1 00:00:00,000 --> 00:00:10,000 Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the 2 00:00:10,000 --> 00:00:15,740 latest trends, tips, and tools in the world of marketing AI, helping you get the best 3 00:00:15,740 --> 00:00:18,560 results from your marketing efforts. 4 00:00:18,560 --> 00:00:22,840 Now let's join our hosts, Paul Avery and Martin Broadhurst. 5 00:00:22,840 --> 00:00:28,480 Welcome to episode 13 of Artificially Intelligent Marketing. 6 00:00:28,480 --> 00:00:31,960 Lucky number 13, Martin, what are your thoughts on that? 7 00:00:31,960 --> 00:00:35,440 I think something's going to go horribly wrong. 8 00:00:35,440 --> 00:00:36,440 Let's hope not. 9 00:00:36,440 --> 00:00:38,320 I think maybe it'll be the best episode ever. 10 00:00:38,320 --> 00:00:40,760 So that's what I'm aiming for. 11 00:00:40,760 --> 00:00:42,880 What have we got for you all lovely people this week? 12 00:00:42,880 --> 00:00:46,440 Well, we've got a few short snippets as usual. 13 00:00:46,440 --> 00:00:51,980 We've got a series of main stories that we're going to cover today, including HubSpot rolling 14 00:00:51,980 --> 00:00:54,080 out campaign assistant. 15 00:00:54,080 --> 00:01:00,720 We're going to look at an attorney in the US who used ChatGPT to build a brief that 16 00:01:00,720 --> 00:01:04,520 was then full of junk, unfortunately, because ChatGPT made a load of stuff up and there 17 00:01:04,520 --> 00:01:05,520 was a big hullabaloo. 18 00:01:05,520 --> 00:01:06,640 We're going to talk about that. 19 00:01:06,640 --> 00:01:13,000 We're going to talk about a sneak peek into the future of OpenAI from Sam Altman. 20 00:01:13,000 --> 00:01:16,800 And we're going to talk about a new paper that was released this week where OpenAI trained 21 00:01:16,800 --> 00:01:19,680 the model based on its reasoning, not its outcome. 22 00:01:19,680 --> 00:01:22,640 We'll talk about a little bit about that later. 23 00:01:22,640 --> 00:01:27,560 And we're also going to lean into our tool of the week this week, which is Recraft AI. 24 00:01:27,560 --> 00:01:28,960 Right. 25 00:01:28,960 --> 00:01:34,520 Let's get them short snippets done first, Martin, because they're quickies this week. 26 00:01:34,520 --> 00:01:39,080 Some of you lovely folks might have seen that Microsoft is rolling out Intelligent Recap 27 00:01:39,080 --> 00:01:41,360 for Teams meetings. 28 00:01:41,360 --> 00:01:47,760 So in essence, Intelligent Recap is bringing a number of the features that a couple of 29 00:01:47,760 --> 00:01:52,760 other tools already have like Tactique and Otter in terms of doing things like smart meeting 30 00:01:52,760 --> 00:01:56,840 summaries, capturing the actions and that type of stuff. 31 00:01:56,840 --> 00:02:01,680 So if you're a Microsoft Teams user, and especially if you're already a premium user, go check 32 00:02:01,680 --> 00:02:02,680 that out. 33 00:02:02,680 --> 00:02:05,020 I'm not, so I haven't been able to play with it, but you definitely should because I bet 34 00:02:05,020 --> 00:02:07,860 it can save you and your team some time. 35 00:02:07,860 --> 00:02:12,800 Then we had a fairly interesting and cool life science company called Benchside that 36 00:02:12,800 --> 00:02:16,640 raised $95 million to help build out, further build out. 37 00:02:16,640 --> 00:02:19,280 It's AI drug discovery platform. 38 00:02:19,280 --> 00:02:24,080 I haven't had a chance to dig into this too much yet, Martin, but it looks really cool. 39 00:02:24,080 --> 00:02:30,240 Like they're collating biological data from across a wide range of sources and using machine 40 00:02:30,240 --> 00:02:37,080 learning and visual recognition to basically make that information searchable and for you 41 00:02:37,080 --> 00:02:40,600 to be able to draw links between, I don't know, different disease states, different 42 00:02:40,600 --> 00:02:43,320 drug trials from the past and that type of stuff. 