1 00:00:00,000 --> 00:00:09,560 Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the 2 00:00:09,560 --> 00:00:15,280 latest trends, tips and tools in the world of marketing AI, helping you get the best 3 00:00:15,280 --> 00:00:18,120 results from your marketing efforts. 4 00:00:18,120 --> 00:00:23,240 Now let's join our hosts, Paul Avery and Martin Broadhurst. 5 00:00:23,240 --> 00:00:28,440 Welcome to Artificially Intelligent Marketing, a podcast for marketers looking to learn about 6 00:00:28,440 --> 00:00:33,240 how they can use AI to improve their work and make it more efficient. 7 00:00:33,240 --> 00:00:35,840 I am joined today by my good friend Martin Broadhurst. 8 00:00:35,840 --> 00:00:37,720 Hi Martin, how are you doing? 9 00:00:37,720 --> 00:00:39,200 I am very well this week. 10 00:00:39,200 --> 00:00:40,200 Yeah, all good. 11 00:00:40,200 --> 00:00:46,920 Been playing with some new tools and presenting a seminar this morning on introduction to 12 00:00:46,920 --> 00:00:49,060 marketing AI, funnily enough. 13 00:00:49,060 --> 00:00:53,120 So yeah, very much front and center of my thinking right now. 14 00:00:53,120 --> 00:00:58,440 A week of mass AI for you then, and hopefully you can distill some of all that good stuff 15 00:00:58,440 --> 00:01:01,400 into the cool stories we've got this week. 16 00:01:01,400 --> 00:01:06,360 Myself, I've been really busy within Biostra, actually doing loads of work, speaking to 17 00:01:06,360 --> 00:01:12,320 some really cool clients with really exciting technologies that are going to change cancer 18 00:01:12,320 --> 00:01:13,840 research and a bunch of other cool stuff. 19 00:01:13,840 --> 00:01:16,600 So it's been quite an exciting week for me speaking to clients to be honest. 20 00:01:16,600 --> 00:01:22,040 But I also noticed, and we've been very active on the old WhatsApps, haven't we, this week 21 00:01:22,040 --> 00:01:25,960 night, because a lot of important and interesting things have happened this week as we alert, 22 00:01:25,960 --> 00:01:27,240 as we'll come to look at. 23 00:01:27,240 --> 00:01:29,880 Firstly, we're going to focus on a handful of stories. 24 00:01:29,880 --> 00:01:35,280 It's going to include the apparent imminent release of GPT-4, might have been a little 25 00:01:35,280 --> 00:01:41,020 slip up or maybe it was deliberate by a Microsoft employee yesterday that we're going to cover. 26 00:01:41,020 --> 00:01:45,640 We're going to look at the rise of CRM AI assistance because that blew up this week 27 00:01:45,640 --> 00:01:46,800 as well. 28 00:01:46,800 --> 00:01:51,080 We're going to talk a little bit about domain specific large language models. 29 00:01:51,080 --> 00:01:55,680 In this case, we're going to look at bio GPT and what that might mean for marketers. 30 00:01:55,680 --> 00:01:58,640 And even though it's not a super marketing application, it is mega cool. 31 00:01:58,640 --> 00:02:03,040 We're going to talk about a paper that was actually released in November last year, but 32 00:02:03,040 --> 00:02:09,360 for some reason caught attention this week where researchers used MRIs and AI tools to 33 00:02:09,360 --> 00:02:12,800 try and predict what images people were looking at, which is kind of cool. 34 00:02:12,800 --> 00:02:14,360 And we'll get into that. 35 00:02:14,360 --> 00:02:17,920 We're also going to have a little look at our tool of the week, which is this week is 36 00:02:17,920 --> 00:02:21,640 runway.ml. 37 00:02:21,640 --> 00:02:26,020 Before we get into all of that, Martin and I wanted to thank everyone for the amazing 38 00:02:26,020 --> 00:02:28,840 feedback that we had when we launched our first episode last week. 39 00:02:28,840 --> 00:02:33,480 We got way more downloads and interest than we were expecting. 40 00:02:33,480 --> 00:02:38,360 And we also had some really good questions about how we do the podcast because a couple 41 00:02:38,360 --> 00:02:45,560 of eagle-eyed eagle-eared listeners spotted that the voice at the beginning of the podcast 42 00:02:45,560 --> 00:02:49,240 didn't sound quite human and of course they were absolutely right. 43 00:02:49,240 --> 00:02:55,300 We used an AI tool to synthesize that because we've been focusing on how do we produce this 44 00:02:55,300 --> 00:02:59,480 entire podcast using as many AI first tools as we can. 45 00:02:59,480 --> 00:03:04,840 So the intro voice was actually created using a tool called play.ht. 46 00:03:04,840 --> 00:03:05,840 So go and check that out. 47 00:03:05,840 --> 00:03:08,800 That was actually a pretty streamlined and easy experience. 48 00:03:08,800 --> 00:03:14,560 Gave it a script, chose which voice we wanted, recaptured it a few times because you get 49 00:03:14,560 --> 00:03:20,120 it in different flavors, in different phrase, not phraseologies, but like tempos to try 50 00:03:20,120 --> 00:03:21,120 and make it feel more natural. 51 00:03:21,120 --> 00:03:23,560 So we used that on the intro music. 52 00:03:23,560 --> 00:03:29,120 We used one of your favorite tools, Martin, that you recommended Ava or AIV. 53 00:03:29,120 --> 00:03:30,840 That was really cool as well. 54 00:03:30,840 --> 00:03:32,800 Gave us some really good options to choose from. 55 00:03:32,800 --> 00:03:37,220 And if we wanted, we could get into the nitty gritty of even changing the, the midi clips 56 00:03:37,220 --> 00:03:41,760 for those of us played with music, you could actually get right in and edit the AI driven 57 00:03:41,760 --> 00:03:42,760 creations. 58 00:03:42,760 --> 00:03:46,280 We didn't to be honest, because we wanted something quick and easy, but we, but we could 59 00:03:46,280 --> 00:03:47,560 have. 60 00:03:47,560 --> 00:03:52,280 On the video editing side, including audio enhancement and captions and stuff, we use 61 00:03:52,280 --> 00:03:53,280 runway.ml. 62 00:03:53,280 --> 00:03:54,680 We'll talk about that a bit later. 63 00:03:54,680 --> 00:03:58,360 And the videos have been recorded using zoom. 64 00:03:58,360 --> 00:04:04,280 You did quite a bit on the sort of RSS feed cover image and sort of summer track text 65 00:04:04,280 --> 00:04:05,720 summaries of the podcast. 66 00:04:05,720 --> 00:04:07,760 Martin, what sort of tools were you using for that? 67 00:04:07,760 --> 00:04:14,040 Yeah, so the AI generated cover art was from, it was, I think it was a stable diffusion 68 00:04:14,040 --> 00:04:19,040 output, but the tool that I used to create it was super machine dot art, which is a really 69 00:04:19,040 --> 00:04:20,040 cool site. 70 00:04:20,040 --> 00:04:27,600 Actually they've implemented lots of different AI text to image models and fine tune some 71 00:04:27,600 --> 00:04:28,600 of themself. 72 00:04:28,600 --> 00:04:33,800 So if you want to use the likes of mid journey stable diffusion, and then some fine tune 73 00:04:33,800 --> 00:04:40,080 variants of those, check out super machine dot art. 74 00:04:40,080 --> 00:04:41,840 Yeah that was the text summarisation. 75 00:04:41,840 --> 00:04:43,520 Sorry to, yeah, I forgot. 76 00:04:43,520 --> 00:04:47,120 That was a, yeah, that was a slightly low winded approach. 77 00:04:47,120 --> 00:04:52,480 I was hoping to use Coheer, but after we spoke about Coheer summarisation last week, I realised 78 00:04:52,480 --> 00:04:58,000 that the, the output quite frankly would have just been too succinct given all of the broad 79 00:04:58,000 --> 00:04:59,640 range of topics that we covered. 80 00:04:59,640 --> 00:05:07,840 So it was a, it was a somewhat manual effort using a transcript generated from RSS.com, 81 00:05:07,840 --> 00:05:14,000 who hosts our podcast, took the transcript and then basically fed it into GPT in bite 82 00:05:14,000 --> 00:05:20,040 sized chalks, taking out the outputs from there and uploading them to the website. 83 00:05:20,040 --> 00:05:21,040 Good stuff. 84 00:05:21,040 --> 00:05:24,000 So for those of you that asked what we were using, that's what we're currently using. 85 00:05:24,000 --> 00:05:28,640 We're going to be exploring and playing with a bunch of other tools as well as we go so 86 00:05:28,640 --> 00:05:32,960 that we get a decent feel for what would be the best tools for us to use. 87 00:05:32,960 --> 00:05:36,160 We also tried podcastle at the beginning. 88 00:05:36,160 --> 00:05:41,240 This is episode two, but actually it's episode three because we recorded it in episode four, 89 00:05:41,240 --> 00:05:46,800 episode one, and we had a failure in podcastle where it only recorded bits of the audio and 90 00:05:46,800 --> 00:05:47,880 we couldn't use it. 91 00:05:47,880 --> 00:05:51,840 So we're not using podcastle anymore. 92 00:05:51,840 --> 00:05:55,640 Once bitten, once shy in this case rather than twice bitten, because we just couldn't 93 00:05:55,640 --> 00:05:56,640 risk it. 94 00:05:56,640 --> 00:05:58,080 So we've had a little bit of a play with that. 95 00:05:58,080 --> 00:06:00,680 That is a powerful tool, really fun and easy to use. 96 00:06:00,680 --> 00:06:04,360 But of course, if we couldn't rely on it, which we felt we couldn't after that first 97 00:06:04,360 --> 00:06:06,960 little slip up, we had to try something else. 98 00:06:06,960 --> 00:06:11,640 So stay tuned for other tools that we bring into the mix as we produce this podcast. 99 00:06:11,640 --> 00:06:12,640 Right. 