1 00:00:00,000 --> 00:00:09,960 Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the 2 00:00:09,960 --> 00:00:15,700 latest trends, tips, and tools in the world of marketing AI, helping you get the best 3 00:00:15,700 --> 00:00:18,540 results from your marketing efforts. 4 00:00:18,540 --> 00:00:22,000 Now let's join our hosts, Paul Avery and Martin Broadhurst. 5 00:00:22,000 --> 00:00:23,600 Hello, everyone. 6 00:00:23,600 --> 00:00:28,360 It's Paul Avery here, which means you're listening to Artificially Intelligent Marketing, 7 00:00:28,360 --> 00:00:33,480 two extremely exciting bits of news today. 8 00:00:33,480 --> 00:00:36,760 And in priority order, the first one is this is episode 21. 9 00:00:36,760 --> 00:00:37,760 How exciting. 10 00:00:37,760 --> 00:00:38,760 We're very excited. 11 00:00:38,760 --> 00:00:41,120 And then the other one is it's Martin Broadhurst. 12 00:00:41,120 --> 00:00:42,120 Birthday today. 13 00:00:42,120 --> 00:00:43,120 Happy birthday, Martin. 14 00:00:43,120 --> 00:00:45,400 You don't look a day over 25. 15 00:00:45,400 --> 00:00:52,520 It's very kind of you to say I feel immeasurably older than 25. 16 00:00:52,520 --> 00:00:58,080 The joints, all of them, every one of them, even my knuckles. 17 00:00:58,080 --> 00:01:04,760 The brain is firing and it's been firing hard all week because of all the wonderful AI stories 18 00:01:04,760 --> 00:01:06,720 that we've been pinging back and forth. 19 00:01:06,720 --> 00:01:07,720 Yes. 20 00:01:07,720 --> 00:01:08,720 Yes, very much so. 21 00:01:08,720 --> 00:01:12,320 Well, thank you for wishing me the happy birthday. 22 00:01:12,320 --> 00:01:14,360 I'm going to have a lovely day. 23 00:01:14,360 --> 00:01:17,160 Not really doing any work except for recording this podcast. 24 00:01:17,160 --> 00:01:20,760 And let's be honest, this doesn't feel like work, does it, Paul? 25 00:01:20,760 --> 00:01:26,120 This is just a good friendly chat informing our good listeners. 26 00:01:26,120 --> 00:01:28,080 But no, it's going to be a good day. 27 00:01:28,080 --> 00:01:30,600 We've got some good stories to cover today. 28 00:01:30,600 --> 00:01:34,400 I've had some good news this week, which was welcome. 29 00:01:34,400 --> 00:01:39,360 I'm now on the editorial board of the Applied Marketing Analytics Journal. 30 00:01:39,360 --> 00:01:44,760 So that's going to give me interesting access to good research papers that we can bring 31 00:01:44,760 --> 00:01:47,120 into the podcast and inform our listeners. 32 00:01:47,120 --> 00:01:52,480 Without doubt, that's going to be some useful content for people coming forward as well. 33 00:01:52,480 --> 00:01:53,480 Lovely stuff. 34 00:01:53,480 --> 00:01:59,080 We've been planning some training for some of our clients coming through that are interested 35 00:01:59,080 --> 00:02:05,840 in finding out how they can leverage the power of AI to inspire their teams to be more creative, 36 00:02:05,840 --> 00:02:07,400 be more efficient, be more effective. 37 00:02:07,400 --> 00:02:09,480 So that's been quite fun as well, hasn't it? 38 00:02:09,480 --> 00:02:10,480 It has. 39 00:02:10,480 --> 00:02:11,480 Yeah. 40 00:02:11,480 --> 00:02:14,680 I'm looking forward to rolling that out with more clients in the coming months. 41 00:02:14,680 --> 00:02:15,680 Absolutely. 42 00:02:15,680 --> 00:02:16,680 Right. 43 00:02:16,680 --> 00:02:20,480 With that, let's jump straight into these stories this week then. 44 00:02:20,480 --> 00:02:29,960 I'm quite heavy on the synthetic voice, synthetic humans this week, starting with PlayHT 2.0, 45 00:02:29,960 --> 00:02:32,400 which is a new conversational AI voice model. 46 00:02:32,400 --> 00:02:38,200 Now you may remember the name PlayHT because when you listen to the intro of the Artificial 47 00:02:38,200 --> 00:02:43,880 Intelligent Marketing podcast, that voice that you hear was synthesized using the first 48 00:02:43,880 --> 00:02:45,640 version of their tool. 49 00:02:45,640 --> 00:02:49,600 But what's really exciting is they have now unveiled their new model. 50 00:02:49,600 --> 00:02:54,920 And it's a real major leap forward in conversational speech generation, at least according to the 51 00:02:54,920 --> 00:02:55,920 company. 52 00:02:55,920 --> 00:03:01,360 It's been trained on over 1 million hours of diverse speech data, and it can mimic voices, 53 00:03:01,360 --> 00:03:08,480 accents, and languages with really amazing realism, even after hearing just a few seconds 54 00:03:08,480 --> 00:03:09,480 of sample audio. 55 00:03:09,480 --> 00:03:15,360 So in essence, what this promises to be able to do is, I think it's six seconds that the 56 00:03:15,360 --> 00:03:20,840 press release says, six seconds of training on your voice and then able to mimic it. 57 00:03:20,840 --> 00:03:24,740 And this includes some examples, which of course we can't really easily test, of someone 58 00:03:24,740 --> 00:03:29,880 speaking in a language such as Spanish and then having that voice used to generate English 59 00:03:29,880 --> 00:03:37,600 but maintaining nuances of that person's Spanish accent, which was quite an incredible demo. 60 00:03:37,600 --> 00:03:41,680 It's also, if that wasn't exciting enough, I'm not thinking it is pretty exciting, it 61 00:03:41,680 --> 00:03:46,560 also introduces the ability to control emotional expression when synthesizing the speech. 62 00:03:46,560 --> 00:03:51,920 So now as a user, you can directly use the model speaking tones conveying happiness, 63 00:03:51,920 --> 00:03:52,920 sadness, fear. 64 00:03:52,920 --> 00:03:58,740 What you basically do is in your prompt, you just say the text you want the synthetic engine 65 00:03:58,740 --> 00:04:03,240 to say, and then you just say scared, and it will sound scared. 66 00:04:03,240 --> 00:04:08,400 And again, all of the demos of this, of course, always cherry pick the best examples. 67 00:04:08,400 --> 00:04:13,160 But if it has the power that it looks like it does in the demos, then goodness me, it's 68 00:04:13,160 --> 00:04:15,960 going to be extremely powerful. 69 00:04:15,960 --> 00:04:20,640 So one of the things that I think has been able to help improve this is compared to the 70 00:04:20,640 --> 00:04:25,600 previous model, which was kind of limited in terms of how it had been built and its 71 00:04:25,600 --> 00:04:30,320 ability to really generate long form synthetic speech outputs. 72 00:04:30,320 --> 00:04:35,800 But thanks to Play H2 2.0, you might have to look at the name that honestly, because 73 00:04:35,800 --> 00:04:39,760 that's almost like trying to say artificially intelligent marketing. 74 00:04:39,760 --> 00:04:44,040 It's got far more parameters, 10 times as many parameters in the model. 75 00:04:44,040 --> 00:04:49,640 And because of how it's been built, it promises real time low latency speech generation, which 76 00:04:49,640 --> 00:04:57,600 again, is a game changer for a lot of applications like customer service and just general conversational 77 00:04:57,600 --> 00:05:03,040 abilities with an engine that you needed to be able to produce synthetic speech. 78 00:05:03,040 --> 00:05:07,840 So yeah, expect to see this popping up in places like virtual assistants, customer service, 79 00:05:07,840 --> 00:05:09,720 even audiobook generation. 80 00:05:09,720 --> 00:05:15,420 There's, I think, probably applications for sales folks out there in terms of creating 81 00:05:15,420 --> 00:05:20,120 customized voice messages and maybe video, which we'll come on to in a bit moment with 82 00:05:20,120 --> 00:05:22,040 a story from Martin. 83 00:05:22,040 --> 00:05:26,540 For marketers, I think it's just going to make it even faster and easier for us to generate 84 00:05:26,540 --> 00:05:34,400 things like podcasts, videos, ads with custom voices, making it easier to give a play my 85 00:05:34,400 --> 00:05:39,080 content button so that you can have people listen to your blog posts and stuff on the 86 00:05:39,080 --> 00:05:40,080 go if that's easier. 87 00:05:40,080 --> 00:05:45,220 And rather than it sounding super robotic, actually sounding like studio quality voice 88 00:05:45,220 --> 00:05:48,400 over work by a human. 89 00:05:48,400 --> 00:05:54,280 In terms of how this technology is really emerging and developing very quickly, Play 90 00:05:54,280 --> 00:05:59,000 HT 1.0 was released eight months ago. 91 00:05:59,000 --> 00:06:03,580 Now we have 2.0 and it is a real leap or at least from the examples looks like a real 92 00:06:03,580 --> 00:06:04,580 leap forward. 93 00:06:04,580 --> 00:06:08,480 So the model at the moment is in closed beta, so you can't go play with this yet, but it 94 00:06:08,480 --> 00:06:14,760 will be made accessible through the company's API and their online sort of Play HT Studio, 95 00:06:14,760 --> 00:06:17,560 which is what we use to generate the intro to this podcast. 96 00:06:17,560 --> 00:06:20,040 Did you have a look at this, Martin? 