1 00:00:00,000 --> 00:00:21,560 Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the 2 00:00:21,560 --> 00:00:27,280 latest trends, tips, and tools in the world of marketing AI, helping you get the best 3 00:00:27,280 --> 00:00:33,600 results from your marketing efforts. Now let's join our hosts, Paul Avery and Martin Broadhurst. 4 00:00:35,680 --> 00:00:39,520 Welcome to the Artificially Intelligent Marketing podcast. 5 00:00:41,040 --> 00:00:49,920 I'm Martin Broadhurst and today I am flying solo. That's right, my regular co-host Paul Avery is not 6 00:00:49,920 --> 00:00:56,960 joining us in this episode. Instead, he had to take a bit of a breather after many weeks 7 00:00:56,960 --> 00:01:03,440 of intensive AI announcements, he just couldn't hack it, and quite frankly he had to just take 8 00:01:03,440 --> 00:01:10,640 a bit of a lie down in the Caribbean sun. So Paul, I hope you are feeling refreshed and ready to go 9 00:01:11,360 --> 00:01:19,760 next week, because if this week is anything to go by, the relentless deluge of AI news and 10 00:01:19,760 --> 00:01:27,680 announcements is not going away anytime soon. In this week's episode we are going to be talking 11 00:01:27,680 --> 00:01:36,800 about four big stories. First thing we're going to look at is Italy banning chat GPT. We'll also be 12 00:01:36,800 --> 00:01:43,360 looking at Metta's announcement of the Segment Anything foundational model for computer vision. 13 00:01:43,360 --> 00:01:50,320 We'll then be looking at the comments made by Sundar Pichai this week as he dismissed the AI 14 00:01:50,320 --> 00:02:00,480 chat bot threat posed by the likes of chat GPT and Bing AI. And then we'll be looking at the AI 15 00:02:00,480 --> 00:02:07,520 index report measuring trends in artificial intelligence before wrapping up this week's 16 00:02:07,520 --> 00:02:14,080 episode with a fascinating long form interview. Let's kick off with the first story of this week 17 00:02:14,080 --> 00:02:24,560 then. Italy has banned chat GPT. Last week's episode of the podcast was called chat GPT under fire. 18 00:02:25,120 --> 00:02:30,800 Little did we know when we recorded that, that little more than 72 hours after we published 19 00:02:30,800 --> 00:02:40,400 the episode a country would ban chat GPT outright. So Italy's data protection authority has criticised 20 00:02:40,400 --> 00:02:47,520 OpenAI for not checking user ages and said that there's been massive personal data collection 21 00:02:48,080 --> 00:02:57,520 during the training of the model. So OpenAI immediately blocked access to Italian users 22 00:02:57,520 --> 00:03:03,840 in response. Now this has obviously upset quite a few Italian users. It didn't take long to find 23 00:03:03,840 --> 00:03:13,680 people on Twitter saying what on earth is going on. So OpenAI has actually this week sat down and had 24 00:03:13,680 --> 00:03:22,560 talks with the Italian data protection authorities and they're presenting remedial measures that 25 00:03:22,560 --> 00:03:29,280 hope to mitigate and allay the concerns of the authorities. The company has sent over a document 26 00:03:29,280 --> 00:03:37,200 outlining measures to address these issues and I think they are hoping that they can get up and 27 00:03:37,200 --> 00:03:44,080 running very soon. Now why does this matter? Well Italy isn't the only country with concerns 28 00:03:44,080 --> 00:03:53,760 regarding chat GPT. In fact Germany this week has signalled that it may also look to block access 29 00:03:53,760 --> 00:04:00,800 to chat GPT. Now what I personally find slightly confusing about this story is as far as I can tell 30 00:04:00,800 --> 00:04:10,240 chat GPT is the only model that has been restricted and Bing AI has not seen any limitations put on it 31 00:04:10,240 --> 00:04:17,200 and as far as I can tell models such as Google's Bard whilst I realise it's only a small audience 32 00:04:17,200 --> 00:04:26,480 at the moment that has not had any restrictions either. So it seems like the authorities in the 33 00:04:26,480 --> 00:04:34,560 EU are really pushing to get the new technology providers in this space to basically increase 34 00:04:34,560 --> 00:04:43,440 transparency. A big concern is obviously around the way that they use people's data, they scrape the 35 00:04:43,440 --> 00:04:51,760 internet and all of the data that they're able to find goes into the training set, yet nobody knows 36 00:04:51,760 --> 00:05:00,400 what that data contains. We do know that if you search for people's names and search for companies 37 00:05:00,400 --> 00:05:06,160 it's able to access and present relevant information about people. You don't need to be 38 00:05:06,640 --> 00:05:14,480 a high profile figure to come up in the outputs of chat GPT. I've tried this with relatively 39 00:05:15,040 --> 00:05:21,680 small businesses and the CEOs of SMEs that I'm aware of and it can tell you who somebody is. 40 00:05:21,680 --> 00:05:27,360 So we know that is using this kind of data and the lack of transparency around the large language 41 00:05:27,360 --> 00:05:33,040 models training data, particularly GPT-4 which OpenAI has been very secretive with, 42 00:05:33,760 --> 00:05:39,040 is a cause for concern for many. So I think this really is just the start of what we can expect 43 00:05:39,040 --> 00:05:46,400 to see from European privacy regulators and no doubt before long they will start to coordinate 44 00:05:46,400 --> 00:05:53,360 their actions across the EU and we can expect to see companies having to make significant 45 00:05:53,360 --> 00:05:59,760 adjustments in the way that they open their books so to speak and let you interrogate their models. 46 00:06:00,800 --> 00:06:06,000 We can expect to see serious changes in this regard in the near future. 47 00:06:07,280 --> 00:06:13,360 What does this mean for marketers? Well watch this space really, the legislation and regulation 48 00:06:13,360 --> 00:06:20,080 around AI is nascent as we discussed in great detail last week. I think we can expect to see 49 00:06:20,080 --> 00:06:30,080 the EU taking a leading role in this. Their policy papers at the moment suggest that companies using 50 00:06:30,080 --> 00:06:35,840 and developing AI models are going to have to be more transparent and more open with the training 51 00:06:35,840 --> 00:06:40,400 data and the weights and biases that they use within their models, all of which I think will 52 00:06:40,400 --> 00:06:46,400 be a good thing in the long term. But when you have a transformative technology like chat GPT that 53 00:06:46,400 --> 00:06:53,040 really does capture the imagination of huge amounts of people all around the world, any restriction 54 00:06:53,040 --> 00:07:01,120 on that can just seem draconian but I'm sure it's all done for the best, with the best of intentions 55 00:07:01,120 --> 00:07:12,240 for the citizens of the EU. While we're talking about tech giants and their AI models that might 56 00:07:12,240 --> 00:07:19,600 come in for some scrutiny, we might as well talk about Meta. This week Facebook's parent company 57 00:07:19,600 --> 00:07:25,760 Meta announced Segment Anything, a new foundational model for computer vision. 