100 00:06:12,640 --> 00:06:15,680 Let us move on to the stories of the week. 101 00:06:15,680 --> 00:06:20,640 Martin, first one, you spotted this last night and pinged it over to me and I nearly dropped 102 00:06:20,640 --> 00:06:21,640 my phone. 103 00:06:21,640 --> 00:06:24,680 Tell us about these GPT four rumors. 104 00:06:24,680 --> 00:06:30,240 Yeah, so this was the talk of Twitter AI last night. 105 00:06:30,240 --> 00:06:32,280 Everyone was jumping on over it. 106 00:06:32,280 --> 00:06:33,280 Yes. 107 00:06:33,280 --> 00:06:44,640 So the CTO of Microsoft Germany said that this was at an internal Microsoft event, Microsoft 108 00:06:44,640 --> 00:06:51,640 Germany event this week, said that GPT four is coming next week. 109 00:06:51,640 --> 00:06:57,320 And specifically said we will introduce GPT four next week. 110 00:06:57,320 --> 00:07:03,240 We will have multimodal models that will offer completely different possibilities. 111 00:07:03,240 --> 00:07:08,240 For example, videos. 112 00:07:08,240 --> 00:07:12,900 So there's been a lot of talk about GPT four for well, for ages, probably since the release 113 00:07:12,900 --> 00:07:18,600 of GPT three, in all honesty, people start getting excited about the next big thing. 114 00:07:18,600 --> 00:07:22,760 It's all been completely under reps. 115 00:07:22,760 --> 00:07:24,800 No details have leaked. 116 00:07:24,800 --> 00:07:26,320 There's been wild speculation. 117 00:07:26,320 --> 00:07:32,480 In fact, before coming onto this part, I thought of then just check out what were the what 118 00:07:32,480 --> 00:07:37,200 were the range of different estimates for the parameters in the model? 119 00:07:37,200 --> 00:07:42,360 So 175 billion parameters in GPT three. 120 00:07:42,360 --> 00:07:47,600 There are guesses or speculation that it's going to have a trillion. 121 00:07:47,600 --> 00:07:50,520 I saw one report saying 100 trillion. 122 00:07:50,520 --> 00:07:55,360 So quite frankly, it is anybody's guess how big the model will actually be when it comes 123 00:07:55,360 --> 00:07:56,360 out. 124 00:07:56,360 --> 00:07:59,120 But I think the interesting thing is multimodal. 125 00:07:59,120 --> 00:08:00,120 Absolutely. 126 00:08:00,120 --> 00:08:01,920 It's not the size of your model. 127 00:08:01,920 --> 00:08:03,760 It's what you do with it, mine, to be honest. 128 00:08:03,760 --> 00:08:08,480 And I think in this case, it'll be interesting to see how relevant that ends up being. 129 00:08:08,480 --> 00:08:11,280 But you're right on the multimodal side. 130 00:08:11,280 --> 00:08:12,640 Wow. 131 00:08:12,640 --> 00:08:13,800 What are we thinking here? 132 00:08:13,800 --> 00:08:16,600 What are we expecting we might get? 133 00:08:16,600 --> 00:08:22,600 Well, in the dream scenario, you know, you just type in a prompt and it spits out a fully 134 00:08:22,600 --> 00:08:25,240 formed video and says, there you go. 135 00:08:25,240 --> 00:08:28,920 There's your there's your 10 second animated GIF. 136 00:08:28,920 --> 00:08:35,600 I would be surprised if it's actual video production capabilities, given where the the 137 00:08:35,600 --> 00:08:41,200 text to video research and development currently is. 138 00:08:41,200 --> 00:08:47,760 You know, if you look at what Google has released with imaging and the text to video, very sure 139 00:08:47,760 --> 00:08:50,280 clips that they're capable of producing, they look cool. 140 00:08:50,280 --> 00:08:52,000 They're certainly doing a good job. 141 00:08:52,000 --> 00:08:55,040 I wouldn't be surprised if they're slightly further along than that. 142 00:08:55,040 --> 00:09:03,600 But I would still be surprised if it had video production capabilities. 143 00:09:03,600 --> 00:09:08,960 That said, I wouldn't be surprised if it was more focused around the input. 144 00:09:08,960 --> 00:09:13,400 There's a bit of speculation reading in some of the groups and the AI communities today 145 00:09:13,400 --> 00:09:17,120 saying it could be that this is multimodal from an input perspective. 146 00:09:17,120 --> 00:09:23,020 So you can upload an image rather than just being text prompts as the input. 147 00:09:23,020 --> 00:09:29,280 You can upload an image with a prompt and saying, describe this scene or describe this 148 00:09:29,280 --> 00:09:31,300 video, transcribe this video. 149 00:09:31,300 --> 00:09:36,320 So it's still not bringing in elements of things like the capabilities of say, whisper 150 00:09:36,320 --> 00:09:40,480 AI, which is the transcription model. 151 00:09:40,480 --> 00:09:42,080 But it's all bringing that into one. 152 00:09:42,080 --> 00:09:45,080 So it's not necessarily producing the video output. 153 00:09:45,080 --> 00:09:49,640 It's editing the inputs or. 154 00:09:49,640 --> 00:09:54,440 So multimodal inputs rather than multimodal outputs potentially at this point. 155 00:09:54,440 --> 00:09:56,760 And I guess we're going to find out next week. 156 00:09:56,760 --> 00:10:01,320 I mean, everything everything that we say here is absolute speculation. 157 00:10:01,320 --> 00:10:02,640 It wouldn't be surprising either, would it? 158 00:10:02,640 --> 00:10:11,600 If it was text based, sorry, image based outputs and they were integrating Dali capabilities 159 00:10:11,600 --> 00:10:16,860 in the outputs, we know that it can already with chat GBD, it can output code, it can 160 00:10:16,860 --> 00:10:20,600 output text, it can output formatted text. 161 00:10:20,600 --> 00:10:27,520 So in tables and things like that, it wouldn't be a stretch to imagine that it's going to 162 00:10:27,520 --> 00:10:32,240 be able to spit out an image in the same way that Dali can spit out an image. 163 00:10:32,240 --> 00:10:35,320 That wouldn't completely blow my mind. 164 00:10:35,320 --> 00:10:40,080 And to be honest, it wouldn't wouldn't be a huge surprise if they did introduce some 165 00:10:40,080 --> 00:10:46,120 video, but in the kind of short form animated GIF, you know, five, ten second long clip 166 00:10:46,120 --> 00:10:47,120 region. 167 00:10:47,120 --> 00:10:48,640 That would be quite cool. 168 00:10:48,640 --> 00:10:54,720 That would potentially revolutionize email communication at Bystrat where we love a GIF, 169 00:10:54,720 --> 00:10:58,520 but we have to use existing GIFs, whereas to be able to describe the GIF that you want 170 00:10:58,520 --> 00:11:05,080 and then get open AI to chat GPT or GPT4 or whatever ends up being to produce that would 171 00:11:05,080 --> 00:11:06,480 be quite fun. 172 00:11:06,480 --> 00:11:11,960 I guess for marketers, this is just an example of how quickly things are developing. 173 00:11:11,960 --> 00:11:14,900 Chat GPT came at end of November, early December. 174 00:11:14,900 --> 00:11:18,280 We're now talking about all these explosion of other tools, some of which we talked about 175 00:11:18,280 --> 00:11:19,280 last week. 176 00:11:19,280 --> 00:11:24,720 And now we're looking at a tool that potentially has multimodal inputs and also potentially 177 00:11:24,720 --> 00:11:26,960 multimodal outputs. 178 00:11:26,960 --> 00:11:32,760 So as a marketer, just further increasing the power of the tools that we have available 179 00:11:32,760 --> 00:11:39,360 all in one place to get our work done, be more creative, get ideas, first versions of 180 00:11:39,360 --> 00:11:44,040 ideas out for us to start thinking about getting feedback on and tweaking. 181 00:11:44,040 --> 00:11:45,360 That could all be pretty cool. 182 00:11:45,360 --> 00:11:47,680 Yeah, very much so. 183 00:11:47,680 --> 00:11:52,000 I remember saying at the start of the year, speaking to friends of people interested in 184 00:11:52,000 --> 00:11:57,840 this space that I would expect a multimodal model to come out before the end of the year. 185 00:11:57,840 --> 00:12:03,000 It wouldn't surprise me if that's when we see the actual release of it. 186 00:12:03,000 --> 00:12:10,160 If we look at the timeline for GPT3, there was something like a 17 month gap between 187 00:12:10,160 --> 00:12:18,880 the paper announcing the initial GPT3 model and then the actual release of the unrestricted 188 00:12:18,880 --> 00:12:21,360 API where anybody could access it. 189 00:12:21,360 --> 00:12:22,920 I think it was about a 17 gap. 190 00:12:22,920 --> 00:12:28,180 So it wouldn't surprise me if there was a decent sized gap between the announcement 191 00:12:28,180 --> 00:12:32,880 and then the actual reality of it. 192 00:12:32,880 --> 00:12:37,680 But when you think about the investors that have come into OpenAI in recent years, Microsoft 193 00:12:37,680 --> 00:12:41,400 being a big one, people are going to be wanting to see developments. 194 00:12:41,400 --> 00:12:47,040 And I think OpenAI getting chat and GPT out in the way that they did has shown that there's 195 00:12:47,040 --> 00:12:49,160 appetite and there's interest in it. 196 00:12:49,160 --> 00:12:53,800 So it wouldn't surprise me if they start releasing things slightly quicker than they did before. 197 00:12:53,800 --> 00:12:57,600 It took them a year from releasing Dali to Dali 2. 198 00:12:57,600 --> 00:12:58,840 Dali was a research paper. 199 00:12:58,840 --> 00:13:02,920 Dali 2 was the lab where you could actually start creating things on it. 200 00:13:02,920 --> 00:13:08,480 I think we're going to see a bit more acceleration from announcement to production. 