43 00:02:43,320 --> 00:02:47,080 So yeah, I haven't dug in too deep to say for sure, but it looks kind of interesting 44 00:02:47,080 --> 00:02:49,760 and I will be looking into that in a bit more detail. 45 00:02:49,760 --> 00:02:56,240 And then the last short snippet is the launch of Nvidia's generative AI service, Nvidia 46 00:02:56,240 --> 00:02:59,340 ACE, which is for game developers. 47 00:02:59,340 --> 00:03:05,680 This was pretty cool because one of the demos was trying to use low latency natural language 48 00:03:05,680 --> 00:03:11,400 generation to have non-playable characters interact with you in a less formulaic way. 49 00:03:11,400 --> 00:03:16,120 So you can imagine you're playing a game and you're playing it through for the stick 50 00:03:16,120 --> 00:03:19,700 time and rather than being able to predict what that character says when you walk up 51 00:03:19,700 --> 00:03:24,040 to them, you no longer can with this type of technology because they'll automatically 52 00:03:24,040 --> 00:03:26,680 adapt to the different things that you might say to them. 53 00:03:26,680 --> 00:03:27,680 You're smiling. 54 00:03:27,680 --> 00:03:34,280 I saw the demo and it was a player walking up to someone, there was an NPC behind the 55 00:03:34,280 --> 00:03:37,640 bar and there was the conversation that took place. 56 00:03:37,640 --> 00:03:38,640 It just looked cool. 57 00:03:38,640 --> 00:03:44,640 I just immediately imagine how much an open world game is going to be expanded by that. 58 00:03:44,640 --> 00:03:46,640 I mean, yeah, incredible really. 59 00:03:46,640 --> 00:03:53,080 We've got supposedly, we've got Apple's VR AR headset announcement next week as well. 60 00:03:53,080 --> 00:03:57,960 So that'll be quite interesting to see how much that leaps the industry forward because 61 00:03:57,960 --> 00:04:01,400 that would be the other bit of technology to plug into your vision there, Martin, wouldn't 62 00:04:01,400 --> 00:04:06,820 it, for maybe being fully immersed and able to not physically, but walk up to the bar 63 00:04:06,820 --> 00:04:10,720 in VR and have that conversation would be pretty cool. 64 00:04:10,720 --> 00:04:11,720 Yeah. 65 00:04:11,720 --> 00:04:12,720 Yeah. 66 00:04:12,720 --> 00:04:18,560 I think the VR games just haven't quite tipped over yet into the mainstream. 67 00:04:18,560 --> 00:04:23,320 There's still something holding people back, but if any company can make it mainstream, 68 00:04:23,320 --> 00:04:26,040 let's be honest, I think our money would be on Apple to do so. 69 00:04:26,040 --> 00:04:28,000 That's what I'm kind of hoping for, if I'm honest. 70 00:04:28,000 --> 00:04:34,080 I've had a Quest 2 since probably the start of the pandemic, so not long after they came 71 00:04:34,080 --> 00:04:36,080 out. 72 00:04:36,080 --> 00:04:37,480 And I really like it. 73 00:04:37,480 --> 00:04:40,040 I think it's a cool piece of tech. 74 00:04:40,040 --> 00:04:43,660 There's a boxing simulator, which is surprisingly fun to play. 75 00:04:43,660 --> 00:04:45,920 There's a golfing simulator, which is awesome. 76 00:04:45,920 --> 00:04:50,720 Playing table tennis in VR basically got me through the pandemic in some respects. 