97 00:06:20,040 --> 00:06:26,320 Yeah, well, I'm thinking about the training and the six seconds worth of training data 98 00:06:26,320 --> 00:06:29,920 required, which is obviously incredibly short amount of time. 99 00:06:29,920 --> 00:06:32,280 So well done them for getting that down. 100 00:06:32,280 --> 00:06:38,200 But as you were talking about the idea of being able to do the emotional nuances, I 101 00:06:38,200 --> 00:06:42,200 was just thinking, do we need six seconds of someone sounding scared to get that into 102 00:06:42,200 --> 00:06:47,360 the model, get someone with a fear of spiders and just throw a tarantula on them and record 103 00:06:47,360 --> 00:06:49,680 that and now we've got their scared model. 104 00:06:49,680 --> 00:06:54,240 I imagine that we don't have to do that though, do we? 105 00:06:54,240 --> 00:06:55,840 That would be good though. 106 00:06:55,840 --> 00:07:00,520 Could you imagine like the quick brown fox jumps over the whatever, whatever, but you've 107 00:07:00,520 --> 00:07:01,640 got to do it. 108 00:07:01,640 --> 00:07:05,720 You have to give the full range of performance, right? 109 00:07:05,720 --> 00:07:06,720 Okay. 110 00:07:06,720 --> 00:07:11,200 If you want to your voice in Denzel Washington style, you got to give the quick brown box 111 00:07:11,200 --> 00:07:14,680 like you've got to be able to, you just move here. 112 00:07:14,680 --> 00:07:19,200 Yeah, you've got to capture those six seconds of if you want it to sound sad, someone's 113 00:07:19,200 --> 00:07:21,720 got to come in and tell you that your dog's just died. 114 00:07:21,720 --> 00:07:22,720 Yeah. 115 00:07:22,720 --> 00:07:27,240 No, the way I understand it is they must have been able to develop in the model, the ability 116 00:07:27,240 --> 00:07:32,640 to sort of understand the sonic signature of someone's voice in that short timeframe 117 00:07:32,640 --> 00:07:34,280 to then be able to reproduce it. 118 00:07:34,280 --> 00:07:38,640 But I think we're going to know a bit more about this when we get our hands on playing 119 00:07:38,640 --> 00:07:42,080 with it, but it sounds pretty cool. 120 00:07:42,080 --> 00:07:44,240 It's not the only game in town though, is it Martin? 121 00:07:44,240 --> 00:07:45,360 What else do we see this with? 122 00:07:45,360 --> 00:07:46,360 It's really not. 123 00:07:46,360 --> 00:07:52,800 Okay, so Hey Jen have unveiled their version two as well. 124 00:07:52,800 --> 00:07:59,160 And theirs goes beyond voice synthesis into full AI generated avatar clones, so video 125 00:07:59,160 --> 00:08:00,640 basically. 126 00:08:00,640 --> 00:08:07,920 The results are incredibly realistic with people questioning on the demos that were 127 00:08:07,920 --> 00:08:14,320 put out on Twitter this week, whether the avatars were actual humans. 128 00:08:14,320 --> 00:08:18,800 I can see why people are questioning it, I think both me and you said we watched the 129 00:08:18,800 --> 00:08:19,800 video. 130 00:08:19,800 --> 00:08:23,960 So on the one that was published on the founders Twitter account, he had two videos showing 131 00:08:23,960 --> 00:08:30,480 side by side and we both watched it thinking that one of them was real and one of them 132 00:08:30,480 --> 00:08:31,480 wasn't. 133 00:08:31,480 --> 00:08:33,480 One of them was AI generated. 134 00:08:33,480 --> 00:08:35,720 It turns out they were both AI generated. 135 00:08:35,720 --> 00:08:41,440 It was a bit mind blowing as we were both trying to find out, we were trying to figure 136 00:08:41,440 --> 00:08:44,320 out the real one, but neither of them were. 137 00:08:44,320 --> 00:08:50,920 That for me is quite a good Turing style test of like, when your brain's doing that, it 138 00:08:50,920 --> 00:08:53,960 basically is believing both of them are real and you're looking for the nuances. 139 00:08:53,960 --> 00:08:54,960 Yeah. 140 00:08:54,960 --> 00:09:00,160 It was a bit of, can they do it for everybody or did they have to like spend a billion hours 141 00:09:00,160 --> 00:09:01,880 of compute power just for the founder? 142 00:09:01,880 --> 00:09:03,080 That's the question for me. 143 00:09:03,080 --> 00:09:07,640 Yeah, I think that's the big piece of the puzzle. 144 00:09:07,640 --> 00:09:13,880 Like if everybody's sat recording video of themselves in, you know, badly the officers 145 00:09:13,880 --> 00:09:18,920 on their office webcam, is that going to generate realistic looking avatars? 146 00:09:18,920 --> 00:09:19,920 We shall see. 147 00:09:19,920 --> 00:09:26,400 But they say that in just two minutes, the founders voice and accent were cloned and 148 00:09:26,400 --> 00:09:29,560 they were obviously cloned to a T so they've done a good job with that. 149 00:09:29,560 --> 00:09:34,320 So not quite the six seconds that we heard about a moment ago, but you know, two minutes 150 00:09:34,320 --> 00:09:36,840 of video footage is not a lot. 151 00:09:36,840 --> 00:09:41,240 That video, I would say would pass for most people. 152 00:09:41,240 --> 00:09:47,640 In fact, I reckon 95% of people would believe those video snippets. 153 00:09:47,640 --> 00:09:53,680 So this synthetic media generation audio video suggests that we're about to see a full explosion 154 00:09:53,680 --> 00:09:57,920 of this kind of content in all sorts of different media landscapes. 155 00:09:57,920 --> 00:10:07,640 Like you spoke about eBooks or audio books and podcasts, video streams, you name it. 156 00:10:07,640 --> 00:10:12,640 And I think the use cases are myriad. 157 00:10:12,640 --> 00:10:18,200 I was thinking about specifically for businesses producing, I don't know, let's say webinar 158 00:10:18,200 --> 00:10:21,040 content, but you're producing some webinar content. 159 00:10:21,040 --> 00:10:23,240 You want to promote it on social media. 160 00:10:23,240 --> 00:10:27,720 A lot of the time you're probably going to stick together a little graphic and post it 161 00:10:27,720 --> 00:10:29,840 on LinkedIn and say, sign up for this. 162 00:10:29,840 --> 00:10:33,000 But we know that actually if you were to do a little video snippet, that's going to be 163 00:10:33,000 --> 00:10:34,600 a bit more engaging. 164 00:10:34,600 --> 00:10:38,960 But to make that engaging little video snippet for your webinar, you're going to have to 165 00:10:38,960 --> 00:10:43,180 write yourself a little script, record that, and then you're either going to record it 166 00:10:43,180 --> 00:10:49,440 as we are now on a webcam, you know, not the best quality conditions in terms of you might 167 00:10:49,440 --> 00:10:54,200 have to record it several times because you cock up the script a couple of times or the 168 00:10:54,200 --> 00:10:57,240 audio is not quite right or this, that or the other. 169 00:10:57,240 --> 00:11:01,960 Well now with a synthetic product, you know, you can make a one minute video and it's right 170 00:11:01,960 --> 00:11:03,200 the first time every time. 171 00:11:03,200 --> 00:11:10,480 It looks like it's recorded on studio quality equipment and you can put it out on social 172 00:11:10,480 --> 00:11:21,000 media and that kind of short, the cost to producing short content has dramatically reduced 173 00:11:21,000 --> 00:11:22,200 or is going to be dramatically reduced. 174 00:11:22,200 --> 00:11:24,520 And we're just going to see more of it. 175 00:11:24,520 --> 00:11:27,400 And I think for marketers, that's going to be a massive boom. 176 00:11:27,400 --> 00:11:28,640 I agree. 177 00:11:28,640 --> 00:11:32,800 But I think, you know, with great power comes great responsibility. 178 00:11:32,800 --> 00:11:38,000 And we've talked about it on the podcast, I had a conversation on the LinkedIn's a month 179 00:11:38,000 --> 00:11:46,520 or so ago about the deluge of crap content that is only ever a heartbeat of way with 180 00:11:46,520 --> 00:11:49,720 the explosion of capabilities of these tools. 181 00:11:49,720 --> 00:11:53,620 So I think it's great that we can produce it easily, but when there's low cost and low 182 00:11:53,620 --> 00:11:59,040 effort to producing content, I think you can often see a low bit of low imagination and 183 00:11:59,040 --> 00:12:00,760 churn like just get out. 184 00:12:00,760 --> 00:12:07,640 And I just, I worry that our ability as marketers to saturate channels, like people always say, 185 00:12:07,640 --> 00:12:10,760 enjoy a particular channel until markets get hold of it and ruin it. 186 00:12:10,760 --> 00:12:11,760 Right. 187 00:12:11,760 --> 00:12:12,760 And I think there's something to that. 188 00:12:12,760 --> 00:12:17,040 I think we've done that to a number of social channels and this just democratizes the ability 189 00:12:17,040 --> 00:12:21,040 to do that because a certain amount of money or skill that you would have needed to produce 190 00:12:21,040 --> 00:12:23,520 this type of content is going to be vastly reduced. 191 00:12:23,520 --> 00:12:28,440 The other thing that comes to mind is on the sort of sales prospecting side, which is maybe, 192 00:12:28,440 --> 00:12:33,440 you know, touches on marketing's role a bit in terms of supporting with content creation. 193 00:12:33,440 --> 00:12:38,560 But I get a lot more outbound crappy emails than I used to. 194 00:12:38,560 --> 00:12:43,400 I'm pretty sure they're all using the same software or the same agency because even the 195 00:12:43,400 --> 00:12:44,800 email templates are the same. 196 00:12:44,800 --> 00:12:46,360 It's like, oh, so dull. 197 00:12:46,360 --> 00:12:48,720 Just leave me alone. 198 00:12:48,720 --> 00:12:53,560 One great way to get around that was to create personalized videos for your prospects, right? 