201 00:13:08,480 --> 00:13:09,480 Right. 202 00:13:09,480 --> 00:13:13,080 So we're going to hear what the cool stuff is next week, but we may not be able to use 203 00:13:13,080 --> 00:13:15,480 it or play with it for some time, you think? 204 00:13:15,480 --> 00:13:16,960 Yeah, I think so. 205 00:13:16,960 --> 00:13:23,560 Which would coincide with an event that Microsoft have got going on as well to do with AI next 206 00:13:23,560 --> 00:13:24,560 week. 207 00:13:24,560 --> 00:13:29,200 So it just makes you wonder whether the kind of stars are aligning and this is actually 208 00:13:29,200 --> 00:13:35,080 something of an orchestrated run up and launch. 209 00:13:35,080 --> 00:13:37,080 Cool, cool. 210 00:13:37,080 --> 00:13:38,080 All right. 211 00:13:38,080 --> 00:13:43,500 Well, we will, I'm sure we'll feature this story next week, assuming we get the news 212 00:13:43,500 --> 00:13:45,120 that we're hoping for. 213 00:13:45,120 --> 00:13:51,880 However, there was some even more additional exciting news items this week. 214 00:13:51,880 --> 00:13:59,000 Let's talk about the rise of CRM AI assistance where all the major players basically launched 215 00:13:59,000 --> 00:14:04,880 one or basically provided us with an announcement about one all in the same week. 216 00:14:04,880 --> 00:14:11,440 Tell us a bit more about Chatspot and Einstein GPT and all the rest, Martin. 217 00:14:11,440 --> 00:14:20,320 Chatspot.ai, brought to you by the coding fingers of Dharmesh Shah, co-founder of HubSpot. 218 00:14:20,320 --> 00:14:24,600 Our background, the way we've actually met was through the HubSpot ecosystem and the 219 00:14:24,600 --> 00:14:26,240 HubSpot partner program. 220 00:14:26,240 --> 00:14:30,120 Both of us have been very involved in that for a number of years now. 221 00:14:30,120 --> 00:14:34,000 Yeah, so it was interesting to see Dharmesh announce this. 222 00:14:34,000 --> 00:14:41,560 It's very much an alpha release, but HubSpot customers can go to Chatspot.ai and connect 223 00:14:41,560 --> 00:14:45,600 their HubSpot portal with the Chatspot system. 224 00:14:45,600 --> 00:14:51,040 And then they've got ChatGPT connected to their CRM. 225 00:14:51,040 --> 00:14:58,440 And given that HubSpot is a wider platform than just a CRM, it's your marketing, it's 226 00:14:58,440 --> 00:15:01,600 your customer support, it's your sales resources as well. 227 00:15:01,600 --> 00:15:08,760 You get access to all of that data through a chat interface. 228 00:15:08,760 --> 00:15:09,760 Very cool. 229 00:15:09,760 --> 00:15:12,200 Have you managed to get on the alpha yet? 230 00:15:12,200 --> 00:15:17,120 Yeah, you sent that WhatsApp link over to me and said, this is cool. 231 00:15:17,120 --> 00:15:20,880 I signed up immediately and within hours I got access. 232 00:15:20,880 --> 00:15:25,080 I think 4,000 users on the wait list for that. 233 00:15:25,080 --> 00:15:27,840 At last count, I saw Dharmesh post on LinkedIn. 234 00:15:27,840 --> 00:15:31,920 So clearly a lot of people interested in getting their hands on it. 235 00:15:31,920 --> 00:15:32,920 What are your thoughts? 236 00:15:32,920 --> 00:15:33,920 I know that you've been playing with it. 237 00:15:33,920 --> 00:15:34,920 Yeah, I've had a play. 238 00:15:34,920 --> 00:15:39,200 I saw Dharmesh put a post on that one of his to-dos every day was to approve another 100 239 00:15:39,200 --> 00:15:40,200 users onto the system. 240 00:15:40,200 --> 00:15:44,800 So I think the fact that we've both got on quite quickly is probably quite lucky to be 241 00:15:44,800 --> 00:15:50,680 honest because it seems to me like they're managing that quite tightly. 242 00:15:50,680 --> 00:15:54,120 If you watch, there's like a 20 minute intro video where Dharmesh takes us through the 243 00:15:54,120 --> 00:16:02,480 system and the possibilities are huge because imagine being able to go into HubSpot, Salesforce, 244 00:16:02,480 --> 00:16:03,840 whatever your tool is. 245 00:16:03,840 --> 00:16:09,480 And rather than like 15, 20, 30 clicks trying to get to different reports or bits of information, 246 00:16:09,480 --> 00:16:14,520 in essence, you just ask the computer a bit like you would like if you're in Star Trek 247 00:16:14,520 --> 00:16:16,120 to just summarize information for you. 248 00:16:16,120 --> 00:16:21,800 And in fact, they've even got a sort of Google style microphone button where you can speak 249 00:16:21,800 --> 00:16:27,040 to Chatspot and ask it stuff and it will auto transcribe what you asked for and then give 250 00:16:27,040 --> 00:16:28,040 you the data. 251 00:16:28,040 --> 00:16:34,440 So some of the great things in Dharmesh's intro video include things like adding a new 252 00:16:34,440 --> 00:16:38,120 contact to HubSpot by just asking HubSpot to add it. 253 00:16:38,120 --> 00:16:41,440 So rather than have to click through to contacts and then click add contact and then write 254 00:16:41,440 --> 00:16:45,000 in someone's name and then connect it to a company profile. 255 00:16:45,000 --> 00:16:49,240 If you use the little microphone button, you can basically just speak in, please add Bob 256 00:16:49,240 --> 00:16:54,400 Bobson from Acme Go and it will do all that for you. 257 00:16:54,400 --> 00:16:58,040 I also really loved the sales prospecting email example. 258 00:16:58,040 --> 00:17:04,520 So writing a sales prospecting email about a new feature released for a product, but 259 00:17:04,520 --> 00:17:10,700 the emails heavily customized by what the CRM data has about that individual contact. 260 00:17:10,700 --> 00:17:12,200 You could imagine doing that at scale. 261 00:17:12,200 --> 00:17:17,840 And the example given really felt like a human had done the due diligence on where do they 262 00:17:17,840 --> 00:17:21,840 work, what are they interested in, how long has their company been going and all that 263 00:17:21,840 --> 00:17:22,840 type of stuff. 264 00:17:22,840 --> 00:17:25,120 And it was all just fed into the email example. 265 00:17:25,120 --> 00:17:30,160 Having played with it, I break it a lot. 266 00:17:30,160 --> 00:17:31,480 It is an alpha. 267 00:17:31,480 --> 00:17:33,300 That's absolutely the expectation. 268 00:17:33,300 --> 00:17:35,960 That's why it's been launched to a small group of users. 269 00:17:35,960 --> 00:17:40,280 And so we can break it and we can feedback, oh, it would be so cool if we could do this. 270 00:17:40,280 --> 00:17:41,860 But by the way, it didn't work. 271 00:17:41,860 --> 00:17:44,060 So I have been breaking it a lot. 272 00:17:44,060 --> 00:17:48,660 That being said, there's a few things I was able to do where I was able to get information 273 00:17:48,660 --> 00:17:51,760 from it that I just didn't think would be easy to get. 274 00:17:51,760 --> 00:17:57,140 So I asked it to give me the location of some of the leads in our database because we manage 275 00:17:57,140 --> 00:18:00,360 who's going to do an intro call based on where the person is based. 276 00:18:00,360 --> 00:18:03,780 So we have a US team, we have a UK team, a European team. 277 00:18:03,780 --> 00:18:08,160 So I'd ask for things like the IP country of a person's name and it would figure out 278 00:18:08,160 --> 00:18:11,560 who the right person was and then tell me what that country was, which was pretty easy 279 00:18:11,560 --> 00:18:13,600 and quick for me to then root in these. 280 00:18:13,600 --> 00:18:16,700 I realized I could do that through workflows and stuff in the backend, but I just wanted 281 00:18:16,700 --> 00:18:19,560 to see if I had to ask it a quick question. 282 00:18:19,560 --> 00:18:23,640 Like I was asking, I don't know, a sales admin support person, could I just get an instant 283 00:18:23,640 --> 00:18:25,960 answer and the answer was yes, I could. 284 00:18:25,960 --> 00:18:27,320 So that was pretty cool. 285 00:18:27,320 --> 00:18:29,980 I've also enjoyed playing with some of the keyword capabilities. 286 00:18:29,980 --> 00:18:34,780 So asking it what keywords do certain competitors appear to be bidding on and it would give 287 00:18:34,780 --> 00:18:37,240 me really good Intel on that, which was fab. 288 00:18:37,240 --> 00:18:41,200 What else was I doing with it? 289 00:18:41,200 --> 00:18:43,460 Suggested keywords for us to go after, which was great. 290 00:18:43,460 --> 00:18:51,460 I was lamenting using ChatGPT for keyword research in our last podcast because it doesn't know 291 00:18:51,460 --> 00:18:53,360 things like search volumes and stuff. 292 00:18:53,360 --> 00:19:00,160 Well, once you connect ChatGPT to HubSpot's data repository and some of the tools in HubSpot, 293 00:19:00,160 --> 00:19:01,700 that's a different game entirely, right? 294 00:19:01,700 --> 00:19:05,360 Because now you're starting to see real search volumes and all of those types of things. 295 00:19:05,360 --> 00:19:10,360 So yes, I'm super excited about what we're going to be able to do with this, but at the 296 00:19:10,360 --> 00:19:14,200 moment they're still just trying to get a feel for what are the types of things that 297 00:19:14,200 --> 00:19:17,880 people are going to ask of it and how do they iron out those bugs? 298 00:19:17,880 --> 00:19:21,560 Yeah, that's very reminiscent of my experience. 