,680 --> 00:52:39,080 And thank you to you, dear listener. 736 00:52:39,080 --> 00:52:41,320 If you find this valuable, you know what I'm going to say. 737 00:52:41,320 --> 00:52:45,920 I say it every week, subscribe, share it with your pals in the marketing land. 738 00:52:45,920 --> 00:52:48,200 If you think that they will find value. 739 00:52:48,200 --> 00:52:49,800 We'll look forward to speaking with you next week. 740 00:52:49,800 --> 00:52:52,520 Martin, have a fab weekend and I will speak to you soon. 741 00:52:52,520 --> 00:52:55,080 See you later. 742 00:52:55,080 --> 00:52:58,520 Thank you for listening to artificially intelligent marketing. 743 00:52:58,520 --> 00:53:04,560 To stay on top of the latest trends, tips and tools in the world of marketing AI, be 744 00:53:04,560 --> 00:53:06,320 sure to subscribe. 745 00:53:06,320 --> 00:53:33,960 We look forward to seeing you again next week.