199 00:12:53,560 --> 00:12:58,320 That showed you really went to the effort of creating something just for them. 200 00:12:58,320 --> 00:13:00,280 People tried to find ways to gain that, right? 201 00:13:00,280 --> 00:13:03,620 By having a thumbnail with a person's name written on it, but then the actual video was 202 00:13:03,620 --> 00:13:05,960 like not customized for them at all. 203 00:13:05,960 --> 00:13:10,920 Well with tech like this, sales reps are going to be able to do that at scale. 204 00:13:10,920 --> 00:13:16,060 And again, without the necessary effort put in to make sure that the content is valuable 205 00:13:16,060 --> 00:13:18,700 for the recipient, it's not going to work. 206 00:13:18,700 --> 00:13:23,560 Like it will work for like four to eight weeks or whatever the timeline it takes for us as 207 00:13:23,560 --> 00:13:29,360 humans to spot that this is just us being fooled into thinking people think we're more 208 00:13:29,360 --> 00:13:31,200 important to them than we really are. 209 00:13:31,200 --> 00:13:34,040 So it could be interesting to see how that plays out. 210 00:13:34,040 --> 00:13:39,680 And I think for both our first two news stories, the suggestion to marketers would be just 211 00:13:39,680 --> 00:13:45,560 because you can produce a lot of content at scale, don't really use it. 212 00:13:45,560 --> 00:13:46,560 Just don't. 213 00:13:46,560 --> 00:13:47,560 Yeah, just don't. 214 00:13:47,560 --> 00:13:49,040 Just use it to produce good content. 215 00:13:49,040 --> 00:13:54,680 If you've got great ideas and you can add value to your audience and actually scale, 216 00:13:54,680 --> 00:13:57,640 opens up new avenues for you to generate value. 217 00:13:57,640 --> 00:13:58,640 Brilliant. 218 00:13:58,640 --> 00:14:03,040 If it's just churn and burn, goodness me, marketers are going to make the world a pretty 219 00:14:03,040 --> 00:14:08,240 dreadful place for us scrolling through LinkedIn, looking for one half decent piece of content 220 00:14:08,240 --> 00:14:10,800 that wasn't just churned out without any thought. 221 00:14:10,800 --> 00:14:13,120 Just because you can doesn't mean you should. 222 00:14:13,120 --> 00:14:14,720 I agree with that. 223 00:14:14,720 --> 00:14:15,720 Right. 224 00:14:15,720 --> 00:14:21,400 So with that synthetic peeps and voice is done, let's move on to chat spot for our new 225 00:14:21,400 --> 00:14:22,400 story. 226 00:14:22,400 --> 00:14:29,240 So users of HubSpot may be aware of chat spot, which is HubSpot sort of large language model 227 00:14:29,240 --> 00:14:33,720 chat botty type tool that we've been following on the podcast over the last couple of months 228 00:14:33,720 --> 00:14:34,720 since it was launched. 229 00:14:34,720 --> 00:14:38,720 I think it's fair to say it launched with some great ideas, but you couldn't really 230 00:14:38,720 --> 00:14:40,920 do much stuff with it. 231 00:14:40,920 --> 00:14:47,240 Well, we're very happy to report that it is starting to really emerge and mature as a 232 00:14:47,240 --> 00:14:50,040 tool that you could actually consider using. 233 00:14:50,040 --> 00:14:53,760 So with this tool, it's had a recent makeover this week. 234 00:14:53,760 --> 00:14:58,120 So it's got a bunch of new features and has different sort of layout and usability to 235 00:14:58,120 --> 00:14:59,120 it. 236 00:14:59,120 --> 00:15:02,840 One of the features really, really cool is the ability to summarize CRM records. 237 00:15:02,840 --> 00:15:08,000 So with a simple command, chat spot can now provide an AI generated summary of a contact 238 00:15:08,000 --> 00:15:13,200 or company's record and not just the stuff it scrapes off the web, but its ability to 239 00:15:13,200 --> 00:15:19,400 actually summarize the notes that you as marketers and sales folks are taking in the system. 240 00:15:19,400 --> 00:15:26,240 So the ability, for example, to figure out what web pages are needed, visited before, 241 00:15:26,240 --> 00:15:27,960 you know, before you have your prospecting call. 242 00:15:27,960 --> 00:15:28,960 So that's pretty awesome. 243 00:15:28,960 --> 00:15:33,160 The interface itself has seen quite a few improvements, going to be a lot easier to 244 00:15:33,160 --> 00:15:37,480 use and things like templates and chat have been separated just to make it a bit more 245 00:15:37,480 --> 00:15:39,640 of an easy experience. 246 00:15:39,640 --> 00:15:43,080 And the tool can also now summarize YouTube videos, which is interesting. 247 00:15:43,080 --> 00:15:47,360 So you can drop a YouTube link into chat spot and get a summary of the video content, which 248 00:15:47,360 --> 00:15:49,000 is a fairly powerful application. 249 00:15:49,000 --> 00:15:52,680 I haven't had a real play with yet, but I want to dig into, have you had a play with 250 00:15:52,680 --> 00:15:53,680 that yet? 251 00:15:53,680 --> 00:15:55,280 Yeah, I tried it on a couple of videos. 252 00:15:55,280 --> 00:15:57,080 They seem to do a good job. 253 00:15:57,080 --> 00:16:03,320 All of these kind of systems offer these summarizations are basically just taking the transcript from 254 00:16:03,320 --> 00:16:06,760 the page and putting that as the input. 255 00:16:06,760 --> 00:16:10,820 So it is as good a job as many of the others that I've used. 256 00:16:10,820 --> 00:16:14,600 And if you're a HubSpot user, maybe even a HubSpot free user, I have to check that, but 257 00:16:14,600 --> 00:16:16,240 chat spots effectively free, right? 258 00:16:16,240 --> 00:16:20,400 So you could absolutely try that and see how you get on. 259 00:16:20,400 --> 00:16:24,620 It will also give you information on podcast stats. 260 00:16:24,620 --> 00:16:27,440 So if you put in a podcast, it will give you a summary of the podcast. 261 00:16:27,440 --> 00:16:32,840 And we checked in artificially intelligent marketing and it found us and it knew that 262 00:16:32,840 --> 00:16:36,920 we'd published it and knew how many episodes there were. 263 00:16:36,920 --> 00:16:39,200 So that was good, good stuff. 264 00:16:39,200 --> 00:16:42,480 It was, it was also slightly sarcastic. 265 00:16:42,480 --> 00:16:43,840 In what sense? 266 00:16:43,840 --> 00:16:47,360 When it said it gave us a rank and it said, where do we rank? 267 00:16:47,360 --> 00:16:49,040 And it said top non. 268 00:16:49,040 --> 00:16:56,560 Yeah, it gave us a rank, rank, but that's okay with us because it's a podcast that's 269 00:16:56,560 --> 00:16:59,740 on the up as you will know, dear listener. 270 00:16:59,740 --> 00:17:01,280 And you will share this with all your friends. 271 00:17:01,280 --> 00:17:06,800 In fact, we're on a mission now for chat spot to give artificially intelligent marketing 272 00:17:06,800 --> 00:17:09,120 a rank that's not top none. 273 00:17:09,120 --> 00:17:14,160 And you dear listener are now a part of this very important social initiative. 274 00:17:14,160 --> 00:17:18,160 Thanks for your help in anticipation. 275 00:17:18,160 --> 00:17:20,000 Why should marketers care about this? 276 00:17:20,000 --> 00:17:25,400 Because as we've been talking about having these types of tools baked into your marketing 277 00:17:25,400 --> 00:17:33,200 and sales CRM is going to give you the type of power to create personalized email programs 278 00:17:33,200 --> 00:17:35,080 for individual leads. 279 00:17:35,080 --> 00:17:40,600 And I really feel like this summarized CRM records tool is a precursor to leveraging 280 00:17:40,600 --> 00:17:43,400 that type of data for other content generation. 281 00:17:43,400 --> 00:17:47,680 So I think it's great as a sales rep, but I think that's what comes next. 282 00:17:47,680 --> 00:17:51,520 I actually had a really practical use case for it this week. 283 00:17:51,520 --> 00:17:57,920 So I've sent out an email campaign this week, or should I say back end of last week, and 284 00:17:57,920 --> 00:18:03,660 I just had chat spot open and I just wanted to check how that campaign had performed open 285 00:18:03,660 --> 00:18:05,320 rate click rate kind of thing. 286 00:18:05,320 --> 00:18:09,480 Normally I would just jump into my dashboard and do that, but I had chat spot open anyway. 287 00:18:09,480 --> 00:18:13,480 So I was just asked it, now it had gone and it told me, it gave me exactly the sort of 288 00:18:13,480 --> 00:18:15,520 report that I would have got if I went into analytic. 289 00:18:15,520 --> 00:18:17,600 And that was super handy. 290 00:18:17,600 --> 00:18:21,160 Just having that there can interrogate the data. 291 00:18:21,160 --> 00:18:25,520 I can ask it to then show me who engaged with it. 292 00:18:25,520 --> 00:18:32,240 And the more I dug into the data and wanted to interrogate it more, it kind of fell down 293 00:18:32,240 --> 00:18:33,240 in a few areas. 294 00:18:33,240 --> 00:18:36,520 But as a starting point, it worked really well. 295 00:18:36,520 --> 00:18:37,520 Yeah. 296 00:18:37,520 --> 00:18:39,080 And it's not many months old. 297 00:18:39,080 --> 00:18:42,120 I think they're doing a lot of work behind the scenes to improve it. 298 00:18:42,120 --> 00:18:47,280 And I think with a lot of these sort of next level applications, it's the teasing of the 299 00:18:47,280 --> 00:18:51,920 progress we're making rather than would we on the podcast today say marketing and sales 300 00:18:51,920 --> 00:18:53,920 folk get chat spot. 301 00:18:53,920 --> 00:18:54,920 You won't need anything else. 302 00:18:54,920 --> 00:18:56,960 It's going to do everything you need. 303 00:18:56,960 --> 00:18:58,320 We're still in the emergence phase. 