299 00:19:21,560 --> 00:19:26,160 I was able to break it remarkably quickly. 300 00:19:26,160 --> 00:19:31,260 I just asked it to tell me what two pipelines, I've got two sales pipelines set up in my 301 00:19:31,260 --> 00:19:32,260 account. 302 00:19:32,260 --> 00:19:37,320 I just asked it to tell me the names of them because I thought I knew the name of one of 303 00:19:37,320 --> 00:19:42,880 them and I said, can you tell me what's my current value in one of the stages in the 304 00:19:42,880 --> 00:19:43,880 pipeline? 305 00:19:43,880 --> 00:19:45,920 And it said that pipeline doesn't exist. 306 00:19:45,920 --> 00:19:49,480 Okay, well tell me what pipelines I've got. 307 00:19:49,480 --> 00:19:57,200 And then it brought up some script that it was like from the hub database, just said 308 00:19:57,200 --> 00:19:59,320 like pipeline one and pipeline two. 309 00:19:59,320 --> 00:20:01,320 That's not helpful. 310 00:20:01,320 --> 00:20:05,200 You know, it's an alpha and it's going to have that. 311 00:20:05,200 --> 00:20:11,280 And I really respect that Darmesh has put it out so quickly and said, go play, tell 312 00:20:11,280 --> 00:20:12,960 us what you want to do. 313 00:20:12,960 --> 00:20:16,800 Because the potential of it is really, really vast. 314 00:20:16,800 --> 00:20:22,520 I also loved the idea that it's like an assistant. 315 00:20:22,520 --> 00:20:29,040 It is like having an assistant, a CRM admin, a data analyst just there. 316 00:20:29,040 --> 00:20:33,280 I went in and just typed in who was my most recent lead? 317 00:20:33,280 --> 00:20:35,880 What was their original source? 318 00:20:35,880 --> 00:20:41,720 What's the value of pipeline, which it didn't quite understand, but just that, you know, 319 00:20:41,720 --> 00:20:44,560 having that available with a different interface. 320 00:20:44,560 --> 00:20:51,400 And I'm sure as this iterates, if you've got a kind of desktop assistant or built into 321 00:20:51,400 --> 00:20:56,200 the mobile app where you can just fire off a question or literally speak into it, as 322 00:20:56,200 --> 00:21:03,240 you say, the microphone control, it is going to be a productivity game changer for everyone. 323 00:21:03,240 --> 00:21:06,440 For sales, for admins, for marketers. 324 00:21:06,440 --> 00:21:16,080 Yeah, I think it's got huge potential and so does Salesforce because they too have entered 325 00:21:16,080 --> 00:21:21,960 this space this week with Einstein GPT. 326 00:21:21,960 --> 00:21:29,520 They've also announced a $250 million fund to invest in generative AI startups alongside 327 00:21:29,520 --> 00:21:30,520 this. 328 00:21:30,520 --> 00:21:34,680 Yeah, I think it's fair to say that this space is hotting up. 329 00:21:34,680 --> 00:21:35,680 Absolutely. 330 00:21:35,680 --> 00:21:42,400 We also got Microsoft's new GPT powered chat bot for Dynamics 365 CRM as well, which was 331 00:21:42,400 --> 00:21:47,560 going to apparently be able to automatically generate answers to customer emails, suggest 332 00:21:47,560 --> 00:21:53,240 marketing and sales campaigns and do things like summarizing team's chat threads and stuff. 333 00:21:53,240 --> 00:21:57,760 So very similar capabilities across a lot of these tools, I expect. 334 00:21:57,760 --> 00:22:02,720 I also think as they evolve, we'll see some modernization there. 335 00:22:02,720 --> 00:22:07,360 So probably it's likely that most of the tools will do the same things and it's what ecosystem 336 00:22:07,360 --> 00:22:09,400 are you already in? 337 00:22:09,400 --> 00:22:15,760 But certainly if you are using a CRM and you're not using one of the big players like a Dynamics 338 00:22:15,760 --> 00:22:22,760 or Salesforce or HubSpot, you need your provider to be thinking about this type of stuff and 339 00:22:22,760 --> 00:22:24,280 building these types of tools in. 340 00:22:24,280 --> 00:22:28,120 Otherwise you're massively going to be left behind, so you can imagine how this could 341 00:22:28,120 --> 00:22:35,640 cause a migration of people using sort of smaller platforms built by smaller companies. 342 00:22:35,640 --> 00:22:40,560 And actually this is the thing that pushes them into HubSpot, into Salesforce, into Dynamics 343 00:22:40,560 --> 00:22:44,040 to be able to get access to this sort of power. 344 00:22:44,040 --> 00:22:46,920 So I think that'll be an interesting thing to see how that plays out. 345 00:22:46,920 --> 00:22:52,520 The other thing I noticed when I was playing with this was I had to think differently and 346 00:22:52,520 --> 00:22:57,540 I've been trying to, how do I articulate how I had to try and think? 347 00:22:57,540 --> 00:23:02,640 And rather than my brain instantly going through a click-based process in my mind, like I need 348 00:23:02,640 --> 00:23:06,640 to go here to do this, to do this, to do this. 349 00:23:06,640 --> 00:23:11,240 In order to get real valuable stuff out of the system, I had to think about insights 350 00:23:11,240 --> 00:23:14,120 I wanted to know, outcomes that I wanted. 351 00:23:14,120 --> 00:23:15,560 It was a different style of thinking. 352 00:23:15,560 --> 00:23:21,960 It was less to-do based and it was more instantly what's the outcome that I want. 353 00:23:21,960 --> 00:23:26,560 Because I'll give an example, sometimes I'll go into the reports tool in HubSpot and I 354 00:23:26,560 --> 00:23:27,800 might look for a specific thing. 355 00:23:27,800 --> 00:23:28,800 Okay. 356 00:23:28,800 --> 00:23:32,280 I know the outcome then, but sometimes I'll play with the filters on the tool to like, 357 00:23:32,280 --> 00:23:35,480 oh, but I wonder if, oh, but I wonder if. 358 00:23:35,480 --> 00:23:39,600 I think that's kind of a discovery-based process where I've got the filters in front of me 359 00:23:39,600 --> 00:23:42,120 and I see them and I'm like, oh, I can filter by that. 360 00:23:42,120 --> 00:23:44,280 And I didn't realize that would be quite interesting to see. 361 00:23:44,280 --> 00:23:50,960 So I think if you're going to be asking a chat bot to do it, there's the potential that 362 00:23:50,960 --> 00:23:55,200 we're going to need to think a bit more creatively about what outcomes we're after. 363 00:23:55,200 --> 00:24:00,880 We may need to learn our systems better to even know what we can ask them to do. 364 00:24:00,880 --> 00:24:06,040 And potentially as the tools get better, the only real limitation could be the creativity 365 00:24:06,040 --> 00:24:11,720 we can come up with in terms of, I wonder if I, you know, computer, I wonder if you 366 00:24:11,720 --> 00:24:14,120 can show me this and this combined in this plot. 367 00:24:14,120 --> 00:24:15,720 And it'll be like, yeah, Paul, no worries. 368 00:24:15,720 --> 00:24:19,280 And it will show me and I can maybe see some insights about how our pipelines progressing 369 00:24:19,280 --> 00:24:23,960 or what our best lead sources are that I may not have even thought to do in the past. 370 00:24:23,960 --> 00:24:30,080 So at the same time, as we see these tools evolve and mature, I think as people will 371 00:24:30,080 --> 00:24:34,280 have to change our approach a bit in terms of how we get the information and insights 372 00:24:34,280 --> 00:24:35,280 that we want. 373 00:24:35,280 --> 00:24:36,280 Yeah. 374 00:24:36,280 --> 00:24:43,160 I find thinking with chat GPT with the chat interface, if you're, so if you look at the 375 00:24:43,160 --> 00:24:48,560 way that the chat GPT API is set up, there is basically an instruction at the start to 376 00:24:48,560 --> 00:24:51,640 say what's the role of the chat in this scenario. 377 00:24:51,640 --> 00:24:58,600 So you'll say you are a travel assistant at a airline helping customers do this, that 378 00:24:58,600 --> 00:24:59,600 the other way. 379 00:24:59,600 --> 00:25:00,880 So that's the kind of input at the start. 380 00:25:00,880 --> 00:25:07,640 I think it helps to go into the chat spot or the CRM assistant AI, whatever, what I 381 00:25:07,640 --> 00:25:09,560 call it, go into it with that mindset. 382 00:25:09,560 --> 00:25:13,020 What do I want this to play at this moment in time? 383 00:25:13,020 --> 00:25:16,400 Because when I first went into chat spot, it's got all of the suggested prompts down 384 00:25:16,400 --> 00:25:18,320 the side and it can kind of do everything. 385 00:25:18,320 --> 00:25:19,640 It can write an email for you. 386 00:25:19,640 --> 00:25:21,160 It can interrogate data. 387 00:25:21,160 --> 00:25:23,720 It can do tasks like adding new contacts. 388 00:25:23,720 --> 00:25:24,720 It can do a lot. 389 00:25:24,720 --> 00:25:28,520 But actually when you go into it, what do I want this assistant to do for me now? 390 00:25:28,520 --> 00:25:29,520 Is it a data analyst? 391 00:25:29,520 --> 00:25:34,080 Do I want it to uncover some lead intelligence? 392 00:25:34,080 --> 00:25:38,440 Do I want it to tell me about the best performing channel? 393 00:25:38,440 --> 00:25:44,800 Do I want it to tell me about which piece of content is having the most influence on 394 00:25:44,800 --> 00:25:47,680 revenue in the last quarter? 395 00:25:47,680 --> 00:25:52,800 You have to go into it thinking about what's the role that it's playing, but yeah, also 396 00:25:52,800 --> 00:25:54,720 knowing what the potential is within the data. 397 00:25:54,720 --> 00:25:56,840 Because ultimately that's what we're doing here. 398 00:25:56,840 --> 00:26:03,520 We're connecting a huge pool of data to this resource. 