58 00:07:26,400 --> 00:07:35,840 So the Segment Anything model, also known as SAM, is a one billion mask dataset. Now the aim of this 59 00:07:35,840 --> 00:07:43,680 as they say in their release paper is to democratise image segmentation. Now if you don't know a great 60 00:07:43,680 --> 00:07:50,400 deal about computer vision, it's not an area that you're familiar with, it's basically the technology 61 00:07:50,400 --> 00:07:58,240 or the machine learning models that allow computers to see and interpret what they are seeing. It 62 00:07:58,240 --> 00:08:04,960 plays a role in things like self-driving cars, they'll have cameras that will detect objects and 63 00:08:04,960 --> 00:08:10,560 classify the objects that they're looking at, things like that is a car, that is a lamppost, 64 00:08:10,560 --> 00:08:18,560 that is a pedestrian, or they can be used in, recently saw a great example where AI computer 65 00:08:18,560 --> 00:08:28,560 vision was being used to identify bugs in an insect farm, and it was looking at the the creatures on 66 00:08:28,560 --> 00:08:35,280 the insect farm and able to identify which of those were injured or maybe had illnesses, a really 67 00:08:35,280 --> 00:08:42,800 interesting innovative area. In more everyday usage you might come across computer vision when 68 00:08:42,800 --> 00:08:48,880 you're unlocking your iPhone, the face unlock feature is using computer vision technology behind 69 00:08:48,880 --> 00:08:57,760 that, but the new Segment Anything foundational model from Meta is designed to enable developers 70 00:08:57,760 --> 00:09:06,640 to incorporate this new technology into their apps at scale, and it's what really sets it apart from 71 00:09:07,840 --> 00:09:15,600 existing datasets is just the sheer volume of the segmentation masks. You can think of a segmentation 72 00:09:15,600 --> 00:09:20,960 mask as being like on a colouring book page where you might just colour in only the shape of the 73 00:09:20,960 --> 00:09:25,360 object that you're interested in, so if there's a cat on that colouring page you would colour in 74 00:09:25,360 --> 00:09:32,000 the cat and leave everything else on the page white, and doing that, feeding that into the computer 75 00:09:32,000 --> 00:09:36,480 will help the computer understand where the object is in the picture and then separate it out from 76 00:09:36,480 --> 00:09:42,080 the background, so the more examples of masks that you've given to the model the better it is at 77 00:09:42,080 --> 00:09:48,640 identifying objects within a scene and being able to separate them out from one another. 78 00:09:48,640 --> 00:09:56,320 In the example that they give in the research announcement they actually say that it has 400 79 00:09:56,320 --> 00:10:06,240 times segmentation masks compared to the next best model, so it's a huge increase in the data that 80 00:10:06,240 --> 00:10:11,760 they've got to work with. Now they give some examples of what you can actually do with all of 81 00:10:11,760 --> 00:10:18,000 this, in the section on the announcement page they say what lies ahead, and they've got some nice 82 00:10:18,000 --> 00:10:24,320 little illustrations where for instance you could be using AR glasses, so you're using augmented 83 00:10:24,320 --> 00:10:31,280 reality glasses and they could prompt the users with reminders and instructions based on what the 84 00:10:31,280 --> 00:10:36,720 person is looking at, and the example that they actually give is somebody working in a kitchen 85 00:10:36,720 --> 00:10:45,360 and in front of them in the kitchen is a chopping board with a cake on the side which has not been 86 00:10:45,360 --> 00:10:55,760 iced, and in the top corner on the augmented reality glasses the user sees videos suggesting 87 00:10:57,040 --> 00:11:04,160 cake decorating tips, because it recognizes that there is an undecorated cake and kind of pulls out 88 00:11:04,160 --> 00:11:10,560 some relevant advice. There's another example where there is an empty dog bowl, a food bowl, 89 00:11:10,560 --> 00:11:17,280 on the floor and the glasses see this bowl and it's empty and it pops up with a notification saying 90 00:11:17,280 --> 00:11:24,640 that Rex was fed 20 minutes ago, so it's context aware. There are industrial applications for this, 91 00:11:24,640 --> 00:11:29,040 many of these that you could think of, but in the examples that they give they talk about using it 92 00:11:29,040 --> 00:11:37,680 for agriculture to be able to assist farmers kind of managing their herds, or you could use it 93 00:11:37,680 --> 00:11:42,320 if you're a biologist, and this is something I'd love to discuss in the future with Paul, 94 00:11:42,720 --> 00:11:51,280 if you were looking at cell division or any kind of life sciences application that really, you know, 95 00:11:51,280 --> 00:11:57,040 you need to be able to count cells rapidly, I don't know why I'm talking about this, 96 00:11:57,040 --> 00:12:02,640 is if I have any idea what I'm on about, this is so much in Paul's domain that I should just 97 00:12:02,640 --> 00:12:08,560 reference the fact that I'm looking at a GIF with some cells moving on it right now, and apparently 98 00:12:08,560 --> 00:12:14,240 it's going to be really useful for that. We'll talk more about that in the future with Paul when 99 00:12:14,240 --> 00:12:21,120 he's back in the studio. Suffice to say this is a really interesting new technology that's going to 100 00:12:21,120 --> 00:12:27,360 open up huge areas of innovation and allow people to do so much more with their existing products. 101 00:12:27,360 --> 00:12:32,240 There's going to be applications in the security industry, you can well imagine how this is going 102 00:12:32,240 --> 00:12:38,400 to be integrated into things like CCTV, how it's going to help things like crowd control, 103 00:12:39,280 --> 00:12:44,880 TV and movie industry is obviously going to be able to make great use of this. Self-driving cars 104 00:12:44,880 --> 00:12:51,120 which do need to be able to see objects, track objects and classify objects are about to get 105 00:12:51,120 --> 00:12:58,240 quite an upgrade as well, so the impact of this is not to be understated and I think it's worth 106 00:12:58,240 --> 00:13:05,760 saying at this point that Meta deserve some credit because this model has actually been released 107 00:13:06,400 --> 00:13:13,360 with the Apache 2 license, which is a permissive open license, so this means that anyone can use 108 00:13:13,360 --> 00:13:18,800 the model and developers can build with it now. The Apache 2 license allows for free use, 109 00:13:18,800 --> 00:13:24,800 modification and distribution of the software so that makes it accessible for a wide range of 110 00:13:24,800 --> 00:13:31,920 applications and research purposes, so hats off to Meta for that. In fact the top comment 111 00:13:31,920 --> 00:13:36,800 that I saw when I read the tweet announcing this was from somebody saying, 112 00:13:36,800 --> 00:13:42,480 that's all great but you should have led with the fact it's an Apache 2 license, wow. 113 00:13:42,480 --> 00:13:48,240 Next up I want to look at a recent interview that Sundar Pichai gave with the Wall Street Journal. 114 00:13:48,720 --> 00:13:56,880 The CEO of Alphabet has come out on a bit of a PR front after Google has faced significant 115 00:13:56,880 --> 00:14:06,720 pushback or criticism for its lack of innovation and basically it's well to address the threat 116 00:14:06,720 --> 00:14:13,920 that it may or may not be facing from the likes of ChatGPT and Bing AI. I mentioned there it's faced 117 00:14:13,920 --> 00:14:20,160 some criticism and that people are saying that it is under threat but that's not how Sundar Pichai 118 00:14:20,160 --> 00:14:26,640 sees it, if anything he actually dismisses and minimises the chatbot threat, he says that they 119 00:14:26,640 --> 00:14:32,640 see opportunity in the space, he says the opportunity is if anything bigger than before. 