304 00:18:58,320 --> 00:18:59,920 We get asked a lot, don't we Martin? 305 00:18:59,920 --> 00:19:01,120 What tool should I use for this? 306 00:19:01,120 --> 00:19:03,120 What tool should I use for that? 307 00:19:03,120 --> 00:19:04,720 It's so dynamic at the moment. 308 00:19:04,720 --> 00:19:08,000 Whatever recommendation we make this week is going to be different probably next week 309 00:19:08,000 --> 00:19:10,800 or the week after, or certainly in a month. 310 00:19:10,800 --> 00:19:16,440 And as much as I realize this is not helpful advice, you have to keep an eye on the tools, 311 00:19:16,440 --> 00:19:20,600 especially the ones that baked into software you already use like Salesforce or HubSpot 312 00:19:20,600 --> 00:19:25,920 to just see what's emerging and what you could tap into. 313 00:19:25,920 --> 00:19:29,920 And I think that's the key because I think if you are a HubSpot user, chat spot is arguably 314 00:19:29,920 --> 00:19:33,280 reaching that tipping point of true utility. 315 00:19:33,280 --> 00:19:38,240 Whereas for a lot of time, even the template chats that you could click a lot of the time 316 00:19:38,240 --> 00:19:40,960 ago, I know what you want, but I can't do it. 317 00:19:40,960 --> 00:19:42,680 Like it's literally that would be its response. 318 00:19:42,680 --> 00:19:46,320 And it's like, well, if you know what I want, just do it. 319 00:19:46,320 --> 00:19:51,280 But, you know, and I appreciate that getting these tools to work is a huge undertaking 320 00:19:51,280 --> 00:19:54,360 and it's just good to see the evolution of chat spot. 321 00:19:54,360 --> 00:19:58,440 And I do think I can see it starting to become more of my workflow now it's reached that 322 00:19:58,440 --> 00:19:59,440 tipping point. 323 00:19:59,440 --> 00:20:00,440 Yep. 324 00:20:00,440 --> 00:20:01,440 Very much agreed. 325 00:20:01,440 --> 00:20:04,120 What's next for us, Martin? 326 00:20:04,120 --> 00:20:10,240 There's some new updates in the open source model world. 327 00:20:10,240 --> 00:20:18,000 So Stability AI have released new Beluga models and they are setting records. 328 00:20:18,000 --> 00:20:25,960 So they released two powerful new AI language models, Stable Beluga 1 and Stable Beluga 329 00:20:25,960 --> 00:20:27,920 2. 330 00:20:27,920 --> 00:20:32,420 They put this in a press release and a little detail I really liked about this story was 331 00:20:32,420 --> 00:20:40,200 that they said previously code named Free Willy, the models were renamed after the gentler 332 00:20:40,200 --> 00:20:42,520 and more harmless Beluga whale. 333 00:20:42,520 --> 00:20:50,600 For those of you that have seen the movie Free Willy, my question, how aggressive will 334 00:20:50,600 --> 00:20:54,200 he was? 335 00:20:54,200 --> 00:20:57,840 I don't think we should cut that audio snippet out either because I don't think I'd do very 336 00:20:57,840 --> 00:20:59,480 well online with that. 337 00:20:59,480 --> 00:21:04,880 But yeah, I actually featured when you were out of Play Mart and the Free Willy models. 338 00:21:04,880 --> 00:21:07,560 It's interesting to see them get rebranded. 339 00:21:07,560 --> 00:21:14,080 Something's driven that outside of the nonsense in the press release, I would suggest. 340 00:21:14,080 --> 00:21:19,080 Yeah, they seem to talk about just the reputation of, they put a line in saying that the reputation 341 00:21:19,080 --> 00:21:23,040 of orcas as being these more aggressive animals. 342 00:21:23,040 --> 00:21:27,680 These models were, they needed to be, well actually they specifically say that the fact 343 00:21:27,680 --> 00:21:35,480 that orcas are called killer whales was not ideal and these models are designed like claw 344 00:21:35,480 --> 00:21:39,800 from anthropic to be helpful and harmless. 345 00:21:39,800 --> 00:21:41,840 So the Beluga whale was chosen. 346 00:21:41,840 --> 00:21:43,640 Anyway, enough about the branding. 347 00:21:43,640 --> 00:21:50,720 The Beluga models were trained on synthesized data inspired by Microsoft's orca model, hence 348 00:21:50,720 --> 00:21:54,240 the Free Willy idea. 349 00:21:54,240 --> 00:22:01,320 Stability AI filtered the training data to ensure fair benchmark comparisons. 350 00:22:01,320 --> 00:22:08,080 It uses only 10% of the data size of the orca paper, but it has through their fine tuning 351 00:22:08,080 --> 00:22:13,920 of this model, it's got really high performance, exceptional performance in all of the benchmarks 352 00:22:13,920 --> 00:22:22,280 and they've outscored many of the other similar models on the open source scene such as GPT 353 00:22:22,280 --> 00:22:26,400 for all and AGI of evals. 354 00:22:26,400 --> 00:22:31,960 So yeah, in fact so much so that stable Beluga two currently ranks number one on the open 355 00:22:31,960 --> 00:22:37,720 large language model leaderboard showing strong language understanding and reasoning capabilities 356 00:22:37,720 --> 00:22:41,940 and they're the things that I always look for is that reasoning capabilities because 357 00:22:41,940 --> 00:22:47,960 when you look at GPT four versus GPT 3.5, it was the reasoning that just took it so 358 00:22:47,960 --> 00:22:48,960 much further along. 359 00:22:48,960 --> 00:22:52,720 You go, wow, now I can do more with this model. 360 00:22:52,720 --> 00:23:00,240 So it's a major advance as with all of these stories in the kind of open source world, 361 00:23:00,240 --> 00:23:05,160 why should we care commercially as organizations? 362 00:23:05,160 --> 00:23:09,000 Why should we care as marketers? 363 00:23:09,000 --> 00:23:14,240 Well, better open source models and more competition in open source models is going to give us 364 00:23:14,240 --> 00:23:15,240 more choice. 365 00:23:15,240 --> 00:23:20,200 That said, this is only available under a research license at the moment. 366 00:23:20,200 --> 00:23:26,400 Like Meta's recently released Lama 2 model, which was given the full commercial license. 367 00:23:26,400 --> 00:23:31,520 So until they give it the full commercial license, it's just something to keep an eye 368 00:23:31,520 --> 00:23:32,520 on. 369 00:23:32,520 --> 00:23:36,840 But yeah, more competition in this space is going to give us more options for creating 370 00:23:36,840 --> 00:23:43,520 chat bots and customer service agents and all of these kinds of useful things that I 371 00:23:43,520 --> 00:23:47,200 think AI is going to do for us in the near future. 372 00:23:47,200 --> 00:23:48,480 Absolutely. 373 00:23:48,480 --> 00:23:52,960 You mentioned the LLM tool and leaderboard GPT for all. 374 00:23:52,960 --> 00:23:57,080 I actually installed that a couple of weeks ago and it was quite interesting to play with 375 00:23:57,080 --> 00:23:58,080 it. 376 00:23:58,080 --> 00:23:59,080 Have you had to play with it, Martin? 377 00:23:59,080 --> 00:24:00,080 I have, yeah. 378 00:24:00,080 --> 00:24:04,000 And you can download all the different models and run them locally. 379 00:24:04,000 --> 00:24:05,000 Yeah. 380 00:24:05,000 --> 00:24:12,320 And I've got like 16 gigs of RAM and I've got like an i7 equivalent, I think. 381 00:24:12,320 --> 00:24:17,840 And they ran, but they ran pretty slowly. 382 00:24:17,840 --> 00:24:24,000 And I think it's really interesting seeing the explosion of this space because I do think 383 00:24:24,000 --> 00:24:28,720 the future for a lot of applications is going to be on board large language models when 384 00:24:28,720 --> 00:24:33,200 they figure out, as computing power on board continues to progress. 385 00:24:33,200 --> 00:24:40,120 We've talked previously about how Macs M1 and M2 chips have things like architecture 386 00:24:40,120 --> 00:24:42,920 built in them to enable this type of stuff. 387 00:24:42,920 --> 00:24:46,160 In fact, I need to get GPT for all installed on my Mac, actually see how much better it 388 00:24:46,160 --> 00:24:47,160 performs. 389 00:24:47,160 --> 00:24:48,160 Yeah. 390 00:24:48,160 --> 00:24:51,200 And I think this is where these types of models... 391 00:24:51,200 --> 00:24:55,640 That was the moment that all these models really suddenly were a thing for me because 392 00:24:55,640 --> 00:25:00,600 you can select from, I don't know, 10, 12 models to download and there's a summary to 393 00:25:00,600 --> 00:25:01,960 try and give you a feel for. 394 00:25:01,960 --> 00:25:03,840 This one's like massive download. 395 00:25:03,840 --> 00:25:10,000 It's really slow, but it would give you great output versus small download, much faster, 396 00:25:10,000 --> 00:25:14,080 but maybe not as creative or good at reasoning and stuff. 397 00:25:14,080 --> 00:25:19,640 So the more that we take big models, make them smaller and somehow tweak the parameters 398 00:25:19,640 --> 00:25:24,620 and make them as good, if not better, the closer we get to running large language models 399 00:25:24,620 --> 00:25:28,160 on our computers, which for some organizations may be needed. 400 00:25:28,160 --> 00:25:33,400 They just may not be comfortable having the data go into the cloud. 