399 00:26:03,520 --> 00:26:07,800 And I dare say that most people don't know what treasure trove of data they're sitting 400 00:26:07,800 --> 00:26:10,320 on with their CRM. 401 00:26:10,320 --> 00:26:11,320 Absolutely agree. 402 00:26:11,320 --> 00:26:14,000 I think that'll be a big part of it. 403 00:26:14,000 --> 00:26:20,840 We should also mention that HubSpot released another cool beta tool at exactly the same 404 00:26:20,840 --> 00:26:26,880 time as this that's kind of been lost a little bit because of how cool Chatspot.ai is, but 405 00:26:26,880 --> 00:26:31,600 it's new content assistant tool, which I haven't been able to get on the beta yet. 406 00:26:31,600 --> 00:26:34,120 HubSpot, if you're listening, it'd be very nice to get on there. 407 00:26:34,120 --> 00:26:38,400 Maybe we'll feature that in tool of the week if you could be so kind. 408 00:26:38,400 --> 00:26:46,160 But I assume it's some sort of GPT-3 style text tool for writing blog posts, summarizing 409 00:26:46,160 --> 00:26:47,480 content, et cetera, et cetera. 410 00:26:47,480 --> 00:26:48,640 Are you on that beta? 411 00:26:48,640 --> 00:26:49,960 Have you managed to play with it? 412 00:26:49,960 --> 00:26:50,960 No, likewise. 413 00:26:50,960 --> 00:26:51,960 I saw it existed. 414 00:26:51,960 --> 00:26:54,480 I haven't gotten there yet. 415 00:26:54,480 --> 00:26:59,760 It looks like a content generation tool, much like something that you would find offline 416 00:26:59,760 --> 00:27:04,200 with the likes of Jasper or something similar, but until I've got my hands on it, couldn't 417 00:27:04,200 --> 00:27:06,200 pass any comment or judgment. 418 00:27:06,200 --> 00:27:07,880 Yeah, me too. 419 00:27:07,880 --> 00:27:08,880 Last point on that, that's really interesting. 420 00:27:08,880 --> 00:27:16,440 I was listening to the podcast with Paul Reutzer and Mike Kaput this week, and they were talking 421 00:27:16,440 --> 00:27:22,640 a little bit about, or maybe it was last week, the consolidation of tools and rapid emergence 422 00:27:22,640 --> 00:27:23,640 of other tools. 423 00:27:23,640 --> 00:27:28,440 And how do you bet as a marketer on what tech stack you're going to build? 424 00:27:28,440 --> 00:27:33,760 I wouldn't take as much as we're all offered these much better deals if we commit to a 425 00:27:33,760 --> 00:27:35,160 year to a software tool. 426 00:27:35,160 --> 00:27:41,280 I wouldn't do that for any tool at the moment, because if I'm a HubSpot customer, will I 427 00:27:41,280 --> 00:27:46,560 need Jasper or Writer or HyperWrite or any of these tools in three or six months? 428 00:27:46,560 --> 00:27:47,560 Maybe not. 429 00:27:47,560 --> 00:27:52,200 Because if I'm doing most of my sales prospecting emails or creating blog posts or whatever 430 00:27:52,200 --> 00:27:57,480 in HubSpot, I may have the writing tools I need in there to do it. 431 00:27:57,480 --> 00:28:02,320 One of the things that is part of ChatSpot is you can ask it to produce images for you 432 00:28:02,320 --> 00:28:04,000 and you can tell it what style you want them in. 433 00:28:04,000 --> 00:28:09,680 So it's almost like a Dorely-esque image generation tool as well, where it will create all of 434 00:28:09,680 --> 00:28:15,080 the images in the same style so that your blog or whatever has images that fit together 435 00:28:15,080 --> 00:28:17,440 into a coherent brand style. 436 00:28:17,440 --> 00:28:23,280 So really interesting to see how that's going to play out and what tools do we all invest 437 00:28:23,280 --> 00:28:24,280 in? 438 00:28:24,280 --> 00:28:25,680 How do we train our team on the different tools? 439 00:28:25,680 --> 00:28:29,560 Because which tool we use might change from week to week. 440 00:28:29,560 --> 00:28:34,600 And I'm starting to drift towards betting on again the big players, the Salesforce, the 441 00:28:34,600 --> 00:28:39,280 HubSpots of the world to try and bring as many of these tools inside their ecosystem 442 00:28:39,280 --> 00:28:43,240 as possible so that I won't hopefully have to buy any other tools because they'll all 443 00:28:43,240 --> 00:28:47,400 just be natively in the tools I'm using. 444 00:28:47,400 --> 00:28:54,400 To this point, or up until this point, the tools that people have been using to interface 445 00:28:54,400 --> 00:29:01,440 with the likes of GPT-3 have effectively been a polished front end. 446 00:29:01,440 --> 00:29:05,320 It's a very simple layer on top of it. 447 00:29:05,320 --> 00:29:12,000 And if you can create a front end and train some specific use cases on, you know, I'm 448 00:29:12,000 --> 00:29:17,920 going to give it, I'm going to give the model, I'm going to give GPT-3 five examples of email 449 00:29:17,920 --> 00:29:25,480 subject lines and then connect into that via the API, I've now got a subject line writer 450 00:29:25,480 --> 00:29:26,680 for my tool. 451 00:29:26,680 --> 00:29:30,680 And if you do that 40 times for 40 different use cases, you've got a tool that you can 452 00:29:30,680 --> 00:29:32,240 productize. 453 00:29:32,240 --> 00:29:33,240 It was a commodity. 454 00:29:33,240 --> 00:29:36,520 It was very easy and cheap to build. 455 00:29:36,520 --> 00:29:41,640 The real power of this, the thing that's going to be the kind of multi-billion, trillion 456 00:29:41,640 --> 00:29:49,560 dollar industry game changer is that connection to that first party data and where you've 457 00:29:49,560 --> 00:29:56,720 got AI that can, you know, it's fine tuned and embedded with your own data and companies 458 00:29:56,720 --> 00:30:02,400 that can make that process super simple. 459 00:30:02,400 --> 00:30:09,520 This week, you know, e-commerce platforms, you've got the likes of Slack integrating 460 00:30:09,520 --> 00:30:12,200 with chat GPT-3. 461 00:30:12,200 --> 00:30:16,320 There's huge amounts of internal company communications and data in Slack channels. 462 00:30:16,320 --> 00:30:20,120 If you can just say to a chat bot, oh, find me the conversation where I discuss this, 463 00:30:20,120 --> 00:30:23,920 this and this, that's a game changer. 464 00:30:23,920 --> 00:30:29,280 And that's where the power of the assistance is going to come. 465 00:30:29,280 --> 00:30:30,480 So yeah, you're absolutely right. 466 00:30:30,480 --> 00:30:37,960 I don't know why anyone would commit to any of these AI tools that aren't doing something 467 00:30:37,960 --> 00:30:38,960 differently. 468 00:30:38,960 --> 00:30:43,720 But to their credit, when I think about things like the AI copywriting tools, the ones that 469 00:30:43,720 --> 00:30:48,240 immediately spring to mind for me are the ones that are doing something different. 470 00:30:48,240 --> 00:30:55,080 JASP has got the business brand tone of voice element, so it's understanding how your company 471 00:30:55,080 --> 00:31:03,200 writes long shot as fact GPT, which can do real time copywriting or about stories published 472 00:31:03,200 --> 00:31:04,200 yesterday. 473 00:31:04,200 --> 00:31:09,200 You know, it's really powerful and very capable. 474 00:31:09,200 --> 00:31:13,840 I look at the likes of Go Charlie and what they're doing with content repurposing where 475 00:31:13,840 --> 00:31:17,360 you can create blogs from video content and audio snippets. 476 00:31:17,360 --> 00:31:24,280 And that's, they're going to be the USPs for those content led tools, I think. 477 00:31:24,280 --> 00:31:25,920 I think they absolutely are. 478 00:31:25,920 --> 00:31:29,480 I think they may have to evolve because I hear those examples you're describing and 479 00:31:29,480 --> 00:31:33,860 I think, well, our website blogger on HubSpot. 480 00:31:33,860 --> 00:31:38,040 So HubSpot knows and can quickly, the bot can quickly just go and see what our brand 481 00:31:38,040 --> 00:31:39,040 style is. 482 00:31:39,040 --> 00:31:41,800 If that's built into the system, it can write like us. 483 00:31:41,800 --> 00:31:46,420 In fact, I could probably even ask it to create a new service page for me based on existing 484 00:31:46,420 --> 00:31:50,440 service pages because they've all kind of got the same template that they're using. 485 00:31:50,440 --> 00:31:55,120 So I guess it could go build those pages, potentially populate the images in the holes 486 00:31:55,120 --> 00:31:59,840 where the images go, maybe even write some copy because it knows what copy style we use 487 00:31:59,840 --> 00:32:01,720 on those pages. 488 00:32:01,720 --> 00:32:08,160 We've got our Facebook ads and Google ads and LinkedIn ads all connected to HubSpot. 489 00:32:08,160 --> 00:32:12,740 So HubSpot has all the data on which of our ads work best. 490 00:32:12,740 --> 00:32:17,440 If it can pull the copy or creative for those ads down, which I assume it can, I don't play 491 00:32:17,440 --> 00:32:21,420 with it enough to really, to really know that off the top of my head. 492 00:32:21,420 --> 00:32:25,680 Well then it can also start to suggest and write LinkedIn ads for me based on the LinkedIn 493 00:32:25,680 --> 00:32:28,720 ads it happens to know have already worked best in the past. 494 00:32:28,720 --> 00:32:33,520 So again, like you said, that first party data lives in a lot of these tools that we're 495 00:32:33,520 --> 00:32:35,000 already using. 