77 00:04:50,720 --> 00:04:55,400 So I think there's loads of niche cases that are really enjoyable, but you're right, if 78 00:04:55,400 --> 00:04:58,080 Apple can bring it to the mainstream, that would be kind of cool. 79 00:04:58,080 --> 00:05:03,180 And when I saw this story, I couldn't help but think with my marketing head, if we can 80 00:05:03,180 --> 00:05:12,760 create NPCs that can be that dynamic and with things like Unreal Engine 5.2 and really significant 81 00:05:12,760 --> 00:05:17,840 improvements in graphics, does this change how things like bots, customer service bots 82 00:05:17,840 --> 00:05:20,520 and sales bots on websites might operate? 83 00:05:20,520 --> 00:05:25,920 Because it might be more like a virtual video call, the bots on those websites rather than 84 00:05:25,920 --> 00:05:30,760 a little chat window, and we might be able to have fairly dynamic interactions with them 85 00:05:30,760 --> 00:05:32,920 a little bit like you would with NPCs. 86 00:05:32,920 --> 00:05:38,520 So if they can develop and enhance and figure out how to make the technology work in games, 87 00:05:38,520 --> 00:05:42,760 we might see it filtering to customer service, sales and marketing. 88 00:05:42,760 --> 00:05:44,760 There's a thought. 89 00:05:44,760 --> 00:05:51,240 Right, let's get into our big old stories for this week. 90 00:05:51,240 --> 00:05:55,920 And we're going to start for number one, which is that HubSpot rolls out Campaign Assistant, 91 00:05:55,920 --> 00:05:59,440 their free AI marketing asset creator. 92 00:05:59,440 --> 00:06:03,280 Has been having a little look at this, Martin, tell us more about it. 93 00:06:03,280 --> 00:06:07,800 This was in beta and actually it still is in beta, but they've opened it up that anybody 94 00:06:07,800 --> 00:06:08,800 can access it now. 95 00:06:08,800 --> 00:06:15,600 And this is a generative AI tool from HubSpot that will generate copy for landing pages, 96 00:06:15,600 --> 00:06:18,240 emails or Google ads. 97 00:06:18,240 --> 00:06:20,240 And it's quite a nice interface. 98 00:06:20,240 --> 00:06:23,160 You just answer a few questions. 99 00:06:23,160 --> 00:06:29,320 It literally has a little box that tells you to describe what it is that you want to promote 100 00:06:29,320 --> 00:06:30,440 or that you're offering. 101 00:06:30,440 --> 00:06:35,320 And then it will ask about the tone and the style so you can select three boxes and pretty 102 00:06:35,320 --> 00:06:36,640 much that's it. 103 00:06:36,640 --> 00:06:40,680 And then it will spit out a result. 104 00:06:40,680 --> 00:06:42,480 It's a hundred percent free to use. 105 00:06:42,480 --> 00:06:48,240 It doesn't look like they're making any restrictions on this at all. 106 00:06:48,240 --> 00:06:50,160 You do need a HubSpot account. 107 00:06:50,160 --> 00:06:55,940 So when you go along to the Campaign Assistant, if you don't have a HubSpot login, you will 108 00:06:55,940 --> 00:07:04,680 be asked to sign up to the tool, at which point you will then be part of that ecosystem. 109 00:07:04,680 --> 00:07:09,680 But yeah, I thought the fact that it is completely free to use is pretty generous. 110 00:07:09,680 --> 00:07:12,080 The outputs are really good. 111 00:07:12,080 --> 00:07:19,360 I thought when I used it myself, I wasn't sure whether they've maybe in their system 112 00:07:19,360 --> 00:07:27,880 prompt in the way that they've configured it, whether they've given it quite tight specification 113 00:07:27,880 --> 00:07:29,920 and examples with fine tuning. 114 00:07:29,920 --> 00:07:36,640 I suspect that they have, but you can actually see the user input as the prompt. 115 00:07:36,640 --> 00:07:41,800 So when you generate the output, you see the input that it gives you. 116 00:07:41,800 --> 00:07:45,960 So it just takes the text that you put in the boxes and puts that as a user prompt as 117 00:07:45,960 --> 00:07:49,200 if you've spoke to ChatGPT and you get the answer out from the agent. 