120 00:14:32,640 --> 00:14:40,080 Now he does say that Google will be enabling users to engage with large language models in a search 121 00:14:40,080 --> 00:14:46,880 context and that they are currently testing new search products, specifically ones that allow 122 00:14:46,880 --> 00:14:53,120 people to ask follow-up questions, so maybe we're going to see a more chat-like interface when it 123 00:14:53,120 --> 00:15:00,800 comes to Google search results. When asked why Google has been able to do so, I think it's because 124 00:15:00,800 --> 00:15:11,360 of why Google hadn't responded or released a chatbot earlier, why it was maybe resting on its laurels, 125 00:15:11,360 --> 00:15:16,400 he said that they were iterating and preparing to ship something and that maybe the timelines 126 00:15:16,400 --> 00:15:24,960 had changed given the moment in the industry that we saw back on November 30th last year when ChatGPT 127 00:15:24,960 --> 00:15:32,160 came to the fore. He does say that there has been incredible user excitement and goes on to say that 128 00:15:32,160 --> 00:15:38,560 this is interesting for the company, he says it has been incredible to see user excitement around 129 00:15:38,560 --> 00:15:45,040 adoption of these technologies and that is a pleasant surprise as well. Now it remains to be 130 00:15:45,040 --> 00:15:52,400 seen exactly how Google does deploy AI in this space, it's a topic of conversation that we've 131 00:15:52,400 --> 00:16:00,480 had in recent weeks talking about what does this mean for search engine optimization, that was an 132 00:16:00,480 --> 00:16:07,120 interesting discussion that I caught up with on Twitter this week as well where Chris Penn from 133 00:16:07,120 --> 00:16:16,480 Trust Insights was saying that the era of unbranded search content is coming to an end, 134 00:16:16,480 --> 00:16:23,440 branded search will be the primary driver of traffic and in a few years time when people are 135 00:16:23,440 --> 00:16:28,720 searching for things that they might ordinarily find on your blog, if you're doing long-form 136 00:16:28,720 --> 00:16:33,760 content marketing and inbound marketing type activity, soon those kind of answers will be 137 00:16:33,760 --> 00:16:42,320 found on the likes of ChatGPT and via chat AIs, so trying to optimize for that will be pointless. 138 00:16:42,320 --> 00:16:49,520 Now whether that comes to pass remains to be seen, I should say apologies to Chris if I've 139 00:16:49,520 --> 00:16:55,600 completely misrepresented that, I feel like I've paraphrased slightly just to get that across 140 00:16:55,600 --> 00:17:03,920 quickly, but yeah going back to Google for just a second, PitchEye says that there are plans in 141 00:17:03,920 --> 00:17:11,360 place for Google Brain, the AI research lab within Google, and DeepMind, the AI research lab 142 00:17:11,360 --> 00:17:21,120 and the alphabet owned research company headed up by Demis Hasabis, there are plans for these two 143 00:17:21,120 --> 00:17:26,720 organizations to collaborate more closely. Now this makes sense, PitchEye does say he expects 144 00:17:26,720 --> 00:17:32,320 a lot more stronger collaboration because some of these efforts will be more compute intensive, 145 00:17:32,320 --> 00:17:37,680 so it makes sense to do it at a certain scale. There has been a little bit of friction between 146 00:17:37,680 --> 00:17:47,200 DeepMind and Google in the past with DeepMind not being entirely aligned with the Google mission, 147 00:17:47,200 --> 00:17:53,200 there's said to be a bit of distrust and a fear of misuse of DeepMind's technology by Google 148 00:17:53,920 --> 00:18:00,080 and or the military. Some sources also mention that there's something of a cultural conflict 149 00:18:00,080 --> 00:18:06,400 between DeepMind's academic researchers and Google's business oriented engineers, so much so 150 00:18:06,400 --> 00:18:12,320 in fact that DeepMind tried to break away from Google and become a separate division under 151 00:18:12,320 --> 00:18:20,160 alphabet, but negotiations on that have failed apparently. So it will be interesting to see how 152 00:18:20,160 --> 00:18:27,520 they bring DeepMind and Google Brain together, certainly feel that DeepMind have, you know, 153 00:18:27,520 --> 00:18:33,840 if I go back looking over the past 10 years or so, DeepMind have always been one of these companies 154 00:18:33,840 --> 00:18:39,840 that quite frankly are pushing the boundaries, they're the ones that are doing great things 155 00:18:39,840 --> 00:18:46,320 in the life sciences space, they've done some amazing work with the likes of AlphaGo, you know, 156 00:18:46,320 --> 00:18:53,120 that was a great DeepMind project. So when we start to see some of their capabilities 157 00:18:54,000 --> 00:18:59,920 incorporated into the Google product suite, that does make you think that there's some exciting 158 00:18:59,920 --> 00:19:05,440 times ahead for Google and you know, maybe they, maybe they should have been looking to make this 159 00:19:05,440 --> 00:19:12,640 move some time before now, but there's nothing like a bit of healthy competition to spur things 160 00:19:12,640 --> 00:19:21,280 along. Our final story of the week comes from the Stanford Institute for Human-Censored Artificial 161 00:19:21,280 --> 00:19:30,240 Intelligence, which has released its 2023 AI Index report. The AI Index collaborates with many 162 00:19:30,240 --> 00:19:35,520 different organizations to track progress in artificial intelligence. These organizations 163 00:19:35,520 --> 00:19:41,360 include the Center for Security, Emerging Technology, Georgetown University, LinkedIn, 164 00:19:41,360 --> 00:19:48,880 NetBase, Quidd, Lightcast and McKinsey. The 2023 report also features more self-collected data 165 00:19:48,880 --> 00:19:58,320 and original analysis than ever, they tell us. So this is a big look at the broadening 166 00:19:59,520 --> 00:20:09,600 space that is AI, and it's tracking everything from investment rates to risks to policy and 167 00:20:09,600 --> 00:20:15,120 regulation, you name it. So I just want to pull out some of the top takeaways from the report to 168 00:20:15,120 --> 00:20:21,920 keep you up to speed with where we're at globally when it comes to AI. So a few key highlights 169 00:20:21,920 --> 00:20:28,720 taken from the report. First of all, industry is racing ahead of academia, and this came as no 170 00:20:28,720 --> 00:20:33,280 surprise to me, and I think anyone that's been paying attention in this space for a few years 171 00:20:33,280 --> 00:20:40,480 now will have seen this, just all of the big announcements these days are coming out of Meta, 172 00:20:40,480 --> 00:20:49,360 Nvidia, OpenAI, Microsoft, etc. So the report says until 2014, most significant machine learning 173 00:20:49,360 --> 00:20:55,920 models were released by academia, since then industry has taken over. In 2022, there were 32 174 00:20:55,920 --> 00:21:03,040 significant industry-produced machine learning models compared to just three produced by academia. 175 00:21:03,040 --> 00:21:08,080 It goes on to state that basically because of the huge amounts of data, the compute power, 176 00:21:08,080 --> 00:21:14,000 and quite frankly, the deep pockets that are required, it makes sense that industrial actors 177 00:21:14,000 --> 00:21:25,440 are going to need to really take up the reins here and drive things forward. Also, when you consider 178 00:21:25,440 --> 00:21:36,560 the massive upside from the profit motive, if you can realize the benefits from AI, it's not a great 179 00:21:36,560 --> 00:21:42,480 surprise to see that the incentive does exist for them to make these investments, because down the 180 00:21:42,480 --> 00:21:49,760 road the returns on investment look to be pretty chunky. In what is something of a theme at the 181 00:21:49,760 --> 00:21:53,840 moment in some of the reports that are coming out, particularly from academia regarding AI, 182 00:21:55,120 --> 00:22:03,120 there is a suggestion that AI is harming the environment, though to be fair this article 183 00:22:03,120 --> 00:22:08,960 does say that AI is both helping and harming the environment. So you know, not all bad. 184 00:22:09,760 --> 00:22:15,360 New research suggests that AI systems can have serious environmental impacts according to one 185 00:22:15,360 --> 00:22:27,280 report. Bloom's training run emitted 25 times more carbon than a single air traveller on a one-way 186 00:22:27,280 --> 00:22:35,920 trip from New York to San Francisco. Which, if I'm honest, doesn't feel like a great deal to me, 187 00:22:35,920 --> 00:22:41,920 I actually think that sounds quite small given the potential benefits of a big large language model 188 00:22:43,440 --> 00:22:51,120 like Bloom. So yeah, I can understand why it needs to be highlighted, but it could be worse. 189 00:22:51,120 --> 00:22:56,880 It does go on to say that new reinforcement learning models like B. Cooler show that AI 190 00:22:56,880 --> 00:23:06,160 systems can be used to optimise energy usage. So yeah, it's not all bad. When it comes to 191 00:23:07,520 --> 00:23:14,000 AI and industry and jobs, a topic that we discussed last week following OpenAI's report that showed 192 00:23:14,000 --> 00:23:21,680 80% of jobs will be impacted by GPT technologies by at least 10%, the report says that the demand 193 00:23:21,680 --> 00:23:27,920 for AI-related professional skills is increasing across virtually every American industrial sector. 