496 00:32:35,000 --> 00:32:39,480 And if the Jaspers and the writers of the world don't have access to that, or we don't 497 00:32:39,480 --> 00:32:44,360 give them access to that, you know, we may find that we don't end up using them. 498 00:32:44,360 --> 00:32:49,000 What you've convinced me through this Martin is that my super duper amazing AI writing 499 00:32:49,000 --> 00:32:53,360 tool that I'm going to build is probably not worth building at this point, but you did 500 00:32:53,360 --> 00:32:58,840 give me a wonderful strap line for it, which is the polished front end, which is going 501 00:32:58,840 --> 00:33:03,200 to be either the strap line for my new AI writing tool or the name of my new band, to 502 00:33:03,200 --> 00:33:07,000 be honest, because in both cases, I really, really like it. 503 00:33:07,000 --> 00:33:12,480 That is a probably a good time to segue on to our last few stories for the week. 504 00:33:12,480 --> 00:33:16,920 The first one, a quick one is about domain specific large language models. 505 00:33:16,920 --> 00:33:21,840 So when we're thinking about ChatGPT and some of these other tools, they've been trained 506 00:33:21,840 --> 00:33:27,200 on a huge amount of information and data across many, many different subject matters. 507 00:33:27,200 --> 00:33:31,600 And as such, they can sometimes struggle when we ask questions, very, very specific questions 508 00:33:31,600 --> 00:33:37,440 about very, very specific topics, where only a little bit of their training data is about 509 00:33:37,440 --> 00:33:38,940 that topic. 510 00:33:38,940 --> 00:33:43,680 So there was a research paper, I think it was actually towards the end of last year 511 00:33:43,680 --> 00:33:51,480 that then Microsoft launched a model based off the back of it called BioGPT. 512 00:33:51,480 --> 00:33:57,000 And in essence, BioGPT was a language model that's been trained on all of the scientific 513 00:33:57,000 --> 00:33:58,560 papers within PubMed. 514 00:33:58,560 --> 00:34:04,040 I don't know how many of them, this huge is like 14 million papers or something like that. 515 00:34:04,040 --> 00:34:12,800 And so in this case, it's a large language model that understands healthcare research, 516 00:34:12,800 --> 00:34:18,080 medical research, primary fundamental biology, chemistry research better than any of the 517 00:34:18,080 --> 00:34:19,860 general models. 518 00:34:19,860 --> 00:34:27,160 And so if you ask it for information on COVID or a particular drug target or a protein or 519 00:34:27,160 --> 00:34:33,320 a gene, it's going to give you a much more accurate answer than any generalized model. 520 00:34:33,320 --> 00:34:35,320 It was built this model based on GPT-2. 521 00:34:35,320 --> 00:34:40,000 Now, I don't have any experience playing with GPT-2, to be honest, but what I saw within 522 00:34:40,000 --> 00:34:45,100 BioGPT was the amount of content that it can create is limited. 523 00:34:45,100 --> 00:34:48,400 So you can start a sentence and it will finish the sentence. 524 00:34:48,400 --> 00:34:52,820 You can even ask it a question and it might give you one or two sentences as an answer. 525 00:34:52,820 --> 00:34:58,160 But beyond that, it doesn't really generate that much more data. 526 00:34:58,160 --> 00:35:01,760 So it's looking like its applications right now are going to be for doing things like 527 00:35:01,760 --> 00:35:08,400 generating descriptions of specific therapies or drug targets, as I said, maybe in designing 528 00:35:08,400 --> 00:35:17,080 slightly better clinical trial protocols, drawing relationships between factors in its 529 00:35:17,080 --> 00:35:21,360 database, so maybe drug-drug interactions like, oh, does this drug interact with this 530 00:35:21,360 --> 00:35:22,360 drug? 531 00:35:22,360 --> 00:35:23,360 Yes, no. 532 00:35:23,360 --> 00:35:26,040 Because when you're prescribing drugs, you've got to be mindful of what other drugs someone 533 00:35:26,040 --> 00:35:30,920 is on in case there's going to be a harmful reaction between the two. 534 00:35:30,920 --> 00:35:35,600 But as a marketer, it's certainly not at all yet where I could, working in the life sciences 535 00:35:35,600 --> 00:35:42,160 like I do, ask it to write really informed blog posts or articles for me because it could 536 00:35:42,160 --> 00:35:44,120 only produce a couple of sentences. 537 00:35:44,120 --> 00:35:46,640 Why is this important for marketers? 538 00:35:46,640 --> 00:35:51,960 Well, A, it shows the power of a domain-specific model, which I definitely think is there. 539 00:35:51,960 --> 00:35:59,000 The information accuracy that you get out of this versus, say, ChatGPT, there is a difference. 540 00:35:59,000 --> 00:36:01,960 This obviously will be applied in other areas. 541 00:36:01,960 --> 00:36:05,440 You can imagine the legal profession's got its own language, for example. 542 00:36:05,440 --> 00:36:11,120 I'm sure different areas of engineering and manufacturing have their own languages taught 543 00:36:11,120 --> 00:36:12,120 too. 544 00:36:12,120 --> 00:36:18,160 So the creation of niche models on information repositories specific to those niches is going 545 00:36:18,160 --> 00:36:19,600 to be valuable. 546 00:36:19,600 --> 00:36:25,680 But until they leverage the power of GPT-3 and, hey, maybe GPT-4, they're not going to 547 00:36:25,680 --> 00:36:31,720 be as broadly usable for marketers as ChatGPT and other tools is what I would say. 548 00:36:31,720 --> 00:36:36,360 What are you thinking on these domain-specific models, Martin? 549 00:36:36,360 --> 00:36:38,840 Do you think we're going to see a lot more of these? 550 00:36:38,840 --> 00:36:40,520 I think it's inevitable. 551 00:36:40,520 --> 00:36:47,120 If for no other reason, I believe that they're less resource-intensive as well. 552 00:36:47,120 --> 00:36:56,480 When you're calling something from ChatGPT or GPT-3, using that whole network, my understanding 553 00:36:56,480 --> 00:37:01,920 is that these don't draw as much resource, so the compute power for every request is 554 00:37:01,920 --> 00:37:02,920 lower. 555 00:37:02,920 --> 00:37:04,320 So I think there's going to be efficiencies. 556 00:37:04,320 --> 00:37:07,640 Anything else better for the planet? 557 00:37:07,640 --> 00:37:10,560 If everybody's using these things all day, every day. 558 00:37:10,560 --> 00:37:15,120 But yeah, I think having more tightly trained without, again, it goes back to what we were 559 00:37:15,120 --> 00:37:18,480 saying with the CRM thing. 560 00:37:18,480 --> 00:37:22,960 Where there's a cluster of data that's relevant for your space, whether it's your business 561 00:37:22,960 --> 00:37:29,840 and your first-party data or your sector, that's what you're going to want to call upon. 562 00:37:29,840 --> 00:37:36,040 It's great that ChatGPT and GPT-3 large language models have knowledge outside of that. 563 00:37:36,040 --> 00:37:39,320 Really, that's what I want to focus on. 564 00:37:39,320 --> 00:37:45,560 So having access to models that are more efficient and more tailored and more knowledgeable in 565 00:37:45,560 --> 00:37:51,040 depth is going to be the game changer for all industries, I think. 566 00:37:51,040 --> 00:37:52,640 Yeah, I agree. 567 00:37:52,640 --> 00:37:57,360 Right, let's keep the biology theme going for the last story of the week. 568 00:37:57,360 --> 00:38:01,040 Again, this is mostly because it's mega-cool rather than there being obvious applications 569 00:38:01,040 --> 00:38:03,600 for marketers today. 570 00:38:03,600 --> 00:38:09,760 But there was a pre-print research paper that was actually published in November, but for 571 00:38:09,760 --> 00:38:15,120 some reason is getting a bit of buzz this week, where AI was used to recreate images 572 00:38:15,120 --> 00:38:19,040 that people had seen by reading their brain scans. 573 00:38:19,040 --> 00:38:22,440 I'm just saying that aloud. 574 00:38:22,440 --> 00:38:23,440 As unbelievable. 575 00:38:23,440 --> 00:38:26,760 If you feel a bit queasy hearing that and you're like, sorry, what? 576 00:38:26,760 --> 00:38:30,380 I think you are probably among the majority there. 577 00:38:30,380 --> 00:38:31,840 So there was a bunch of reports this week. 578 00:38:31,840 --> 00:38:37,240 The one I read was in Science Magazine about how researchers use stable diffusion, which 579 00:38:37,240 --> 00:38:43,720 is one of the AI models for image generation, to interpret brain activity and recreate images 580 00:38:43,720 --> 00:38:44,880 that people were looking at. 581 00:38:44,880 --> 00:38:48,760 So basically they used a data set where four people looked at 10,000 images while their 582 00:38:48,760 --> 00:38:51,640 brain was being scanned using fMRI. 583 00:38:51,640 --> 00:38:54,160 And the results were scarily accurate. 584 00:38:54,160 --> 00:38:58,040 So in one example, they had someone look at a teddy bear. 585 00:38:58,040 --> 00:39:03,280 The pure model is able to generate the shapes and perspective that makes it look a bit like 586 00:39:03,280 --> 00:39:05,480 a teddy bear, but not really. 