118 00:07:49,200 --> 00:07:50,200 Right. 119 00:07:50,200 --> 00:07:56,680 But I did come across a little frustration at the moment with it. 120 00:07:56,680 --> 00:07:58,960 And that's basically, it's kind of a standalone tool. 121 00:07:58,960 --> 00:08:01,920 So it isn't fully integrated into the wider HubSpot suite. 122 00:08:01,920 --> 00:08:08,000 So if you are a HubSpot user already, just be aware that this isn't like you're going 123 00:08:08,000 --> 00:08:12,320 to go into the email section, create a new email, and then within that space, you're 124 00:08:12,320 --> 00:08:16,920 going to be able to generate some text and just populate it into your campaign. 125 00:08:16,920 --> 00:08:17,920 It's standalone. 126 00:08:17,920 --> 00:08:21,800 You just have to copy the text and then paste it into your campaign. 127 00:08:21,800 --> 00:08:26,960 And actually what I found quite frustrating in my use case of it when I was testing it 128 00:08:26,960 --> 00:08:33,480 is I copied the text to add it into a campaign, I dropped it in, and actually the text didn't 129 00:08:33,480 --> 00:08:35,800 add the kind of formatting. 130 00:08:35,800 --> 00:08:39,760 So the markup was included in the copy, if that makes... 131 00:08:39,760 --> 00:08:45,880 So it said headline and then said the headline and it said paragraph one and then added paragraph 132 00:08:45,880 --> 00:08:46,880 one. 133 00:08:46,880 --> 00:08:53,360 So just adding it into the actual content editor within the email interface was just 134 00:08:53,360 --> 00:08:54,360 a bit clunky. 135 00:08:54,360 --> 00:08:56,240 You just had to edit that. 136 00:08:56,240 --> 00:08:59,720 But other than that, I mean, bearing in mind, this is a beta product and it's completely 137 00:08:59,720 --> 00:09:02,680 free and looks like it's going to remain free for everyone. 138 00:09:02,680 --> 00:09:05,280 I think they've done a really solid job. 139 00:09:05,280 --> 00:09:07,520 Yeah, that's awesome. 140 00:09:07,520 --> 00:09:10,040 I had a bit of a play with this in the beta. 141 00:09:10,040 --> 00:09:16,400 I think I was using it to come up with outlines and copy for blog posts, just a couple of 142 00:09:16,400 --> 00:09:18,320 experiments. 143 00:09:18,320 --> 00:09:26,320 And certainly my feeling at the time was, and I think this maybe was pre-GPT4, I can't 144 00:09:26,320 --> 00:09:31,720 remember, but my feeling at the time was I'm a bit surprised at the high quality of both 145 00:09:31,720 --> 00:09:36,760 the outlines and the content I was getting in terms of how informed it was on the topics. 146 00:09:36,760 --> 00:09:42,800 Now, of course, I was testing marketing-based topics for the Byastrata blog and you would 147 00:09:42,800 --> 00:09:47,280 hope that generative tools would know a bit about marketing given their hubspots. 148 00:09:47,280 --> 00:09:54,280 So I don't know whether they've trained their tools on data they've got or other sources. 149 00:09:54,280 --> 00:09:57,920 But at the time, I remember thinking, this is better than I'm currently getting out of 150 00:09:57,920 --> 00:10:01,080 say ChatGPT or other tools. 151 00:10:01,080 --> 00:10:02,320 So that was quite good. 152 00:10:02,320 --> 00:10:07,640 Did you have any experiences where you could sort of ascertain how good is the information 153 00:10:07,640 --> 00:10:09,320 that this is producing? 154 00:10:09,320 --> 00:10:12,120 How well written is it or anything like that? 155 00:10:12,120 --> 00:10:17,000 Now I thought that it did stay quite close to the style. 156 00:10:17,000 --> 00:10:22,960 In a couple of instances, for instance, I actually tried to create a marketing email 157 00:10:22,960 --> 00:10:24,800 promoting the podcast. 