401 00:25:33,400 --> 00:25:38,400 I can imagine healthcare applications where the regulations and restrictions on how you 402 00:25:38,400 --> 00:25:42,680 use data, where you move data, where you store data. 403 00:25:42,680 --> 00:25:44,720 That could be a bit of a game changer for them. 404 00:25:44,720 --> 00:25:50,000 So I think there's also applications in the marketing space for those companies that may 405 00:25:50,000 --> 00:25:56,400 not be able to ever use cloud-based tools for this because the sharing of data with 406 00:25:56,400 --> 00:26:01,280 them is just too complicated to set up and takes too long. 407 00:26:01,280 --> 00:26:06,200 Although one thing to note on that, and I think this came out of Maycon conference for 408 00:26:06,200 --> 00:26:07,200 me. 409 00:26:07,200 --> 00:26:14,560 I realized that not a lot of people realize that Anthropic and the Claude model is HIPAA 410 00:26:14,560 --> 00:26:15,560 compliant. 411 00:26:15,560 --> 00:26:16,560 Is that right? 412 00:26:16,560 --> 00:26:23,800 Now, I also know, I'm sure you've mentioned that on a previous podcast actually, but yeah. 413 00:26:23,800 --> 00:26:24,800 Okay. 414 00:26:24,800 --> 00:26:31,960 Well, let's skip and talk a bit more about Anthropic's new AI model because they've released 415 00:26:31,960 --> 00:26:39,400 an improved version of Claude Instant 1.2, so some more model news here. 416 00:26:39,400 --> 00:26:43,560 In essence, this new model incorporates some of the strengths of the recently released 417 00:26:43,560 --> 00:26:48,400 Claude 2 that we've talked about on the podcast, showing improvements in areas like maths, 418 00:26:48,400 --> 00:26:51,480 coding, reasoning, and those types of things. 419 00:26:51,480 --> 00:26:56,960 It also generates longer, better structured responses and follows your formatting instructions 420 00:26:56,960 --> 00:27:04,200 for users of chat GPT-4 and 3.5 Claude Instant 1.2 is probably a reasonable option now at 421 00:27:04,200 --> 00:27:10,800 this point, especially because it has a larger context window of about 75,000 words like 422 00:27:10,800 --> 00:27:12,840 Claude 2. 423 00:27:12,840 --> 00:27:14,440 This is pretty cool. 424 00:27:14,440 --> 00:27:18,600 The model is also apparently less prone to hallucination and a bit more resistant to 425 00:27:18,600 --> 00:27:19,600 misuse. 426 00:27:19,600 --> 00:27:22,720 Why is this interesting for marketers? 427 00:27:22,720 --> 00:27:30,520 I think for us on the podcast, seeing that Anthropic and to a certain extent, OpenAI, 428 00:27:30,520 --> 00:27:34,560 are continuing to invest in developing and improving older models. 429 00:27:34,560 --> 00:27:38,120 There was a bit of research that came out a couple of weeks ago that we talked about 430 00:27:38,120 --> 00:27:40,600 on the podcast looking... 431 00:27:40,600 --> 00:27:44,440 The argument of the research was that GPT-4 had got worse, but what kind of got lost in 432 00:27:44,440 --> 00:27:49,200 the shuffle was it looks like GPT-3.5 had got better. 433 00:27:49,200 --> 00:27:55,200 If Claude Instant 1.2 is a significant improvement over 1.1, these are smaller, faster models 434 00:27:55,200 --> 00:27:58,720 whose performance is improving. 435 00:27:58,720 --> 00:28:02,640 When speed is important, you can start to rely on them for some use cases, potentially 436 00:28:02,640 --> 00:28:09,120 content creation use cases, summarizing call transcripts for interviewing SMEs and producing 437 00:28:09,120 --> 00:28:11,360 blog posts and all of those types of things. 438 00:28:11,360 --> 00:28:13,280 I think that's really interesting. 439 00:28:13,280 --> 00:28:19,120 For people who are drifting into power use and playing with APIs like we do, Martin, 440 00:28:19,120 --> 00:28:25,400 the ability to access cheaper models and get good outputs more quickly is important and 441 00:28:25,400 --> 00:28:27,880 therefore it will be important to the software developers. 442 00:28:27,880 --> 00:28:33,200 That could see powerful software tools for marketers that cost even less because these 443 00:28:33,200 --> 00:28:36,360 smaller models are improving a lot. 444 00:28:36,360 --> 00:28:43,040 Yeah, I found with Claude 2.0 that on one particular use case that I was... 445 00:28:43,040 --> 00:28:49,920 It was a workflow that I was building out in Zapier and I just couldn't get it to run 446 00:28:49,920 --> 00:28:50,920 in time. 447 00:28:50,920 --> 00:28:56,320 I kept hitting timeouts because it was a little bit slow and Zapier was just like, oh, we 448 00:28:56,320 --> 00:28:59,560 didn't get a response in time, the end. 449 00:28:59,560 --> 00:29:05,040 So I switched over to Claude Instant, which worked, but the outputs weren't brilliant. 450 00:29:05,040 --> 00:29:08,360 I had to adjust my expectations slightly on that. 451 00:29:08,360 --> 00:29:09,360 So this is great. 452 00:29:09,360 --> 00:29:17,720 I'm very pleased that we've now got a model which is faster and going to give us what 453 00:29:17,720 --> 00:29:18,720 I want. 454 00:29:18,720 --> 00:29:19,720 Absolutely. 455 00:29:19,720 --> 00:29:25,360 We've been hitting those same timeout issues when we use a tool called GPT for Sheets. 456 00:29:25,360 --> 00:29:29,560 Well worth having a look at and a play with for those of you that are into figuring out 457 00:29:29,560 --> 00:29:34,600 how to better use more exotic approaches to improve your marketing outputs. 458 00:29:34,600 --> 00:29:40,440 We use it to summarize interesting stories that we find where we just paste the story 459 00:29:40,440 --> 00:29:46,000 into a cell in the sheet and then we ask Claude 2.0 or some other model in the next column 460 00:29:46,000 --> 00:29:48,800 over to summarize it for us. 461 00:29:48,800 --> 00:29:52,440 We've been getting timeout issues with GPT 4.0 and Claude 2.0. 462 00:29:52,440 --> 00:29:55,880 Whereas, if we can now summarize them with Claude 1.2.0, it's going to be cheaper in 463 00:29:55,880 --> 00:30:01,720 terms of API credits, but also faster and we won't hit that problem. 464 00:30:01,720 --> 00:30:04,000 So yeah, interesting stuff. 465 00:30:04,000 --> 00:30:08,480 I do love playing with different tools that access all of this stuff through API. 466 00:30:08,480 --> 00:30:11,640 It really opens up your creative juices. 467 00:30:11,640 --> 00:30:12,640 It does. 468 00:30:12,640 --> 00:30:17,160 And when you start using the API, the speed of the, when it is quick, the speed is so 469 00:30:17,160 --> 00:30:23,920 much better than using the chat GPT interface or similar. 470 00:30:23,920 --> 00:30:24,920 Absolutely. 471 00:30:24,920 --> 00:30:25,920 Right. 472 00:30:25,920 --> 00:30:28,000 Next news item is with you Martin Longshot. 473 00:30:28,000 --> 00:30:29,000 Yeah. 474 00:30:29,000 --> 00:30:31,920 So Longshot is an AI copywriting tool. 475 00:30:31,920 --> 00:30:34,960 It's one that I've been on board with since basically day one of their launch. 476 00:30:34,960 --> 00:30:40,960 I got an early, got access, had a subscription to the tool since then, and you know, similar 477 00:30:40,960 --> 00:30:48,240 to any AI copywriting tool, they've got templates for writing blog posts and blog introductions 478 00:30:48,240 --> 00:30:51,080 and meta descriptions and social media posts and all of that kind of stuff. 479 00:30:51,080 --> 00:30:53,080 And it does a very good job. 480 00:30:53,080 --> 00:30:57,680 But what I really like about Longshot is that they've always been trying to address customer 481 00:30:57,680 --> 00:31:02,200 needs beyond just content generation, because you can now do that anywhere. 482 00:31:02,200 --> 00:31:07,800 In fact, they were the first tool that I really saw that came out with a fact checking tool. 483 00:31:07,800 --> 00:31:17,840 They had fact GPT, which would do, you could get it to write AI generated content about 484 00:31:17,840 --> 00:31:19,080 things that had just happened. 485 00:31:19,080 --> 00:31:22,920 For instance, in the demo, they talked about the world cup final and being able to write 486 00:31:22,920 --> 00:31:24,720 about the world cup final the day after it. 487 00:31:24,720 --> 00:31:30,800 You didn't have to wait until the models powering it were trained on that kind of data. 488 00:31:30,800 --> 00:31:32,000 So they did a very good job there. 489 00:31:32,000 --> 00:31:36,240 And they've just released a new tool called BotShot. 490 00:31:36,240 --> 00:31:41,680 And BotShot is where you can now create a bot, customer service bot or some sort of 491 00:31:41,680 --> 00:31:46,760 bot that sits on your website or elsewhere that is trained on your data. 492 00:31:46,760 --> 00:31:49,920 And it's a no code solution for doing this. 493 00:31:49,920 --> 00:31:55,560 So we talk about being able to train large language models on your existing proprietary 494 00:31:55,560 --> 00:31:56,560 company data. 495 00:31:56,560 --> 00:32:02,480 Well, now you can do that in a no code way using the BotShot interface. 496 00:32:02,480 --> 00:32:08,040 They released it last week and they say that you can have your own chat GPT for your website 497 00:32:08,040 --> 00:32:11,600 and deploy it in under two minutes. 