587 00:39:05,480 --> 00:39:11,240 And then when supported by a short image caption like teddy bear, the system can produce an 588 00:39:11,240 --> 00:39:17,200 image that was scarily close to the actual image the person was looking at. 589 00:39:17,200 --> 00:39:21,600 That text bit is critical though, because I think it sounds super, super, super, super 590 00:39:21,600 --> 00:39:22,600 cool and it is. 591 00:39:22,600 --> 00:39:28,760 But when you learn that the system using the MRI was able to get the approximate shapes 592 00:39:28,760 --> 00:39:35,760 and colors right, but needed that little bit of help to interpret them as a teddy bear, 593 00:39:35,760 --> 00:39:41,280 or whatever it may be, I think that's a critical part that helped it to improve much better. 594 00:39:41,280 --> 00:39:42,680 But still incredibly cool. 595 00:39:42,680 --> 00:39:45,400 And here's some stuff from that science article. 596 00:39:45,400 --> 00:39:50,880 So Iris Groen, who's a neuroscientist not involved in the work, thought that the accuracy 597 00:39:50,880 --> 00:39:57,760 of the new method was very, very impressive and that these diffusion models have an unprecedented 598 00:39:57,760 --> 00:40:02,200 ability to generate realistic images, which I think is really cool and a science that 599 00:40:02,200 --> 00:40:06,600 could create new opportunities for cognitive neuroscience research. 600 00:40:06,600 --> 00:40:10,960 One of the scientists who worked on the study suggested that further refinements to the 601 00:40:10,960 --> 00:40:15,080 technology could be used to intercept imagined thoughts and dreams. 602 00:40:15,080 --> 00:40:17,600 Just going to let that sink in for a moment. 603 00:40:17,600 --> 00:40:19,800 Or could allow scientists, this is incredible. 604 00:40:19,800 --> 00:40:21,160 I'm not laughing because it's funny. 605 00:40:21,160 --> 00:40:23,520 I'm laughing because I literally can't believe it. 606 00:40:23,520 --> 00:40:28,800 Or could allow scientists to understand how differently other animals perceive reality. 607 00:40:28,800 --> 00:40:35,760 Well, if I just slipped into the Twilight Zone, are we really here discussing this? 608 00:40:35,760 --> 00:40:42,920 That is, that sounds quite speculative to me, but the fact that they can even interpret 609 00:40:42,920 --> 00:40:50,200 those brain patterns to figure out approximate shapes and colors is pretty awesome. 610 00:40:50,200 --> 00:40:52,080 And we saw this story this week, didn't we Martin? 611 00:40:52,080 --> 00:40:53,560 And we were like, that can't be real. 612 00:40:53,560 --> 00:40:54,560 We've got to be nonsense. 613 00:40:54,560 --> 00:40:55,560 Yeah. 614 00:40:55,560 --> 00:41:01,000 We did everything we could on the web to try and find the reasons why this was junk. 615 00:41:01,000 --> 00:41:05,560 But once we started seeing more reputable providers like science picking it up and we dove into 616 00:41:05,560 --> 00:41:10,360 some subreddits and reddits on the topic and we read the paper and we're obviously not 617 00:41:10,360 --> 00:41:15,280 experts in this area, but we were really trying to find how to find the reasons why this wasn't 618 00:41:15,280 --> 00:41:16,280 so exciting. 619 00:41:16,280 --> 00:41:21,600 And the only real one we could find is you need to support that image interpretation 620 00:41:21,600 --> 00:41:25,120 with a bit of a text prompt, which I think makes it a tiny bit less exciting than it 621 00:41:25,120 --> 00:41:28,640 sounds on the surface, but still pretty cool. 622 00:41:28,640 --> 00:41:30,000 Incredibly cool. 623 00:41:30,000 --> 00:41:34,120 And we're at day one of this technology. 624 00:41:34,120 --> 00:41:35,680 It's doing this today. 625 00:41:35,680 --> 00:41:42,120 Just think about what we were saying with a GPT-3 to GPT-4, we're potentially talking 626 00:41:42,120 --> 00:41:48,080 about a model that goes from 175 billion parameters to a trillion, say. 627 00:41:48,080 --> 00:41:51,720 That's in the space of three years of development. 628 00:41:51,720 --> 00:41:59,520 Fast forward five, 10 years, the requirement to add that little bit of text saying teddy 629 00:41:59,520 --> 00:42:02,960 bear, I don't know, going to be there, do you? 630 00:42:02,960 --> 00:42:06,760 I suppose you can get the more data you can get on this and the better that you can train 631 00:42:06,760 --> 00:42:07,760 them. 632 00:42:07,760 --> 00:42:10,640 The only thing I can think of in terms of real world application is we can't all walk 633 00:42:10,640 --> 00:42:16,560 around with a miniature fMRI on our head recording what we're doing, but I'm sure they potentially 634 00:42:16,560 --> 00:42:18,400 be able to find other mechanisms to work with. 635 00:42:18,400 --> 00:42:20,200 Thanks, Silly, at this point. 636 00:42:20,200 --> 00:42:27,000 I do not need a real time live stream of my thoughts and feelings just being interpreted 637 00:42:27,000 --> 00:42:29,980 and potentially leaked to the world. 638 00:42:29,980 --> 00:42:34,080 Seeing you as I do mine, I would think that A, that would get you in a lot of trouble 639 00:42:34,080 --> 00:42:38,080 and B, it should be illegal just based on some of the things that flow around in there. 640 00:42:38,080 --> 00:42:41,960 Yeah, no, it's really going to change the dynamic of those Zoom calls. 641 00:42:41,960 --> 00:42:46,520 Right, okay, that's quite enough nonsense there. 642 00:42:46,520 --> 00:42:50,000 Let's move on to the very last part of our section this week, which is our tool of the 643 00:42:50,000 --> 00:42:51,000 week. 644 00:42:51,000 --> 00:42:54,480 This week, we are focusing on Runway.ML. 645 00:42:54,480 --> 00:42:56,480 So what is it and what can it do? 646 00:42:56,480 --> 00:43:02,720 Runway is a multimodal tool that is focused on AI supported image and video generation, 647 00:43:02,720 --> 00:43:05,200 basically, and also image and video enhancement. 648 00:43:05,200 --> 00:43:10,280 It can do a few other things like audio cleanup on a video, transcription and all that good 649 00:43:10,280 --> 00:43:11,280 stuff. 650 00:43:11,280 --> 00:43:14,440 But there's really so, so, so many tools. 651 00:43:14,440 --> 00:43:17,840 And I think what makes Runway different from a lot of the other tools that you can use 652 00:43:17,840 --> 00:43:20,640 is just how many different things you can do with it. 653 00:43:20,640 --> 00:43:22,320 So it's got trained image generators. 654 00:43:22,320 --> 00:43:23,840 We'll talk about those a bit in the moment. 655 00:43:23,840 --> 00:43:28,920 It's got a remove background tool, a text to image tool, an image to image tool, image 656 00:43:28,920 --> 00:43:31,960 expansion tools, erase and replace. 657 00:43:31,960 --> 00:43:34,760 It's got backdrop changes. 658 00:43:34,760 --> 00:43:37,200 It's got text to 3D texture. 659 00:43:37,200 --> 00:43:42,280 On the 3D, on the video side, you can use inpainting to remove people and objects from 660 00:43:42,280 --> 00:43:46,080 videos that you wish weren't in there, you weren't in there when you captured the video. 661 00:43:46,080 --> 00:43:48,240 It can do artificial bokeh effects. 662 00:43:48,240 --> 00:43:49,240 It can blur faces. 663 00:43:49,240 --> 00:43:50,920 It could be super slow motion. 664 00:43:50,920 --> 00:43:52,720 It can clean your audio. 665 00:43:52,720 --> 00:43:55,840 Loads of stuff. 666 00:43:55,840 --> 00:44:02,080 When you think about the fact that it is only $12 a month per user, to be able to access 667 00:44:02,080 --> 00:44:05,520 all of those tools is kind of spectacular. 668 00:44:05,520 --> 00:44:08,240 And there are many of them, I haven't played with all of them. 669 00:44:08,240 --> 00:44:09,680 There are just too many. 670 00:44:09,680 --> 00:44:14,440 But many of them are really, really powerful and work nearly as advertised in almost all 671 00:44:14,440 --> 00:44:16,440 the cases I've played with so far. 672 00:44:16,440 --> 00:44:21,920 And I've seen some examples where people really pushed it to the limit removing, say, subjects 673 00:44:21,920 --> 00:44:25,360 from videos where they're panning the video across and they actually managed to pretty 674 00:44:25,360 --> 00:44:29,880 much remove a whole subject from it, which is again, kind of crazy when you think about 675 00:44:29,880 --> 00:44:30,880 it. 676 00:44:30,880 --> 00:44:31,880 So that's really cool. 677 00:44:31,880 --> 00:44:38,040 A couple of the things that I think are especially special about RunYML that make it, I mean, 678 00:44:38,040 --> 00:44:40,280 it's worth checking out just for all that stuff. 679 00:44:40,280 --> 00:44:45,320 But some of the stuff that's really especially cool is the fact that you can train the models 680 00:44:45,320 --> 00:44:46,320 on your own images. 681 00:44:46,320 --> 00:44:47,920 So you've got the personal portraits. 682 00:44:47,920 --> 00:44:48,920 I know there's a... 683 00:44:48,920 --> 00:44:52,960 Do you remember the name of the app that was doing the rounds a month or so ago where everybody 684 00:44:52,960 --> 00:44:57,280 got obsessed with making themselves look like famous characters from different stuff? 685 00:44:57,280 --> 00:45:00,880 No, it was trending for about a week, wasn't it? 686 00:45:00,880 --> 00:45:02,200 Everyone was going mad for it. 687 00:45:02,200 --> 00:45:03,520 But no, I don't remember the name. 688 00:45:03,520 --> 00:45:05,640 So you can basically do the same thing in Runway. 