158 00:10:24,800 --> 00:10:30,400 And in the input text where I described what I was promoting, I explained the podcast and 159 00:10:30,400 --> 00:10:38,240 I said, I present it with you and I said that this is Paul Avery, CEO of Byastrata. 160 00:10:38,240 --> 00:10:44,800 And then when it wrote the actual copy, it kind of got confused because it was attached 161 00:10:44,800 --> 00:10:49,040 to my company's HubSpot portal. 162 00:10:49,040 --> 00:10:54,920 So it talked about it being from my company, but then talked about our CEO, Paul Avery, 163 00:10:54,920 --> 00:10:56,640 CEO of Byastrata. 164 00:10:56,640 --> 00:11:04,600 So it just kind of got a bit lost with those kind of little minor details. 165 00:11:04,600 --> 00:11:10,120 But other than that, I actually thought the way that it structured the email, so it had 166 00:11:10,120 --> 00:11:13,800 appropriate paragraph length, it had sub headers, it didn't... 167 00:11:13,800 --> 00:11:16,640 Yeah, I thought it did a really good job. 168 00:11:16,640 --> 00:11:18,440 It is powered by OpenAI. 169 00:11:18,440 --> 00:11:26,040 It does have a disclaimer on it saying you have to basically be aware that you're kind 170 00:11:26,040 --> 00:11:31,080 of sending data to them and not to share any sensitive information. 171 00:11:31,080 --> 00:11:36,000 It does explicitly say that in the input prompt. 172 00:11:36,000 --> 00:11:37,480 That's probably useful to know. 173 00:11:37,480 --> 00:11:38,480 Thanks, Martin. 174 00:11:38,480 --> 00:11:39,800 So there you have it. 175 00:11:39,800 --> 00:11:44,880 If you're a HubSpot user or you're in the habit of coming up with business ideas and 176 00:11:44,880 --> 00:11:50,200 throwing up a landing page in HubSpot to see if you can get anybody to sign up or anybody 177 00:11:50,200 --> 00:11:53,920 interested, you can now set up a free HubSpot account and start playing with some of those 178 00:11:53,920 --> 00:11:55,960 tools. 179 00:11:55,960 --> 00:12:00,160 For people that have never used any generative AI, this campaign assistant just makes it 180 00:12:00,160 --> 00:12:01,520 dead simple. 181 00:12:01,520 --> 00:12:04,200 And the fact that it's free is a lovely bonus. 182 00:12:04,200 --> 00:12:05,280 Agreed. 183 00:12:05,280 --> 00:12:06,280 So there you are, people. 184 00:12:06,280 --> 00:12:08,480 Go and have a play. 185 00:12:08,480 --> 00:12:11,520 Let's move on to story number two, Martin. 186 00:12:11,520 --> 00:12:18,720 So this is an interesting one and we're using this as a launchpad for a wider discussion 187 00:12:18,720 --> 00:12:24,840 on the pros and cons and risks and rewards of using generative AI tools. 188 00:12:24,840 --> 00:12:30,000 So I'd be very surprised if anybody hasn't seen this news item now because it blew up 189 00:12:30,000 --> 00:12:34,880 across many different news networks outside of the fact that it was just AI. 190 00:12:34,880 --> 00:12:43,560 But attorney Stephen Schwartz in the States submitted a brief as part of a lawsuit against 191 00:12:43,560 --> 00:12:50,320 Avianca, a Colombian airline, and used ChatGPT to generate the brief. 192 00:12:50,320 --> 00:12:55,100 I've got the SuperWidey report, but I've got the New York Times article here in front 193 00:12:55,100 --> 00:13:02,160 of me at the moment saying that it was a 10-page brief that cited more than a half a dozen 194 00:13:02,160 --> 00:13:08,480 relevant court decisions, but unfortunately, none of them were real. 195 00:13:08,480 --> 00:13:11,920 It's so stupid. 196 00:13:11,920 --> 00:13:15,360 It's just such as, I can't believe you did it. 197 00:13:15,360 --> 00:13:16,680 It's kind of crazy. 