194 00:23:27,920 --> 00:23:34,400 So no great surprise there. And it goes on to say across every sector in the United States for which 195 00:23:34,400 --> 00:23:41,440 there is data, with the exception of agriculture, forestry, fisheries and hunting, the number of AI 196 00:23:41,440 --> 00:23:52,880 related job posts has increased on average from 1.7% in 2021 to 1.9% in 2022. So employers in the 197 00:23:52,880 --> 00:23:58,960 United States are increasingly looking for workers with AI-related skills, unless of course you're in 198 00:23:58,960 --> 00:24:04,880 the world of hunting. And to be honest, I'd prefer that we just keep artificial intelligence 199 00:24:04,880 --> 00:24:11,040 out of hunting. That just sounds like a bad, I've seen Terminator, that just doesn't seem like a 200 00:24:11,040 --> 00:24:18,960 wise move at this stage. Or if we are going to do it, just keep it a GPT-3. The report also goes on 201 00:24:18,960 --> 00:24:24,800 to say that policymakers' interest in AI is on the rise, and this was a big topic of discussion 202 00:24:24,800 --> 00:24:31,920 on last week's show after the UK government announced their white paper looking at AI policy 203 00:24:31,920 --> 00:24:39,200 and regulation. So the report says that an AI index analysis of the legislative records of 127 204 00:24:39,200 --> 00:24:45,120 countries shows that the number of bills containing artificial intelligence that were passed into law 205 00:24:45,120 --> 00:24:54,320 grew from just 1 in 2016 to 37 in 2022. An analysis of the parliamentary records on AI 206 00:24:54,320 --> 00:25:01,280 in 81 countries likewise shows that mentions of AI in global legislative proceedings have increased 207 00:25:01,280 --> 00:25:08,240 nearly 6.5 times since 2016. So this is no great surprise at all, and you know that was the lead 208 00:25:08,240 --> 00:25:17,680 story on today's pod, Italy banning GPT, and we're going to see more of this in the future no doubt. 209 00:25:17,680 --> 00:25:26,320 As the EU brings its own AI legislation to parliament and puts it on the books, and America 210 00:25:26,320 --> 00:25:34,000 does the same and China is putting it through the law books as we speak, we can just expect more and 211 00:25:34,000 --> 00:25:42,400 more of this interest in regulating AI to come to the fore. So watch this space, it will be interesting 212 00:25:42,400 --> 00:25:49,520 to see where we land on that one. And one final takeaway, there is more on this by the way, so 213 00:25:51,040 --> 00:25:55,760 if you want to read the full report it's about 380 pages long and I'm not going to pretend I've read 214 00:25:55,760 --> 00:26:02,800 all of it in the few days since it came out, but there are a few more highlights, but the last one 215 00:26:02,800 --> 00:26:09,920 I want to talk about is something about the sentiment that we feel towards artificial 216 00:26:09,920 --> 00:26:16,480 intelligence. How do the general public feel about AI? Well according to the report Chinese 217 00:26:16,480 --> 00:26:21,440 citizens are among those who feel the most positively about AI products and services, 218 00:26:22,000 --> 00:26:32,240 whereas Americans not so much. In a 2022 Ipsos survey 78% of Chinese respondents, the highest 219 00:26:32,240 --> 00:26:40,800 proportion surveyed, agreed with the statement that products and services using AI have more 220 00:26:40,800 --> 00:26:49,840 benefits than drawbacks. After Chinese respondents Saudi Arabia 76% and India 71% felt the most 221 00:26:49,840 --> 00:26:58,160 positive about AI, so you know you're really 78% of Chinese, 76% Saudi, 71% India, that's a lot of 222 00:26:58,160 --> 00:27:04,640 positivity floating around there about what AI can do for people, whereas Americans are a little bit 223 00:27:04,640 --> 00:27:13,120 more skeptical, only 35% of those sampled agreed that products or services using AI had more 224 00:27:13,120 --> 00:27:21,680 benefits than drawbacks. What that means for adoption of new technologies or how companies 225 00:27:21,680 --> 00:27:29,120 need to think about rolling out AI infused products and services remains to be seen, but 226 00:27:29,120 --> 00:27:37,120 it does suggest that there's a degree of scepticism around the power and potential of AI in the US and 227 00:27:37,120 --> 00:27:45,040 maybe the wider Western societies as well, that means that we're going to have to act cautiously 228 00:27:45,040 --> 00:27:51,760 and not be so techno-optimistic, and I definitely fall into the side of being a techno-optimist, 229 00:27:51,760 --> 00:27:58,960 just something to be aware of if you're evangelizing the benefits of AI, maybe you'll 230 00:27:58,960 --> 00:28:05,760 get a more positive response or a positive reception if you go and do that in China. 231 00:28:07,280 --> 00:28:14,560 Well that's it for the main stories this week, note all of the week this week because as you're 232 00:28:14,560 --> 00:28:22,480 going to hear in the interview that Paul conducted, you're going to hear quite a lot about the power 233 00:28:22,480 --> 00:28:30,720 of GPT-4, so without further ado I'm going to hand off to Paul who conducted this interview looking 234 00:28:30,720 --> 00:28:39,280 at how chat GPT-4 can be used to boost the productivity of an e-commerce company, over to you 235 00:28:39,280 --> 00:28:46,960 Paul. So a bit of a special edition artificially intelligent marketing podcast today, because we 236 00:28:46,960 --> 00:28:54,000 are lucky enough to be able to speak with a good friend of mine, Emil Lamprecht. Emil is CEO of 237 00:28:54,000 --> 00:28:59,920 Growth Mechanics, he is a co-founder of multiple businesses within that portfolio, he's a startup 238 00:28:59,920 --> 00:29:07,840 expert having built a number of startup incubators over the years and mentored over 800 startups, 239 00:29:07,840 --> 00:29:13,280 he's a marketing expert himself having previously been CMO at Athletic Greens which is now a hundred 240 00:29:13,280 --> 00:29:18,080 million dollar company as well as a number of other marketing gigs and bits and pieces he's done 241 00:29:18,080 --> 00:29:23,120 over the years and he's just an overall business expert and good guy. I guess the most important 242 00:29:23,120 --> 00:29:29,120 thing is he is a fellow AI nerd, so today we're going to talk about some cool stuff. Emil thank 243 00:29:29,120 --> 00:29:34,640 you so much for joining me, I appreciate it. Lovely to be here Paul, thanks for inviting me and it's 244 00:29:34,640 --> 00:29:38,640 going to be very hard not to call you by your shorthand, I just realized I almost called you 245 00:29:38,640 --> 00:29:46,080 AVO right off the bat there. I'll have to get used to calling you Paul on this conversation. 