689 00:45:05,640 --> 00:45:10,200 So if you wanted to see yourself as Princess Leia from Star Wars or Jon Snow from Game 690 00:45:10,200 --> 00:45:14,520 of Thrones or maybe even the Hulk from The Avengers you now can do, there's an animal 691 00:45:14,520 --> 00:45:20,120 version of it so you can chuck in pictures of your pets and you can do pretty much the 692 00:45:20,120 --> 00:45:22,160 same thing with animal pictures. 693 00:45:22,160 --> 00:45:25,460 And then there's a model where you can basically train it on any set of images that you have. 694 00:45:25,460 --> 00:45:30,280 So you could imagine a time where you might give it your icon repository as a brand or 695 00:45:30,280 --> 00:45:35,040 your brand image set and then use that to generate image from text prompts where you 696 00:45:35,040 --> 00:45:38,480 had much more confidence that the images that you got would be in your brand colors and 697 00:45:38,480 --> 00:45:40,520 that type of stuff. 698 00:45:40,520 --> 00:45:42,000 That's pretty, pretty cool. 699 00:45:42,000 --> 00:45:46,240 I also love the Infinite Image and the Image Expander tools. 700 00:45:46,240 --> 00:45:52,240 A lot of the tools you can generate using AI at the moment are square, but most of the 701 00:45:52,240 --> 00:45:54,880 images that I need are not square. 702 00:45:54,880 --> 00:45:59,800 I need more widescreen style images for websites or for social media images or for banners 703 00:45:59,800 --> 00:46:01,960 on blog posts and all that type of stuff. 704 00:46:01,960 --> 00:46:07,040 So that makes it very easy to generate the square image maybe using another tool if you 705 00:46:07,040 --> 00:46:09,120 get different types of results or different types of tools. 706 00:46:09,120 --> 00:46:13,600 But you can quickly then expand it or change the aspect ratio of the image to fit what 707 00:46:13,600 --> 00:46:16,040 you're actually trying to do. 708 00:46:16,040 --> 00:46:19,720 As I've mentioned, the inpainting to remove objects from people from videos looks really 709 00:46:19,720 --> 00:46:22,520 powerful and surprisingly accurate. 710 00:46:22,520 --> 00:46:24,080 That's worth playing with. 711 00:46:24,080 --> 00:46:29,280 And Runway have just released their Gen1 tool for creating videos. 712 00:46:29,280 --> 00:46:33,720 Now I think this is quite early in terms of the types of video production that we can 713 00:46:33,720 --> 00:46:35,880 expect from AI augmented tools. 714 00:46:35,880 --> 00:46:40,840 But I did see one example on Twitter where someone had basically filmed some scenes on 715 00:46:40,840 --> 00:46:49,120 their phone in their house and then fed Runway's Gen1 tool some image stylization and that 716 00:46:49,120 --> 00:46:55,800 raw video from their phone and created like a stylized cartoon of someone as Indiana Jones. 717 00:46:55,800 --> 00:47:00,480 Dig it out on the socials a bit or by Googling it if you can, because it's actually really 718 00:47:00,480 --> 00:47:06,220 kind of spectacular and it gives the feeling that it could probably have been produced 719 00:47:06,220 --> 00:47:07,520 in a couple of hours max. 720 00:47:07,520 --> 00:47:13,040 In fact, I suspect the bit that took the longest was actually filming the sort of raw scenes 721 00:47:13,040 --> 00:47:14,600 in the house on the phone. 722 00:47:14,600 --> 00:47:20,240 So if you think about how you could very quickly knock up storyboard and creative ideas for 723 00:47:20,240 --> 00:47:26,360 your next video ads to support your campaign, or dare I say it, for some of the campaigns 724 00:47:26,360 --> 00:47:30,440 that you might run, the video that you might even get from Gen1 might be good enough to 725 00:47:30,440 --> 00:47:34,440 run on as part of a Facebook ad campaign or something like that. 726 00:47:34,440 --> 00:47:38,360 That is really rather cool. 727 00:47:38,360 --> 00:47:39,820 It's got loads of stuff it can do. 728 00:47:39,820 --> 00:47:42,200 Having played with it, a lot of the stuff works really well. 729 00:47:42,200 --> 00:47:47,160 We decided to use Runway as the tool we were going to trial for editing the first video 730 00:47:47,160 --> 00:47:48,160 version of the podcast. 731 00:47:48,160 --> 00:47:53,840 So for last week's episode one, the interface is very intuitive and that's great. 732 00:47:53,840 --> 00:47:55,000 So it's easy to use. 733 00:47:55,000 --> 00:47:59,000 But the biggest drawback that I found with it was it was too slow. 734 00:47:59,000 --> 00:48:02,560 When I was trying to edit, you know, the podcast about 45 minutes long, when I was trying to 735 00:48:02,560 --> 00:48:07,080 edit the automatically generated captions, it did a brilliant job of getting those captions 736 00:48:07,080 --> 00:48:09,320 pretty much perfect. 737 00:48:09,320 --> 00:48:15,360 But there were probably around 40 to 50 small areas, usually the names of people or tools 738 00:48:15,360 --> 00:48:17,320 or technologies. 739 00:48:17,320 --> 00:48:21,720 Editing that, it would take like five seconds to make each edit because it would update 740 00:48:21,720 --> 00:48:22,720 so slowly. 741 00:48:22,720 --> 00:48:25,360 Oh no, I couldn't handle that. 742 00:48:25,360 --> 00:48:28,000 So that was not quite what I was looking for. 743 00:48:28,000 --> 00:48:36,120 The export function also took hours to produce a 45 minute 720p video where if I think of 744 00:48:36,120 --> 00:48:41,600 I was using a piece of technology that was on board on my PC like Adobe Premiere Pro 745 00:48:41,600 --> 00:48:46,480 or something like it would be 10%, 20% of the time of that. 746 00:48:46,480 --> 00:48:49,880 So if you're in a rush and you really need to produce video quickly, I'm not sure. 747 00:48:49,880 --> 00:48:54,360 At least with my current setup, maybe it was my browser, maybe it was my processor, I don't 748 00:48:54,360 --> 00:48:57,300 know, a little bit too slow for me on that. 749 00:48:57,300 --> 00:49:01,920 So huge amounts of potential, especially on the video and probably short, especially on 750 00:49:01,920 --> 00:49:08,280 the images, I should say, and the short video side in terms of editing the long form video, 751 00:49:08,280 --> 00:49:11,440 not so good quite yet. 752 00:49:11,440 --> 00:49:14,800 So at the moment, all of this is done in the browser? 753 00:49:14,800 --> 00:49:16,360 Yes, all in the browser. 754 00:49:16,360 --> 00:49:19,280 And that probably is the main limitation, I would have thought. 755 00:49:19,280 --> 00:49:24,880 Even Descript has, so Descript is another tool that you can use for really easy AI supported 756 00:49:24,880 --> 00:49:29,040 editing, that actually has a tool that you download. 757 00:49:29,040 --> 00:49:31,560 So maybe that's what's going to be needed. 758 00:49:31,560 --> 00:49:35,840 But yeah, I think if you're looking for a powerful one in one image and video creation 759 00:49:35,840 --> 00:49:40,200 tool, you should definitely check out Runway because I think it's just got so many capabilities 760 00:49:40,200 --> 00:49:42,280 for a ridiculously low price point. 761 00:49:42,280 --> 00:49:48,000 Yeah, 12 pound or $12 a month, there's nothing for all of that. 762 00:49:48,000 --> 00:49:55,000 I've seen some tools like, for instance, the kind of portrait face replicator thing, training 763 00:49:55,000 --> 00:49:56,600 a model on your own face. 764 00:49:56,600 --> 00:50:02,280 I've seen tools going for $10 and $12 per training set. 765 00:50:02,280 --> 00:50:06,120 So just to train your face once would cost you that. 766 00:50:06,120 --> 00:50:10,720 So getting all of that bundled in is quite the deal. 767 00:50:10,720 --> 00:50:14,480 Yes, I do think that every time you train on an image set, you have to pay a small fee 768 00:50:14,480 --> 00:50:16,480 in addition, if I'm honest. 769 00:50:16,480 --> 00:50:20,640 But yeah, the ongoing access is relatively cheap. 770 00:50:20,640 --> 00:50:21,640 Cool, cool. 771 00:50:21,640 --> 00:50:24,160 Well, that's our tool of the week. 772 00:50:24,160 --> 00:50:27,680 And that brings us to a close for this week's podcast. 773 00:50:27,680 --> 00:50:29,840 Always a pleasure chatting with you, Martin. 774 00:50:29,840 --> 00:50:31,200 Loads of great stuff this week. 775 00:50:31,200 --> 00:50:32,520 Covered some ground. 776 00:50:32,520 --> 00:50:33,520 Yeah, definitely. 777 00:50:33,520 --> 00:50:35,440 We absolutely have done. 778 00:50:35,440 --> 00:50:39,840 Thanks so much to those of you who have listened today and got to the end. 779 00:50:39,840 --> 00:50:41,320 High fives to you. 780 00:50:41,320 --> 00:50:45,640 And we'll look forward to having a good chat next week and hopefully having you join us. 781 00:50:45,640 --> 00:50:47,680 Have a great week, Martin. 782 00:50:47,680 --> 00:50:48,680 You too. 783 00:50:48,680 --> 00:50:49,680 See you soon. 784 00:50:49,680 --> 00:50:50,680 Cheers. 785 00:50:50,680 --> 00:50:51,680 Bye. 786 00:50:51,680 --> 00:50:58,100 Thank you for listening to Artificially Intelligent Marketing. 787 00:50:58,100 --> 00:51:04,160 To stay on top of the latest trends, tips and tools in the world of marketing AI, be 788 00:51:04,160 --> 00:51:05,920 sure to subscribe. 789 00:51:05,920 --> 00:51:17,200 We look forward to seeing you again next week.