198 00:13:16,680 --> 00:13:21,280 So a couple of things here just to let you know that just so that we all realize how 199 00:13:21,280 --> 00:13:24,800 fallible we are and how easily this could happen, right? 200 00:13:24,800 --> 00:13:30,400 So Mr. Schwartz, according to the New York Times, has practiced law in New York for three 201 00:13:30,400 --> 00:13:31,400 decades. 202 00:13:31,400 --> 00:13:35,760 This is not an intern or someone fresh out of grad school. 203 00:13:35,760 --> 00:13:41,640 This is someone who has been playing this game for a very long time and I'm sure understands 204 00:13:41,640 --> 00:13:46,760 the importance of verifiable sources, right? 205 00:13:46,760 --> 00:13:51,880 And in fairness to Mr. Schwartz, he also threw himself, as the New York Times says, on the 206 00:13:51,880 --> 00:13:57,880 mercy of the court on Thursday saying that he used the Artificial Intelligence Program 207 00:13:57,880 --> 00:14:05,000 to do his legal research, in quotes, a source that has revealed itself to be unreliable, 208 00:14:05,000 --> 00:14:09,320 no SH1T, Sherlock. 209 00:14:09,320 --> 00:14:15,660 So I think this just goes to show just how easily you can end up if you just leverage 210 00:14:15,660 --> 00:14:23,280 the stuff that these models push out as fact, you can get yourself into a sticky situation 211 00:14:23,280 --> 00:14:25,640 pretty quickly. 212 00:14:25,640 --> 00:14:28,600 Who could have predicted that that would happen? 213 00:14:28,600 --> 00:14:34,160 It's so stupid, it kind of makes my brain hurt. 214 00:14:34,160 --> 00:14:38,480 But fair play to him for holding his hands up and saying, yep, you caught me. 215 00:14:38,480 --> 00:14:42,600 Can you imagine being the opposition legal team in that instance and you're spending 216 00:14:42,600 --> 00:14:48,840 one at a time crawling and trawling all of the court documents and case law and going, 217 00:14:48,840 --> 00:14:49,840 we just can't find it. 218 00:14:49,840 --> 00:14:52,360 Has he just made it up? 219 00:14:52,360 --> 00:14:54,080 That would be strategic though, wouldn't it? 220 00:14:54,080 --> 00:14:57,680 It's like, I need to keep them busy while I figure out my real argument. 221 00:14:57,680 --> 00:15:02,520 Here, go look up these sources that sound plausible but are not real. 222 00:15:02,520 --> 00:15:08,880 And then that moment when his client must have found out in the courtroom, I'm sorry, 223 00:15:08,880 --> 00:15:09,880 you did what? 224 00:15:09,880 --> 00:15:16,440 I'm paying you how many dollars an hour and you just, oh my dear Lord. 225 00:15:16,440 --> 00:15:20,600 Yeah, so it does go to show that anybody can be duped. 226 00:15:20,600 --> 00:15:28,800 And it goes back to that point about when ChatGPT was first launched, people were saying 227 00:15:28,800 --> 00:15:33,760 that ChatGPT is a bull. 228 00:15:33,760 --> 00:15:36,880 It tells BS. 229 00:15:36,880 --> 00:15:40,000 It's convincing and you read it and you think it's real. 230 00:15:40,000 --> 00:15:50,640 I had an incident recently where I was testing the plugin, the new browsing plugin in ChatGPT 231 00:15:50,640 --> 00:15:54,960 and it was after the Sam Altman Senate hearing. 232 00:15:54,960 --> 00:15:58,880 And one of the clips that I turned or one of the bits of the conversation I wanted to 233 00:15:58,880 --> 00:16:11,560 kind of dial into was about nutritional labeling for AI systems and I asked ChatGPT to explain 234 00:16:11,560 --> 00:16:16,000 this a bit more and I thought, well, it's been mentioned at this hearing so it must 235 00:16:16,000 --> 00:16:19,640 be a topic that a few people have written about. 