246 00:29:46,080 --> 00:29:51,920 You call me whatever you like as long as it's nice, okay. That's probably going to be the hardest bit. 247 00:29:51,920 --> 00:29:56,000 That's the hardest bit yeah. We're going to talk about two main things for the mill today, the 248 00:29:56,000 --> 00:30:01,760 first one we're going to talk about is a Ninja marketing application of AI that he's developed, 249 00:30:01,760 --> 00:30:05,440 which I think you're all going to find really interesting. So we're going to talk about that 250 00:30:05,440 --> 00:30:09,680 and hopefully it will inspire you as marketers to think about other ways you might use AI outside of 251 00:30:09,680 --> 00:30:14,160 the general, oh I'm going to put some stuff into chat GPT, which of course many of us have played 252 00:30:14,160 --> 00:30:18,960 with. The other thing we're going to talk about is we're going to tap into Emil's role as a business 253 00:30:18,960 --> 00:30:24,960 owner and founder and just get his thoughts on how is the emergence of AI really going to drive 254 00:30:24,960 --> 00:30:30,880 how businesses need to operate and market themselves over the next 18 months, 24 months and beyond 255 00:30:30,880 --> 00:30:35,280 because obviously things are moving really quickly. So Emil, let's start with this Ninja 256 00:30:35,280 --> 00:30:39,040 application that you were telling me about because it sounded really cool to take us through what you 257 00:30:39,040 --> 00:30:45,760 were doing and why and how you benefited. Yeah, so I mean, interestingly enough, or I think for me, 258 00:30:45,760 --> 00:30:50,560 what's most interesting about it is we did this before the chat GPT hype, obviously aspects of 259 00:30:50,560 --> 00:30:55,760 open AI and use of their APIs have been around for a while and one of the cool things is that 260 00:30:55,760 --> 00:31:02,160 there have been some tools that implemented some form of availability to that early on. So like 261 00:31:02,160 --> 00:31:08,480 we use a tool for a lot of things called coda.io, which if you're familiar with Airtable or Notion 262 00:31:08,480 --> 00:31:16,080 is comparable to sort of both of those things or a combination of those things. And part of our 263 00:31:16,080 --> 00:31:23,760 portfolio is e-commerce and in one case of business we've recently acquired is e-commerce with a very 264 00:31:23,760 --> 00:31:29,920 large portfolio. We're talking thousands of products across multiple languages, across four 265 00:31:29,920 --> 00:31:37,840 languages. So English, Spanish, Portuguese and French, which okay, sure, we're, you know, a small 266 00:31:37,840 --> 00:31:44,400 agile team and it's a nice idea to grow that to have 100 translators to constantly monitor and 267 00:31:44,400 --> 00:31:50,640 update every detail of product schema for every language all the time in the highest version of 268 00:31:50,640 --> 00:31:59,200 quality, but there are limits to that being pragmatic or realistic. And so last autumn, 269 00:31:59,760 --> 00:32:05,840 sort of August, September, we were looking at, okay, how could we, to a certain level of quality 270 00:32:05,840 --> 00:32:13,760 and efficiency leverage existing technology to try and manage and possibly even write or rewrite and 271 00:32:13,760 --> 00:32:22,160 republish product schema, whether that be titles, metadata, material information, warnings and 272 00:32:22,160 --> 00:32:30,320 recommendations, descriptions, et cetera, et cetera. And we said, well, oh, okay, we play with 273 00:32:30,320 --> 00:32:35,520 coda and there's currently a little bit of availability built by one of the coda team members 274 00:32:35,520 --> 00:32:46,240 of using the OpenAI API, which at that point was GPT 3.0. I think, I don't think it was at 3.5 275 00:32:46,240 --> 00:32:54,400 at that point. And let's see what we can do. And, you know, the advantage of being able to build 276 00:32:54,400 --> 00:33:02,800 and manage a database and an easy to use UI like these tools now allow you to, we sort of figured 277 00:33:02,800 --> 00:33:11,920 out a way to import, rewrite or translate as needed entire sets of product schema, concatenate 278 00:33:11,920 --> 00:33:17,680 it into, you know, end results, including recommendation links to associated product types 279 00:33:17,680 --> 00:33:23,840 and things like this, and then push that all the way back into the front end of the store and sort 280 00:33:23,840 --> 00:33:33,120 of dynamically manage our product portfolio and all of that product metadata off site, but to a 281 00:33:33,120 --> 00:33:39,680 certain extent automatically, which was quite a win. I mean, I calculated that at one point that 282 00:33:39,680 --> 00:33:46,000 by using this method of translation, we were probably losing, you know, maybe like a 20, 25% 283 00:33:46,000 --> 00:33:54,560 quality amounts, but we were gaining what would normally take in human hours, about 1200 hours 284 00:33:54,560 --> 00:33:59,920 to execute in any one go. And, you know, we're able, we still have to manage it in small, 285 00:33:59,920 --> 00:34:04,720 it's a lot of data, right? So you can't just do a portfolio of 10,000 products in one go. It takes a 286 00:34:04,720 --> 00:34:11,840 progressive amount of import, export, processing, et cetera. But, you know, let's say we do 287 00:34:11,840 --> 00:34:16,240 a hundred products a week or a thousand products a week over the course of a year, we're saving 288 00:34:16,240 --> 00:34:22,560 thousands of human working hours from that. So let me just play that back to make sure I've 289 00:34:22,560 --> 00:34:29,280 understood. You've got this e-commerce website where you're selling 5,000 plus products across 290 00:34:29,280 --> 00:34:35,680 a bunch of different languages. And then what you've done is you've used access to GPT's API 291 00:34:35,680 --> 00:34:41,520 to find a way of effectively automating updates to the product descriptions and stuff across 292 00:34:41,520 --> 00:34:48,320 all those products in all those languages with GPT taking care effectively of the writing of those 293 00:34:48,320 --> 00:34:52,640 and the translation of those so that humans didn't have to. Is that a fair summary of what you did? 294 00:34:52,640 --> 00:34:59,200 Yeah, yeah, exactly. And there's actually, you know, of course, since the sort of hype of AI began 295 00:34:59,200 --> 00:35:03,040 in November, December, there's of course an increased availability of similar things on 296 00:35:03,040 --> 00:35:07,520 most e-commerce things natively now, right? So there's WordPress plugins, there's Shopify apps, 297 00:35:07,520 --> 00:35:11,920 there's big commerce apps, there's, you know, even press to shop modules that exist for this 298 00:35:11,920 --> 00:35:19,360 type of stuff within certain limitations. I think even that the fact that those exist is, you know, 299 00:35:19,360 --> 00:35:25,120 a way to do some of the work either automated or, you know, push button generation. I think the 300 00:35:25,120 --> 00:35:31,680 difference between that and what we built was that we can manage entire sections of the portfolio 301 00:35:31,680 --> 00:35:38,320 also because we control the prompt structure. So I'm having, you know, we're writing the prompt 302 00:35:38,320 --> 00:35:44,960 into the kind of database command that processes each product so that for each language, for each 303 00:35:44,960 --> 00:35:50,880 product type is producing a prompt specific in style and cadence and interest to that particular 304 00:35:50,880 --> 00:35:55,120 product in that context, which is not something you can really do with just sort of the plug and 305 00:35:55,120 --> 00:36:01,600 play apps that you find for any e-commerce system at the moment. So I don't know if those apps will 306 00:36:01,600 --> 00:36:08,880 find a way to mature to that. To be honest, it seems unlikely because, you know, marketers always 307 00:36:08,880 --> 00:36:14,480 want the laziest version of things. So I'm not even sure it will occur to most marketers to try and 308 00:36:14,480 --> 00:36:20,640 be that specific with generated plugins on e-commerce. But for us, it was really important 309 00:36:20,640 --> 00:36:27,280 to kind of have some stylistic control. And, you know, even with the revolution of, you know, 310 00:36:27,280 --> 00:36:31,040 the thousands of AI tools that have appeared in the last two months, nothing comes close that I've 311 00:36:31,040 --> 00:36:36,000 seen so far to being able to produce the result we're doing with a very simple API call through 312 00:36:36,000 --> 00:36:41,600 a Coda doc. So how long did it take you? I realize there's an optimization part to this where you 313 00:36:41,600 --> 00:36:45,760 make sure it works as you want, but how long did it take you to basically build that? 314 00:36:48,800 --> 00:36:53,760 That's its own question. Like, is it a day? Is it a week? Is it six months of work to build something 315 00:36:53,760 --> 00:36:58,240 like this? What does it look like? You know, I think to set up the initial version, just to prove 316 00:36:58,240 --> 00:37:03,840 the concept, I think it probably took, you know, between me and another team