498 00:32:11,600 --> 00:32:17,040 So if the workflow is as good as it looks, and I think the workflow generally that they 499 00:32:17,040 --> 00:32:20,680 put together is very good. 500 00:32:20,680 --> 00:32:26,480 This is an interesting tool and an interesting way to add new capabilities to your website 501 00:32:26,480 --> 00:32:31,800 without having to really get into things like fine tuning and using vector embeddings and 502 00:32:31,800 --> 00:32:33,960 things like that. 503 00:32:33,960 --> 00:32:39,640 It's interesting to consider how they must do it to get around all of some of those issues. 504 00:32:39,640 --> 00:32:45,480 Is it like customer facing only for like on your website or is it, can you use it almost 505 00:32:45,480 --> 00:32:46,480 like an internal tool? 506 00:32:46,480 --> 00:32:48,400 I think you can use it as an internal thing. 507 00:32:48,400 --> 00:32:49,400 Yeah. 508 00:32:49,400 --> 00:32:54,120 Ultimately, you would just choose where the bot sits and how, you know, the kind of address 509 00:32:54,120 --> 00:32:58,520 that people would interface through it. 510 00:32:58,520 --> 00:33:05,400 I think the way that they do it and my understanding of it is that you give it the data and it 511 00:33:05,400 --> 00:33:08,280 manages the vector embeddings piece for you. 512 00:33:08,280 --> 00:33:12,680 So they've built the workflow that just means that you don't have to be the technical person 513 00:33:12,680 --> 00:33:16,200 that figures all of that kind of stuff out. 514 00:33:16,200 --> 00:33:17,200 Okay. 515 00:33:17,200 --> 00:33:26,000 One of my sort of business fantasies is the ability to codify the training that I've given 516 00:33:26,000 --> 00:33:33,480 people who've joined our agency over the years to in essence create a pool chatbot that the 517 00:33:33,480 --> 00:33:40,680 team can ask bot pool instead of real pool and hopefully get half decent answers out 518 00:33:40,680 --> 00:33:45,240 of it in my stead, right? 519 00:33:45,240 --> 00:33:51,080 Getting me up to focus on other tasks, which could be done through knowledge bases, pulling 520 00:33:51,080 --> 00:33:55,480 a ton of emails out or, you know, other different ways. 521 00:33:55,480 --> 00:34:01,640 And anything that's going to make that type of application a bit easier is very welcome. 522 00:34:01,640 --> 00:34:05,240 And a lot of companies, obviously a lot bigger than Biostrata probably have massive knowledge 523 00:34:05,240 --> 00:34:09,120 bases already that they can open this stuff up to. 524 00:34:09,120 --> 00:34:10,360 So good story. 525 00:34:10,360 --> 00:34:15,360 I think I've got to go and have a little play with bot to shot by long shot. 526 00:34:15,360 --> 00:34:16,680 It's good job. 527 00:34:16,680 --> 00:34:17,680 It's early in the morning, isn't it? 528 00:34:17,680 --> 00:34:20,800 There's a few tongue twisters in here today. 529 00:34:20,800 --> 00:34:22,760 Next story, we're talking Zoom. 530 00:34:22,760 --> 00:34:27,840 So probably a lot of you are and have used Zoom, we're on Zoom right now, the popular 531 00:34:27,840 --> 00:34:33,160 video conferencing platform, which has updated its terms of service to allow the use of some 532 00:34:33,160 --> 00:34:36,200 customer data to train its AI models. 533 00:34:36,200 --> 00:34:41,720 Now these new terms, which came in in July 27 specify that Zoom can use service generated 534 00:34:41,720 --> 00:34:47,280 data such as product usage, telemetry and diagnostic data for machine learning purposes. 535 00:34:47,280 --> 00:34:52,560 However, supposedly the terms state the audio, video or chat content will not be used for 536 00:34:52,560 --> 00:34:56,440 AI training without the customer's consent. 537 00:34:56,440 --> 00:35:00,680 So an interesting one on this one, because I think this was spotted in the terms and 538 00:35:00,680 --> 00:35:06,360 conditions, like it took a couple of weeks for an eagle-eyed person to nab this because 539 00:35:06,360 --> 00:35:10,920 of course we're reporting on this today on the 14th of August. 540 00:35:10,920 --> 00:35:20,320 So that is, I think, indicative of the type of information that we're providing to software 541 00:35:20,320 --> 00:35:23,280 providers that they're potentially using to train AI models on. 542 00:35:23,280 --> 00:35:29,080 And we'll just come to that in a moment, because if you're using Zoom's two new AI features, 543 00:35:29,080 --> 00:35:33,640 which came out in June, which is a meeting summary tool and a chat message composition 544 00:35:33,640 --> 00:35:39,560 tool to use these features, you must sign a consent form allowing Zoom to use your data 545 00:35:39,560 --> 00:35:45,960 to improve the performance and accuracy of these AI services, a la in essence, we're 546 00:35:45,960 --> 00:35:49,320 going to use your content to better train our models. 547 00:35:49,320 --> 00:35:53,000 So if you're opting into using these cool new features, that's the bit where it says 548 00:35:53,000 --> 00:35:56,960 we don't use your data unless you give us permission. 549 00:35:56,960 --> 00:35:58,800 That's the bit where you give them permission. 550 00:35:58,800 --> 00:36:01,520 It's not, hey, we want to train our tools on your data. 551 00:36:01,520 --> 00:36:03,160 No, it's packaged in a here. 552 00:36:03,160 --> 00:36:04,320 Do you want this cool thing? 553 00:36:04,320 --> 00:36:06,680 Oh yeah, I want to play with that cool thing. 554 00:36:06,680 --> 00:36:09,160 By the way, we could use your data now. 555 00:36:09,160 --> 00:36:10,160 Right? 556 00:36:10,160 --> 00:36:13,520 So you've got to be really hot on keeping an eye on these things. 557 00:36:13,520 --> 00:36:20,040 I think this is where, particularly in areas such as the EU, companies are going to get 558 00:36:20,040 --> 00:36:21,040 really unstuck. 559 00:36:21,040 --> 00:36:27,000 They're going to have to be really, really explicit about separating this out in the 560 00:36:27,000 --> 00:36:28,000 terms and conditions. 561 00:36:28,000 --> 00:36:36,240 So we know that for instance, with OpenAI, if you put data into chat GPT, they're using 562 00:36:36,240 --> 00:36:41,600 that model to train the future models. 563 00:36:41,600 --> 00:36:44,400 If you use the API, they're not. 564 00:36:44,400 --> 00:36:45,640 They're keeping those separate. 565 00:36:45,640 --> 00:36:49,640 And I think the likes of Zoom are going to have to be really explicit about, look, we 566 00:36:49,640 --> 00:36:55,120 can give you this cool new feature and we will take your audio and your feeds and we 567 00:36:55,120 --> 00:37:01,200 can do the summaries and we can do the meeting notes and we can do the sales coaching like 568 00:37:01,200 --> 00:37:05,800 Gong or whatever it is that they're planning to do in the future while separating out the 569 00:37:05,800 --> 00:37:12,380 are you opting into our further training enhancement of our own AI models? 570 00:37:12,380 --> 00:37:17,920 Because I just think that regulators like the EU will come down on them like a ton of 571 00:37:17,920 --> 00:37:18,920 bricks. 572 00:37:18,920 --> 00:37:19,920 Yeah. 573 00:37:19,920 --> 00:37:21,360 It's quite a messy one, isn't it? 574 00:37:21,360 --> 00:37:22,360 From that perspective. 575 00:37:22,360 --> 00:37:24,480 We've spoke about in the podcast before. 576 00:37:24,480 --> 00:37:27,320 If the product's free, it's because you are the product. 577 00:37:27,320 --> 00:37:33,240 But in essence here, even if you're paying for this product, you're still the product, 578 00:37:33,240 --> 00:37:34,240 right? 579 00:37:34,240 --> 00:37:37,240 Because a lot of us are paying for Zoom. 580 00:37:37,240 --> 00:37:45,600 And I think ultimately, there's so much data needed to train these models that wherever 581 00:37:45,600 --> 00:37:49,480 these companies can get access to it, it becomes a very valuable commodity for them to access 582 00:37:49,480 --> 00:37:53,180 and that we're going to see more and more of this and probably has happened for many 583 00:37:53,180 --> 00:37:57,560 years and now our data is being used and we're just maybe not as aware of it and now we're 584 00:37:57,560 --> 00:37:59,200 becoming aware of it. 585 00:37:59,200 --> 00:38:04,500 I think it's also going to make it very interesting to see how this plays out when Copilot gets 586 00:38:04,500 --> 00:38:10,560 rolled out to Office 365 and Google's Workspace version because ultimately, the power of a 587 00:38:10,560 --> 00:38:15,720 lot of those tools is going to be predicated on being able to in essence, natural language 588 00:38:15,720 --> 00:38:20,600 search through all your own documents and have a chatbot like to create content off 589 00:38:20,600 --> 00:38:21,840 the back of them for you. 590 00:38:21,840 --> 00:38:26,440 While by default, they're going to have to access all your stuff. 591 00:38:26,440 --> 00:38:28,600 How is that going to be managed? 592 00:38:28,600 --> 00:38:30,320 Is there going to be a training element to that? 593 00:38:30,320 --> 00:38:34,640 If you opt into using your information and allowing it to be part of the training set 594 00:38:34,640 --> 00:38:38,240 for future runs, can you get a discount on your subscriptions? 595 00:38:38,240 --> 00:38:39,680 Who knows, right? 596 00:38:39,680 --> 00:38:45,840 Because it's very interesting to see how all of that's going to play out. 597 00:38:45,840 --> 00:38:46,840 Agent Bench. 