236 00:16:19,640 --> 00:16:27,480 And it came back to me and said, what you're referring to with nutrition labeling is actually 237 00:16:27,480 --> 00:16:34,000 model cards and model cards were created by Google and they're a way of explaining what 238 00:16:34,000 --> 00:16:37,600 an AI model does, blah, blah, blah, blah, blah. 239 00:16:37,600 --> 00:16:39,960 And I just knew that wasn't true. 240 00:16:39,960 --> 00:16:44,320 But if I didn't know that I would have just taken it as red because actually it sounded 241 00:16:44,320 --> 00:16:48,720 plausible and if you're someone that doesn't follow these things closely. 242 00:16:48,720 --> 00:16:52,600 So I had to say to it, I kind of gave it a telling off and said, no, that's just not 243 00:16:52,600 --> 00:16:53,600 right. 244 00:16:53,600 --> 00:17:01,000 This is a very new concept, but actually what was interesting upon telling it off and telling 245 00:17:01,000 --> 00:17:06,800 it to try again, it must have refined its kind of search parameters because it ended 246 00:17:06,800 --> 00:17:12,080 up coming back saying, oh, this concept was mentioned at this Senate hearing early this 247 00:17:12,080 --> 00:17:13,080 week. 248 00:17:13,080 --> 00:17:15,740 I can't find any other information about it. 249 00:17:15,740 --> 00:17:21,040 This must be a new concept and therefore just kind of watch this space. 250 00:17:21,040 --> 00:17:24,000 It's so interesting, isn't it, that you... 251 00:17:24,000 --> 00:17:29,040 Because I've seen people chastise chat GPT for being wrong and then it's like, oh no, 252 00:17:29,040 --> 00:17:30,040 you're right. 253 00:17:30,040 --> 00:17:31,040 I'm wrong. 254 00:17:31,040 --> 00:17:32,040 Thanks for letting me know. 255 00:17:32,040 --> 00:17:34,720 And it's like, well, if you knew, why did you do it? 256 00:17:34,720 --> 00:17:40,080 And that's because it doesn't know because using the word no is a misnomer. 257 00:17:40,080 --> 00:17:44,000 It's not a correct use of a term here because it's just based on its probability of what 258 00:17:44,000 --> 00:17:46,120 it thinks the next term should be. 259 00:17:46,120 --> 00:17:54,040 Even if it can browse the web, there's going to be this issue is going to be in the system. 260 00:17:54,040 --> 00:17:58,840 In the case of Mr. Schwartz, according to the New York Times article, he told the judge, 261 00:17:58,840 --> 00:18:05,520 Judge Castell, that he asked the program to verify the cases were real and it said yes. 262 00:18:05,520 --> 00:18:06,520 Right? 263 00:18:06,520 --> 00:18:08,760 Well, that's due diligence. 264 00:18:08,760 --> 00:18:11,800 I'm sure you pay them big money for. 265 00:18:11,800 --> 00:18:16,080 So I think in summary, and this kind of reiterates, but it's a perfect real life example of things 266 00:18:16,080 --> 00:18:21,400 that we said on this podcast several times, many times at this point, which is having 267 00:18:21,400 --> 00:18:26,600 an expert in the loop to verify the information that's coming out of these tools is absolutely 268 00:18:26,600 --> 00:18:27,600 critical. 269 00:18:27,600 --> 00:18:34,440 Otherwise, you will make yourself and or your company look like fools and you either need 270 00:18:34,440 --> 00:18:39,980 to understand the topics that you're prompting it about so that you can provide that critical 271 00:18:39,980 --> 00:18:42,280 thinking and analysis to what comes out. 272 00:18:42,280 --> 00:18:46,880 Or you need to be prepared to do rigorous due diligence on quite a lot of stuff that's 273 00:18:46,880 --> 00:18:51,880 coming through from them, but certainly anything that you think sounds a bit dubious. 274 00:18:51,880 --> 00:18:54,000 Yeah. 275 00:18:54,000 --> 00:18:56,040 Right. 