member playing around 317 00:37:03,840 --> 00:37:11,200 with it, it probably took us two days a day to get like, oh, yeah, this could work to get to that 318 00:37:11,200 --> 00:37:19,200 point of, oh, yeah, this could work. Right. To get it really nailed down, it took a number of iterations 319 00:37:19,200 --> 00:37:23,920 and those iterations are still happening. I mean, today we realized, for instance, just as an example, 320 00:37:23,920 --> 00:37:28,320 a Portuguese friend of mine was going over some longer form stuff that we had produced through 321 00:37:28,320 --> 00:37:37,520 the system and informed me in a very kind, polite way. Anytime it writes something in Portuguese 322 00:37:37,520 --> 00:37:44,960 over a certain length, it assumes a Brazilian style of Portuguese. Interesting. Which I, not being a 323 00:37:44,960 --> 00:37:49,920 native Portuguese speaker is not something I picked up on when I was reading various Portuguese 324 00:37:49,920 --> 00:37:55,920 texts coming out of the system. So, you know, but because we built the system the way it is, 325 00:37:55,920 --> 00:38:00,960 I can go back to where we have isolated the language prompts and instead of saying in Portuguese, 326 00:38:00,960 --> 00:38:07,840 I'm going to edit that to say in Portuguese as you would find in Portugal, not in Brazil, and, 327 00:38:07,840 --> 00:38:13,840 you know, customize that prompt even further to that sort of specificity. So these sorts of 328 00:38:13,840 --> 00:38:20,240 micro iterations are happening all the time. You know, we've managed since sort of starting 329 00:38:20,240 --> 00:38:25,920 to translate the product portfolio, we've managed to expand that to, okay, we also have to translate 330 00:38:25,920 --> 00:38:31,600 quite a lot of product information into social posts in different languages for different social 331 00:38:31,600 --> 00:38:36,320 accounts that we run. And so we've now started building out those layers and then we're also 332 00:38:36,320 --> 00:38:39,760 then extending those integrations and connecting that to schedulers and saying, okay, well, can you 333 00:38:39,760 --> 00:38:44,080 go through a generative and human review process that then pushes all the way straight through to 334 00:38:44,080 --> 00:38:49,040 the schedule and then you're done. And that's all based on this one sort of centralized system we've 335 00:38:49,040 --> 00:38:54,560 built. So again, very automated GPT is writing social posts, promoting the products, you want 336 00:38:54,560 --> 00:38:57,840 a human reviewing them for quality, but then once they're approved, they get pushed through and 337 00:38:57,840 --> 00:39:02,160 automatically built into the scheduled stream of posts that are going to go out across the social 338 00:39:02,160 --> 00:39:06,720 profiles for the business. Is that what you're describing? Yeah, exactly. Yeah. And I, you know, 339 00:39:06,720 --> 00:39:11,280 particularly for the social stuff, you know, for the product stuff on the mass that we have, 340 00:39:12,000 --> 00:39:16,800 you know, we will mature to having a human layer on top of that generator form to sort of approve 341 00:39:16,800 --> 00:39:21,680 and review before it pushes all the way back through, because I just, it does make a huge 342 00:39:21,680 --> 00:39:28,960 difference. You know, at first glance, text looks the same, but once you get used to seeing AI text, 343 00:39:28,960 --> 00:39:34,640 even once you get really good at stylistic prompts, you can still always tell it's AI. And so 344 00:39:34,640 --> 00:39:38,800 we'll search engines and everything else, right? So it's important to have that human layer. 345 00:39:39,520 --> 00:39:44,000 For now, we just have that human layer on the social side, but you know, we've gotten very good 346 00:39:44,000 --> 00:39:48,800 at how the sort of generative human review publication cycle should work just in the past 347 00:39:48,800 --> 00:39:55,120 couple of weeks. And so, you know, what would take normally would have taken someone probably 348 00:39:55,120 --> 00:40:00,800 a couple hours to write, you know, multilingual versions of social posts about a series of 349 00:40:00,800 --> 00:40:06,160 products, takes a few minutes, then some editing and quick review, and then pushes straight all the 350 00:40:06,160 --> 00:40:12,080 way through. Awesome. It also comes to mind that, because you've got control over the prompts, 351 00:40:12,080 --> 00:40:16,720 which some of these sort of third party plugins might not have, I guess you could end up doing 352 00:40:16,720 --> 00:40:22,400 some quite interesting stuff running things like A-B tests to see if having a certain type of style 353 00:40:22,400 --> 00:40:27,200 of description works better than another, but you can also run those A-B tests at quite, 354 00:40:27,200 --> 00:40:31,120 quite an interesting scale across different product groups, because you don't have to 355 00:40:31,120 --> 00:40:34,000 manually do all the writing of the different descriptions. 356 00:40:34,480 --> 00:40:39,600 Yeah, exactly. I mean, we're definitely not at the point to do that scientifically enough yet, 357 00:40:39,600 --> 00:40:46,320 but it's, if you look at the admin doc of things to do, it's there. It's definitely on the list. 358 00:40:46,320 --> 00:40:50,720 We'd love to get to the point where we could do that. And part of that is a performance question, 359 00:40:50,720 --> 00:40:55,840 right? Like, especially when you're processing this many requests, it's a lot of work. 360 00:40:55,840 --> 00:41:03,200 And so, we're using the API one call at a time, right? So we dump in 100 products for each product, 361 00:41:03,200 --> 00:41:08,560 it's making 28 separate requests and then collating those requests into a final 362 00:41:08,560 --> 00:41:13,760 series of outcomes, right? So it's a lot of, a lot happens for every product still, 363 00:41:14,640 --> 00:41:18,800 which takes a lot of time to process. And then the database has to actually load that visually 364 00:41:18,800 --> 00:41:23,760 so that we can interact with it for the human review and so forth. So there's some parts of 365 00:41:23,760 --> 00:41:29,040 the human review and so forth. So there's some performance questions and being able to really 366 00:41:29,040 --> 00:41:37,680 do this at scale with all of those considerations in place. But again, even to just take your top 367 00:41:37,680 --> 00:41:42,400 hundred products of a portfolio of several thousand and manage to do this process, 368 00:41:42,400 --> 00:41:47,120 which we've now done, that and beyond has proven invaluable. 369 00:41:47,760 --> 00:41:53,360 Amazing. So there you have it, marketers. What creative things can you think of to help scale 370 00:41:53,360 --> 00:42:00,880 up the product? And then, I think, outside of some of the sort of commercial tools that become 371 00:42:00,880 --> 00:42:06,320 available, Emil's told us a wonderful story there about figuring out how to massively scale product 372 00:42:06,320 --> 00:42:11,920 description changes across a wide array of e-commerce products in multiple languages, 373 00:42:11,920 --> 00:42:18,160 and then using AI to also write social posts to promote them. Hopefully that will inspire people 374 00:42:18,160 --> 00:42:22,480 to go away and think, hmm, wonder how I could apply that type of approach to my own work. 375 00:42:22,480 --> 00:42:24,480 I'm not sure if you've ever thought about that with me. 376 00:42:24,480 --> 00:42:29,280 No, of course. I actually have one more point to add on to that, which comes to mind now, 377 00:42:29,280 --> 00:42:38,960 which is just that at the status that this business is for us, we just wouldn't have done it. 378 00:42:38,960 --> 00:42:40,160 Right. 379 00:42:40,160 --> 00:42:48,240 Right. This isn't something that budget-wise, scale-wise, operationally, we could have done 380 00:42:48,240 --> 00:42:53,680 it. It would have just waited until hopefully the business, through other means, had advanced 381 00:42:53,680 --> 00:42:57,920 and matured and gotten to the point where it could then invest in that transition or the process. 