598 00:38:46,840 --> 00:38:50,880 Tell us about Agent Bench, Martin. 599 00:38:50,880 --> 00:38:56,020 So on the discussion last week or in the interview last week with Brennan Woodruff, we spoke 600 00:38:56,020 --> 00:39:01,080 about AI agents and the capabilities of agents. 601 00:39:01,080 --> 00:39:05,160 So this is large language models being able to complete tasks, being able to actually 602 00:39:05,160 --> 00:39:12,520 do things, not just tell you the next word or the next token, but actually go and execute 603 00:39:12,520 --> 00:39:15,000 things for you. 604 00:39:15,000 --> 00:39:20,400 And how well AI agents work has never really been benchmarked until now. 605 00:39:20,400 --> 00:39:27,080 We've now got Agent Bench, which is a new benchmark developed by researchers. 606 00:39:27,080 --> 00:39:33,040 That's a twisting words like researchers. 607 00:39:33,040 --> 00:39:37,440 What a complicated word for 10.30 in the morning. 608 00:39:37,440 --> 00:39:44,280 Researchers at Ohio State University, UC Berkeley and Tsinghua University in China. 609 00:39:44,280 --> 00:39:50,640 So Agent Bench evaluates large language models across eight different environments, all of 610 00:39:50,640 --> 00:39:53,880 which simulate real world situations. 611 00:39:53,880 --> 00:40:00,040 So they look at the ability to interact with operating systems, databases, knowledge graphs. 612 00:40:00,040 --> 00:40:03,760 There's a digital card game, which they have to play. 613 00:40:03,760 --> 00:40:09,120 There's some lateral thinking puzzles, and householding, which is kind of interaction 614 00:40:09,120 --> 00:40:10,120 with the real world. 615 00:40:10,120 --> 00:40:16,720 So it's almost like a 3D kitchen environment, I think that one is, almost like robotics 616 00:40:16,720 --> 00:40:18,280 or controlling robotics. 617 00:40:18,280 --> 00:40:23,680 I think it's kind of virtual household environment. 618 00:40:23,680 --> 00:40:27,520 Web shopping and web browsing. 619 00:40:27,520 --> 00:40:29,680 So they're the different categories. 620 00:40:29,680 --> 00:40:35,080 And the goal is ultimately to test the large language models capabilities for coding, logical 621 00:40:35,080 --> 00:40:40,360 reasoning, knowledge acquisition, and actual interacting with the world. 622 00:40:40,360 --> 00:40:49,200 So the researchers benchmarked 25 large language models, both API accessible large language 623 00:40:49,200 --> 00:40:57,200 models, such as GPT-4, Claude, Claude Instant, and then a bunch of open source models as 624 00:40:57,200 --> 00:40:59,640 well. 625 00:40:59,640 --> 00:41:05,200 And it should come as no surprise to hear that the API based models outperformed the 626 00:41:05,200 --> 00:41:11,600 open source models significantly, but there was one standout model amongst the list. 627 00:41:11,600 --> 00:41:17,040 Paul, I don't know if you want to have a guess at which model outperformed the others. 628 00:41:17,040 --> 00:41:18,280 Couldn't possibly, yes. 629 00:41:18,280 --> 00:41:21,720 It wouldn't be an open AI model, almost certainly. 630 00:41:21,720 --> 00:41:22,720 That's where you're wrong. 631 00:41:22,720 --> 00:41:28,880 It was an open AI model and it was their state of the art model GPT-4. 632 00:41:28,880 --> 00:41:30,200 All you are blown away. 633 00:41:30,200 --> 00:41:38,720 So it scored on the agent bench overall score, it scored 4.41. 634 00:41:38,720 --> 00:41:46,080 And that compares to Claude, which scored 2.77 in second place. 635 00:41:46,080 --> 00:41:51,600 So significant outperformance or overperformance there. 636 00:41:51,600 --> 00:41:54,080 It smashed the rest of the field, let's be honest. 637 00:41:54,080 --> 00:41:55,400 Yeah, it really did. 638 00:41:55,400 --> 00:42:04,080 There's a graphic that shows how it performs in each of those areas and it absolutely dominates 639 00:42:04,080 --> 00:42:11,120 in every area except bizarrely the web shopping. 640 00:42:11,120 --> 00:42:23,600 GPT-3.5 Turbo outperforms GPT-4, but in every other field GPT-4 outperforms and scores really 641 00:42:23,600 --> 00:42:24,600 highly. 642 00:42:24,600 --> 00:42:34,760 I think generally speaking though, the researchers came away with the conclusion that none of 643 00:42:34,760 --> 00:42:42,280 them, even GPT-4 with its great performance relative to the rest, they're not ready for 644 00:42:42,280 --> 00:42:48,320 public consumption in terms of actually rolling out to the real world as autonomous agents 645 00:42:48,320 --> 00:42:49,320 just yet. 646 00:42:49,320 --> 00:42:53,360 But I think what it demonstrates is that it's coming. 647 00:42:53,360 --> 00:43:05,080 We are on the precipice of some very capable autonomous agents coming to the market and 648 00:43:05,080 --> 00:43:06,080 that's going to change. 649 00:43:06,080 --> 00:43:08,120 I mean, it's interesting to think about it from the marketer's perspective, but just 650 00:43:08,120 --> 00:43:14,680 as consumers, as knowledge workers, what this is going to enable us to do is, it's kind 651 00:43:14,680 --> 00:43:15,960 of hard to imagine right now. 652 00:43:15,960 --> 00:43:22,280 And I think for me, the AI agent space is where if I kind of fast forward, if I throw 653 00:43:22,280 --> 00:43:29,880 it ahead two or three years time, that's going to be the real big difference to now and then. 654 00:43:29,880 --> 00:43:34,280 We look back to, I don't know, let's say like 2005, what's really different. 655 00:43:34,280 --> 00:43:39,120 Well, smartphones and apps is the kind of thing that you go, oh yeah, well we had these 656 00:43:39,120 --> 00:43:43,040 flip phones in 2005. 657 00:43:43,040 --> 00:43:49,560 In like 2030, I think we'll be looking back at 2023 and going, we didn't even have those 658 00:43:49,560 --> 00:43:56,040 AI agents that are just constantly doing tasks for us behind the scenes now. 659 00:43:56,040 --> 00:43:58,960 We talked about this a little bit last week. 660 00:43:58,960 --> 00:44:06,120 HyperWrite, which started out as a writing support tool, really like Reowriter, Jasper, 661 00:44:06,120 --> 00:44:12,680 all these other tools, but they've got their AI assistant now, haven't they, that's appearing 662 00:44:12,680 --> 00:44:16,040 to gain more and more capability. 663 00:44:16,040 --> 00:44:21,120 In essence, you ask it to do things for you like book a flight for you or find a recipe 664 00:44:21,120 --> 00:44:26,500 or on the website, they give examples like organize my Gmail inbox and draft responses 665 00:44:26,500 --> 00:44:29,560 for me, find engineering candidates on LinkedIn. 666 00:44:29,560 --> 00:44:36,520 So really stringing together some of those tasks, which I think is interesting and actually 667 00:44:36,520 --> 00:44:44,360 is a fairly handy segue to our last story of the week, which is about MetaGPT, which 668 00:44:44,360 --> 00:44:48,000 aims to replace entire software companies. 669 00:44:48,000 --> 00:44:55,120 So this is a really interesting one that we saw on the superhuman newsletter this week, 670 00:44:55,120 --> 00:45:01,000 which is an AI system that basically works by assigning different language models, specific 671 00:45:01,000 --> 00:45:07,180 roles like engineer and project manager, and you give MetaGPT a simple prompt and it activates 672 00:45:07,180 --> 00:45:12,440 a chain of AI agents that collaborate and build sites and more. 673 00:45:12,440 --> 00:45:17,600 So the example that was given was that a developer created an entire flappy bird game in just 674 00:45:17,600 --> 00:45:24,720 10 minutes using the tool and only providing one line of code and MetaGPT did the rest. 675 00:45:24,720 --> 00:45:31,240 So to your previous point, mine, this is an interesting emerging area of chaining together 676 00:45:31,240 --> 00:45:37,560 tools to develop more comp and deliver more complex outcomes and a bit of a glimpse into 677 00:45:37,560 --> 00:45:43,160 the future where creating apps and other digital products becomes so much easier and faster, 678 00:45:43,160 --> 00:45:49,200 even for people who don't have necessarily the training skills or background for that 679 00:45:49,200 --> 00:45:51,520 type of stuff. 680 00:45:51,520 --> 00:45:56,600 I've been using chat GPT to create a little bit of script for me that nobody else in the 681 00:45:56,600 --> 00:45:59,440 world would need, but I can't code. 682 00:45:59,440 --> 00:46:02,160 I can't use Python, but now I can. 683 00:46:02,160 --> 00:46:07,040 I still can't write the code, but I can get code generated for the specific task that 684 00:46:07,040 --> 00:46:08,960 I need and executed and off I go. 685 00:46:08,960 --> 00:46:14,520 Now, if I can do that, everyone's going to be able to do that very soon. 686 00:46:14,520 --> 00:46:19,880 And I think this example with MetaGPT is a great example. 687 00:46:19,880 --> 00:46:24,720 I love the fact that he's created these specific roles like project manager and engineer. 688 00:46:24,720 --> 00:46:27,520 It's kind of company in a box, isn't it? 689 00:46:27,520 --> 00:46:33,620 It'd be interesting to see how this performs on, because this is AI agent, but it's still 690 00:46:33,620 --> 00:46:34,960 very specialized. 691 00:46:34,960 --> 00:46:42,680 This is exclusively focused on the task of creating software. 