382 00:42:58,560 --> 00:43:03,440 So the most important thing about here is not that this is the best end result. Like I said, 383 00:43:03,440 --> 00:43:09,200 quality-wise, it's probably 70 out of 180 out of 100. But if you're going to kind of look at that 384 00:43:09,200 --> 00:43:17,360 from an Occam's Razor's lens, if we can do that now and get 80% of the results, that value at this 385 00:43:17,360 --> 00:43:22,480 stage is so much higher than what we would eventually be able to invest in human-wise if 386 00:43:22,480 --> 00:43:28,560 we ever got to that point in the business. Especially for smaller low-budget or low-operating 387 00:43:28,560 --> 00:43:35,280 teams or smaller operating teams, it's about maximizing efficiency. It's about how do you look 388 00:43:35,280 --> 00:43:39,840 at it through the assistant lens and say, okay, if I could get an assistant to do anything, 389 00:43:39,840 --> 00:43:43,600 what would it be? And then, yes, you have to design that. And there's a lot of system 390 00:43:43,600 --> 00:43:47,680 design and logic that went into that process. You do have to be good at that. But if you're 391 00:43:47,680 --> 00:43:50,720 good at it and creative enough, you could get some pretty excellent outcomes. 392 00:43:51,280 --> 00:43:57,360 Yeah. No, I appreciate you adding that extra detail and that concept of, if I had an assistant, 393 00:43:58,000 --> 00:44:02,560 what would I ask them, I think is a muscle that a lot of us are going to have to develop 394 00:44:02,560 --> 00:44:06,560 when we're looking at our business and marketing challenges. And maybe we can use that as a segue 395 00:44:06,560 --> 00:44:11,520 to, I know you're a deep thinker and you've been doing a lot of thinking. We've had a number of 396 00:44:11,520 --> 00:44:17,600 discussions recently about what the emergence of certainly generative AI tools, but a lot of other 397 00:44:17,600 --> 00:44:23,280 AI enabled tools is going to do in terms of having an impact on how people run their businesses and 398 00:44:23,280 --> 00:44:28,080 market their businesses. So what do you see happening? How do you think this is going to 399 00:44:28,080 --> 00:44:32,400 play out over the next months and years as these tools take hold? 400 00:44:33,360 --> 00:44:39,520 Well, first things first, we all know marketers will ruin everything because 401 00:44:39,520 --> 00:44:48,880 because no one's better at it than we are. So that's number one. But the reason I start 402 00:44:48,880 --> 00:44:53,440 with that is because that's actually the lesson. So like right now, what we see happening in the 403 00:44:53,440 --> 00:45:04,000 AI market is people creating wonderfully intelligent or simple AI tools to more efficiently do or 404 00:45:04,000 --> 00:45:12,000 maximize current standards of marketing, which very quickly means that those standards will 405 00:45:12,000 --> 00:45:19,360 become more saturated and very much more quickly irrelevant. That is not the lens we should be 406 00:45:19,360 --> 00:45:26,000 looking at this is it's not about how do we use AI to maximize the kind of performance or automated 407 00:45:26,000 --> 00:45:31,920 pieces of marketing that work now. It's about when this shift takes place in it might be 18 months 408 00:45:31,920 --> 00:45:36,720 or two years or four years doesn't really matter. What is then the marketing things that will 409 00:45:36,720 --> 00:45:45,280 survive or emerge on the other end of that? And that's where I've, as a portfolio, we have 410 00:45:45,280 --> 00:45:50,000 different companies that operate in different ways. And everyone's feeling a bit of the recession, 411 00:45:50,000 --> 00:45:54,720 everyone's feeling the AI tech pressure on top of that, which makes it more amplified. 412 00:45:54,720 --> 00:45:58,640 So everyone in our teams are going, whoa, whoa, whoa, whoa, wait, wait, what do we have to do? 413 00:45:58,640 --> 00:46:06,160 So I just start printing content like mad using chat GBT. Is that what I should be doing? Like 414 00:46:06,160 --> 00:46:11,760 writing 10 articles a day instead of one good one? That's literally the types of questions that 415 00:46:11,760 --> 00:46:17,520 marketers and even non-marketers in our team are asking right now. And I think that's probably 416 00:46:17,520 --> 00:46:25,280 what a lot of people are feeling some version of conscious or not in the space right now. And so 417 00:46:25,280 --> 00:46:29,440 we've had to kind of look a little bit deeper and start to think, okay, 418 00:46:31,520 --> 00:46:36,320 what works now that will probably work no matter what happens with AI? That's question number one. 419 00:46:36,320 --> 00:46:42,960 Okay. So what will probably not be affected fundamentally as a marketing style or tactic 420 00:46:42,960 --> 00:46:50,080 regardless of AI? And then two, once AI makes certain types of obvious marketing, 421 00:46:50,080 --> 00:46:57,760 probably not work as well. What will be the new tactics that might emerge from that? And that's 422 00:46:57,760 --> 00:47:04,960 a bit more sort of hedging and risk betting and like kind of trying to use your creative brain 423 00:47:04,960 --> 00:47:12,080 as a marketer to think through to the next systemic move. And then the third one is, 424 00:47:12,720 --> 00:47:18,560 regardless of number two, if people start to change their technical behavior, 425 00:47:18,560 --> 00:47:26,640 right? So if plugins on chat GPT get really sticky, which means people move away from 426 00:47:26,640 --> 00:47:34,160 dedicated solution, point and click interaction, they move away from direct, you know, two to five 427 00:47:34,160 --> 00:47:40,720 word search, and they actually do eventually start to adopt a more lingual form of interacting with 428 00:47:40,720 --> 00:47:47,120 multiple services. What will then the knock on effects of that be on other channels or stable 429 00:47:47,120 --> 00:47:53,520 systems, stable quote unquote systems that we use, right? So we did a little bit of very rough math, 430 00:47:53,520 --> 00:48:00,560 right? Please, please no one quote me on these numbers, please. But some very rough math on what 431 00:48:00,560 --> 00:48:10,880 happens if an average of around two or just over 2% of the English speaking market changes from 432 00:48:10,880 --> 00:48:23,920 changes from a dedicated UI based interaction and using search engines to using a lingual platform, 433 00:48:23,920 --> 00:48:30,000 whether that be chat GPT and plugins or something like it, right? 2% moves from the way everyone 434 00:48:30,000 --> 00:48:35,120 operates at the moment to exclusively that English market. So even that you could consider as like 435 00:48:35,120 --> 00:48:41,280 early adopters, that's not a huge amount percentage wise, right? Yeah. It's like, you know, 436 00:48:41,280 --> 00:48:47,760 a couple hundred million people move their everyday behavior over there. What happens? Well, 437 00:48:47,760 --> 00:48:55,440 one of the things that happens is ads almost double in price for almost every single niche. 438 00:48:56,320 --> 00:49:02,880 Right. Right. Because that is the 2% that is digitally active enough that they are sought out 439 00:49:02,880 --> 00:49:11,200 by most advertising categories. And so if you just remove 2% of that population, that entire sort of 440 00:49:11,200 --> 00:49:17,840 stable platform of that of ad performance and how that's calculated and generated starts to buckle. 441 00:49:18,640 --> 00:49:23,680 And so you suddenly get, you know, we're already seeing crazy price increases in most ad platforms 442 00:49:23,680 --> 00:49:29,680 the past two years. This isn't necessarily new. It's just, but it's potential to be accelerated 443 00:49:29,680 --> 00:49:35,600 massively by a very small transitional adoption of the population to this new type of behavior 444 00:49:36,320 --> 00:49:41,360 could have massive implications. Now, if Google ads doubles in price, particularly for e-commerce, 445 00:49:41,360 --> 00:49:47,600 half of e-commerce classes, it's the primary margin on it anymore. No, it's the primary channel, 446 00:49:47,600 --> 00:49:53,520 like ROAS on Google ads and Google shopping ads is like the primary channel for the majority of 447 00:49:53,520 --> 00:50:02,640 mid to at scale e-commerce platforms. So they either have to kind of get way ahead of that 448 00:50:02,640 --> 00:50:07,280 and start changing how those tactics work now so that in two, three years, they're, you know, 449 00:50:07,280 --> 00:50:11,840 it's a much less reliant channel or they're going to wait too long and it's going to start to collapse 450 00:50:11,840 --> 00:50:15,120 and they're going to struggle and a lot of them are going to die. What does that do to the sort of 451 00:50:15,120 --> 00:50:20,720 economic stability of business and having a startup and all of that? These are sort of the 452 00:50:20,720 --> 00:50:27,280 existential crises that we're considering in this equation, right? But it's those sort of three 453 00:50:28,240 --> 00:50:33,840 core points that we've been looking at. Yeah, I completely agree. I think the first one is 454 00:50:33,840 --> 00:50:39,200 definitely the immediate reaction I see in a lot of places because it's practical and applicable 455 00:50:39,200 --> 00:50:44,480 quickly, right? You can log into ChatGPT or some other tool. You can start to have a play with it. 456 00:50:44,480 --> 00:50:47,840 You can get it generating content for you very quickly. And all of a sudden you feel like 457 00:50:47,840 --> 00:50:52,240 I should be doing this no matter what, because surely it's better and more efficient. And we've 458 00:50:52,240 --> 00:50:57,200 talked on the podcast previously and I know you and I have spoken about what happens when you get 459 00:50:57,200 --> 00:51:02,800 a deluge of AI generated content that may not be that well curated or edited and it just is junk. 460 00:51:02,800 --> 00:51:08,560 And, you know, hopefully the cream rises to the top and most of that is at best AI enabled. I 461 00:51:08,560 --> 00:51:13,120 don't think it will be AI generated. So I think there's some stuff in that. In terms of looking 462 00:51:13,120 --> 00:51:19,920 at the wider ramifications like you just described, I don't think most marketers are even thinking 463 00:51:19,920 --> 00:51:23,840 about number one, if I'm honest. I think some are and I think that proportion is growing. 464 00:51:24,480 --> 00:51:28,880 The number of people who are thinking about number two is very low and the number of people 465 00:51:28,880 --> 00:51:33,760 who are looking like number three, it's probably just you, Emil. It's probably just me. Yeah, 466 00:51:33,760 --> 00:51:37,680 it's probably just me having an existential crisis about something happening in four years. That's 467 00:51:37,680 --> 00:51:43,360 that sounds about on point. Right. But I think, you know, when you're trying to predict the future 468 00:51:43,360 --> 00:51:47,760 and manage your business and your marketing strategy, you have to keep these things in mind. 469 00:51:47,760 --> 00:51:53,120 So I think that is hopefully going to give our listeners real food for thought. You know, 470 00:51:53,120 --> 00:52:00,800 where are you listeners on your journey from one to two to three there? Maybe we can summarize 471 00:52:00,800 --> 00:52:05,520 those in a word, Emil, right? Like number one is using tools to produce content now. 472 00:52:05,520 --> 00:52:13,120 Yeah. Number two is how do things change as people start to use these tools and how people 473 00:52:13,120 --> 00:52:17,840 actually search for information changes? And then three is what happens to the rest of the business 474 00:52:17,840 --> 00:52:25,040 ecosystems as a knock on from that factor in terms of how the ad existing ad platforms naturally 475 00:52:25,040 --> 00:52:31,680 adapt to changes in user behavior, the emergence of new ad platforms, et cetera. So is that a fair 476 00:52:31,680 --> 00:52:39,520 summary? Yeah, I think the tweak in review of one and two and at this point, I've been waffling so 477 00:52:39,520 --> 00:52:44,960 much. I've probably lost my own threads there a couple of times, but you know, the one thing 478 00:52:44,960 --> 00:52:51,600 that's important, most important to recall is just like the tools that are being pushed upon us to 479 00:52:51,600 --> 00:52:58,480 use now are those that amplify current strategies, but that doesn't insulate you against once 480 00:52:58,480 --> 00:53:02,160 everyone is using those strategies and they're saturated, no longer relevant. Right? So that's 481 00:53:02,160 --> 00:53:09,200 kind of where point two is leading. Yeah. And that then leads to the inevitable existential crisis of 482 00:53:09,200 --> 00:53:15,360 point three, but like it's really more about like, if you're going to be a good marketer, 483 00:53:15,360 --> 00:53:21,120 you have to stop being a lazy marketer and you have to think, okay, if everyone does this, 484 00:53:21,120 --> 00:53:26,400 what am I going to do? That's a great question. And what are the things that are still going to 485 00:53:26,400 --> 00:53:35,280 work regardless of this? And so I think an inevitable question from someone listening to this, 486 00:53:35,280 --> 00:53:42,640 I hope would be, well, what are those things? You know, and we could go probably on a separate 487 00:53:42,640 --> 00:53:47,280 rant about what some of those things may be. I can leave a little like nugget hanging here, 488 00:53:47,280 --> 00:53:54,080 which is that if it's human, it's probably going to stay. Right? There's a level to that, that is 489 00:53:54,080 --> 00:54:00,160 real. Right? And you have to think of human on many levels. Partnerships are human. Agreed. 490 00:54:00,160 --> 00:54:06,320 Communities that exist within current constructs of self-identity and alignment are human. So 491 00:54:06,320 --> 00:54:11,920 access to those with our partnership are human. Hint, hint, hint. Which we already knew from the 492 00:54:11,920 --> 00:54:15,760 past couple of years, right? We see performance-based marketing kind of collapsing. 493 00:54:15,760 --> 00:54:21,120 We're returning to sort of more of a brand human driven thing anyway. Yeah. And now AI is just 494 00:54:21,120 --> 00:54:26,080 pushing that all the way. I love that. I think we're about on time now, Emil. So let's leave that 495 00:54:26,080 --> 00:54:31,600 dangling in people's minds. I want to come back and maybe we'll do this again in a month or so, 496 00:54:31,600 --> 00:54:35,760 so it's time to really dig into some of those other number three scenarios, because I think 497 00:54:35,760 --> 00:54:39,760 they're really interesting. But I think for now, hopefully everyone that's given you something to 498 00:54:39,760 --> 00:54:44,240 chew on. Emil, thank you so much for your time today. I always love hanging out with you. Good 499 00:54:44,240 --> 00:54:53,200 to speak with you. And yeah, hopefully we'll do this again soon. Yeah, no, it's been a pleasure as 500 00:54:53,200 --> 00:54:59,360 always. I love doing this. So thanks for hearing out my existential ramblings and thanks to your 501 00:54:59,360 --> 00:55:05,520 audience and to your audience as well, like chase us down with questions, especially in the potential 502 00:55:05,520 --> 00:55:10,480 that we might do a follow-up conversation. I'd love to hear what listeners are actually thinking 503 00:55:10,480 --> 00:55:15,280 about or feeling when they start considering this for themselves. Great stuff. Yeah, completely agree. 504 00:55:15,280 --> 00:55:18,720 I'm sure you'll find this on the socials, on the LinkedIn's and the Twitter's when we put it out 505 00:55:18,720 --> 00:55:24,240 there. So we'd love to hear your questions on those channels. Thanks, Emil. I'll speak to you soon. 506 00:55:24,240 --> 00:55:43,360 All right, we'll do. Thanks, Paul. Take care. Thank you for listening to Artificially Intelligent 507 00:55:43,360 --> 00:55:49,680 Marketing. To stay on top of the latest trends, tips and tools in the world of marketing AI, 508 00:55:49,680 --> 00:55:53,840 be sure to subscribe. We look forward to seeing you again next week.