692 00:46:42,680 --> 00:46:50,840 But using the agent bench scorecard, I'd like to see how MetaGPT performs in the benchmark 693 00:46:50,840 --> 00:46:56,120 tests because it isn't going to perform very well presumably on things like householding 694 00:46:56,120 --> 00:47:01,640 and lateral thinking puzzles maybe, but would perform pretty well on things like operating 695 00:47:01,640 --> 00:47:04,120 systems, databases, knowledge graphs. 696 00:47:04,120 --> 00:47:05,120 Yeah. 697 00:47:05,120 --> 00:47:11,520 I think as well, the focus here quite obviously, but it's worth stating on digital outcomes. 698 00:47:11,520 --> 00:47:12,520 Right? 699 00:47:12,520 --> 00:47:18,160 Until robotics catches up, we're going to be able to create perhaps autonomous chains 700 00:47:18,160 --> 00:47:24,760 of agents that can do interesting things automatically and autonomously in the digital world. 701 00:47:24,760 --> 00:47:29,640 But until we overcome that interface problem to the physical world, that's where it's 702 00:47:29,640 --> 00:47:30,760 going to be restricted to. 703 00:47:30,760 --> 00:47:35,720 That lots of things happen in the digital world, most knowledge workers these days operate 704 00:47:35,720 --> 00:47:38,480 almost exclusively in the digital world. 705 00:47:38,480 --> 00:47:40,600 So it's not like they won't have a massive impact. 706 00:47:40,600 --> 00:47:42,640 They will. 707 00:47:42,640 --> 00:47:49,320 But I think once the robotics catches up, that's when we hit the bottom of another exponential 708 00:47:49,320 --> 00:47:51,240 curve from there. 709 00:47:51,240 --> 00:47:57,160 I really love your point as well, Martin, about using GPT-4, which you can do today 710 00:47:57,160 --> 00:47:59,120 to help you with code. 711 00:47:59,120 --> 00:48:05,800 You're one of the best copy-pasters I know, asking great questions of GPT-4 and then copy-pasting 712 00:48:05,800 --> 00:48:08,440 the results and getting stuff to run for you. 713 00:48:08,440 --> 00:48:09,440 I do you a disservice. 714 00:48:09,440 --> 00:48:10,440 I'm joking, obviously. 715 00:48:10,440 --> 00:48:11,440 Fine. 716 00:48:11,440 --> 00:48:13,240 You're a fantastic problem solver. 717 00:48:13,240 --> 00:48:14,240 It's not. 718 00:48:14,240 --> 00:48:15,240 Do you know what? 719 00:48:15,240 --> 00:48:19,960 You're not far off though, because yeah, you can write a good prompt and you can go back 720 00:48:19,960 --> 00:48:23,480 and forth and refine, but ultimately it comes down to just asking a couple of questions 721 00:48:23,480 --> 00:48:25,320 and then copying and pasting. 722 00:48:25,320 --> 00:48:28,400 And sometimes when it doesn't give you the output you want, literally just go back in 723 00:48:28,400 --> 00:48:34,560 and saying, well, that didn't work because it said this and running those time and time 724 00:48:34,560 --> 00:48:35,560 again. 725 00:48:35,560 --> 00:48:36,560 Absolutely. 726 00:48:36,560 --> 00:48:40,840 And if you think about it at the moment, most marketers and sales folk are relying on off-the-shelf 727 00:48:40,840 --> 00:48:44,520 software and SaaS platforms to do the stuff they need to do. 728 00:48:44,520 --> 00:48:49,800 But as these types of, well, as the approaches that you take already are empowering you to 729 00:48:49,800 --> 00:48:56,400 create snippets of code to do things you want to do that are very specific to you, obviously 730 00:48:56,400 --> 00:49:00,480 you're an early adopter and you're having to think around problems and be quite creative. 731 00:49:00,480 --> 00:49:07,640 But once these become packaged in products or out of the box type tools, the imagination 732 00:49:07,640 --> 00:49:11,800 becomes the only limit really in terms of as a marketer, oh, I wish I could do this 733 00:49:11,800 --> 00:49:12,800 thing. 734 00:49:12,800 --> 00:49:14,880 Why isn't there a tool out there that does it? 735 00:49:14,880 --> 00:49:16,560 Build your own. 736 00:49:16,560 --> 00:49:21,100 Ask your agent team to go build it for you and then give them some feedback when it comes 737 00:49:21,100 --> 00:49:22,920 back and iterate it through. 738 00:49:22,920 --> 00:49:25,360 So interesting times. 739 00:49:25,360 --> 00:49:29,760 I really wonder how long do you think it will take, completely unfair question for you Ryan, 740 00:49:29,760 --> 00:49:35,400 how long do you think it will take before we can start to use some of these agents to 741 00:49:35,400 --> 00:49:37,520 really get useful stuff done? 742 00:49:37,520 --> 00:49:48,800 I think we'll be looking at probably about two to three years when they're really fully 743 00:49:48,800 --> 00:49:49,800 functional. 744 00:49:49,800 --> 00:49:56,600 At the moment, we're starting to see the glimmers of it and we're going to see buggy systems, 745 00:49:56,600 --> 00:50:01,000 trial and error, iteration over time, but I think, you know, throw it out to a three 746 00:50:01,000 --> 00:50:07,000 years time and we'll have, it's probably a similar timeframe actually to the likes of 747 00:50:07,000 --> 00:50:13,560 GPT-5 and the kind of multimodality coming in. 748 00:50:13,560 --> 00:50:14,920 Yeah. 749 00:50:14,920 --> 00:50:24,280 I mean, code interpreters interesting because it takes you through its reasoning and tries 750 00:50:24,280 --> 00:50:25,280 different things. 751 00:50:25,280 --> 00:50:30,720 And I don't know if that's like an artifact or like it's for show to make the human feel 752 00:50:30,720 --> 00:50:35,280 like inside the loop of what it's doing or there actually needs to like reason with itself. 753 00:50:35,280 --> 00:50:36,280 And it's like, oh, okay. 754 00:50:36,280 --> 00:50:40,680 Well, to access, to analyze data you just gave me like that, we probably need to format 755 00:50:40,680 --> 00:50:46,200 it in this way or run this type of analysis and it doesn't need to say that out loud. 756 00:50:46,200 --> 00:50:51,720 One assumes obviously the showing the text part to me is absolutely for me, but I wonder 757 00:50:51,720 --> 00:50:54,000 if that's how it actually reasons. 758 00:50:54,000 --> 00:50:55,840 Like it goes through a reasoning process. 759 00:50:55,840 --> 00:50:58,880 With you know, chain of thought prompting is a thing, isn't it? 760 00:50:58,880 --> 00:51:04,200 You know, asking large language models to show their reasoning and show their working. 761 00:51:04,200 --> 00:51:09,080 The documentation, if you've ever read the documentation of Claude on how to get most 762 00:51:09,080 --> 00:51:13,400 out of the model, just look at the documentation and it says, ask it to show it's working, 763 00:51:13,400 --> 00:51:16,400 but it gives you examples of how to put that into a prompt itself. 764 00:51:16,400 --> 00:51:21,840 You can't just ask it a question and then at the end of it, put, show you're working 765 00:51:21,840 --> 00:51:25,680 there's a specific structure that you have to put in, in order for it to then take the 766 00:51:25,680 --> 00:51:29,840 steps which gets a better reasoned output. 767 00:51:29,840 --> 00:51:35,520 So I would imagine that part of code interpreters set up is the same thing. 768 00:51:35,520 --> 00:51:41,720 It's been been trained to go through step by step and show it's working in order to 769 00:51:41,720 --> 00:51:43,880 get the outputs that it requires. 770 00:51:43,880 --> 00:51:44,880 Yeah. 771 00:51:44,880 --> 00:51:48,400 Well, we will see if your prediction comes true. 772 00:51:48,400 --> 00:51:51,960 Personally, I hope we get it a little bit faster than that, but I think it's probably 773 00:51:51,960 --> 00:51:53,840 more realistic what you said. 774 00:51:53,840 --> 00:51:54,840 Right. 775 00:51:54,840 --> 00:52:00,000 As it's your birthday, I think we should let you go off and enjoy the rest of your day. 776 00:52:00,000 --> 00:52:03,480 As always, thanks for your time on the podcast. 777 00:52:03,480 --> 00:52:05,240 And look forward to the next one. 778 00:52:05,240 --> 00:52:07,640 Thanks for cracking stuff for our lovely listeners. 779 00:52:07,640 --> 00:52:10,520 If you're not subscribed yet, hit subscribe because we're going to provide you with these 780 00:52:10,520 --> 00:52:12,800 insights every week. 781 00:52:12,800 --> 00:52:18,600 And of course, we're now on our mission to have Chatspot effectively recognize artificially 782 00:52:18,600 --> 00:52:22,560 intelligent marketing as a podcast worth ranking. 783 00:52:22,560 --> 00:52:26,640 So please share this with your pals, your friends, anyone in marketing or business land 784 00:52:26,640 --> 00:52:31,840 that you would feel would benefit from some of the insights and news and information and 785 00:52:31,840 --> 00:52:34,840 product that we talk about on the podcast. 786 00:52:34,840 --> 00:52:37,240 Other than that, have a lovely week, everyone. 787 00:52:37,240 --> 00:52:38,240 Thanks again, Martin. 788 00:52:38,240 --> 00:52:39,240 Bye bye. 789 00:52:39,240 --> 00:52:40,240 Cheers. 790 00:52:40,240 --> 00:52:41,240 Bye. 791 00:52:41,240 --> 00:52:45,240 Thank you for listening to Artificially Intelligent Marketing. 792 00:52:45,240 --> 00:52:51,280 To stay on top of the latest trends, tips and tools in the world of marketing AI, be 793 00:52:51,280 --> 00:52:53,040 sure to subscribe. 794 00:52:53,040 --> 00:53:07,040 We look forward to seeing you again next week.