276 00:18:56,040 --> 00:19:03,200 Let us move on to the next story then, which in this case is a little perhaps a little 277 00:19:03,200 --> 00:19:10,160 bit of insight into where open AI is going, where GPTs are going. 278 00:19:10,160 --> 00:19:18,800 So this comes from humanloop.com where Raza Habib has documented an interview with Sam 279 00:19:18,800 --> 00:19:25,120 Altman, the CEO of OpenAI and a number of other developers to talk about where OpenAI 280 00:19:25,120 --> 00:19:26,240 is going. 281 00:19:26,240 --> 00:19:30,200 And this article is fascinating because it's got loads of like juicy bits in it. 282 00:19:30,200 --> 00:19:34,480 So we're going to pull some out here, aren't we, Martin, in real time based on the things 283 00:19:34,480 --> 00:19:37,240 that we found the most interesting. 284 00:19:37,240 --> 00:19:39,800 So where do you want to start? 285 00:19:39,800 --> 00:19:41,160 I think start at the top. 286 00:19:41,160 --> 00:19:49,000 So the OpenAI heavily GPT, sorry, GPT, GPU limited at present, which is not a massive 287 00:19:49,000 --> 00:19:50,000 surprise. 288 00:19:50,000 --> 00:19:55,080 We know that the amount of compute power required for these large language models is massive. 289 00:19:55,080 --> 00:19:59,360 It actually links back to what could have been one of the short snippets this week, 290 00:19:59,360 --> 00:20:05,800 which is Nvidia, developers of GPU technology, just about. 291 00:20:05,800 --> 00:20:11,960 I'm not sure if they quite broke the trillion dollar mark this week, but their investors 292 00:20:11,960 --> 00:20:17,320 clearly see a lot of opportunity for them in the AI space at the moment. 293 00:20:17,320 --> 00:20:19,680 So their valuation went through the roof. 294 00:20:19,680 --> 00:20:29,840 But yeah, GPU capacity is what is delaying the short term plans of OpenAI at the moment. 295 00:20:29,840 --> 00:20:33,520 And apparently Sam acknowledged their concern and explained that most of the issue was the 296 00:20:33,520 --> 00:20:39,120 result of GPU shortages, specifically around frustrations that users have when it comes 297 00:20:39,120 --> 00:20:44,320 to the speed of GPT-4 because people get a bit frustrated with that. 298 00:20:44,320 --> 00:20:55,320 But also some of the functionality that was teased when they announced GPT-4, the launch 299 00:20:55,320 --> 00:21:02,300 event and Greg Brockman did the demo with turning the pencil sketch and a notebook into 300 00:21:02,300 --> 00:21:05,160 a real working website. 301 00:21:05,160 --> 00:21:10,840 That multimodality is not coming anytime soon based on what this report says. 302 00:21:10,840 --> 00:21:19,640 We're not going to see that until 2024 simply because they need more GPUs to come online. 303 00:21:19,640 --> 00:21:23,240 Slightly frustrating because there's a couple of projects I've been approached about recently 304 00:21:23,240 --> 00:21:30,840 that would utilize that capability quite beautifully. 305 00:21:30,840 --> 00:21:33,640 But yeah, we must wait. 306 00:21:33,640 --> 00:21:40,160 Who would have thought that GPU access would end up slowing a lot of this stuff down versus 307 00:21:40,160 --> 00:21:44,960 the scramble for regulation if indeed, I mean, this is just one data point, but if indeed 308 00:21:44,960 --> 00:21:49,800 that is proving a major bottleneck, that's quite interesting. 309 00:21:49,800 --> 00:21:53,860 The other thing that we're seeing in the world of open source is people finding clever ways 310 00:21:53,860 --> 00:21:59,320 to make these large language models be able to produce high quality outputs, but without 311 00:21:59,320 --> 00:22:04,400 anywhere near the amount of compute power required, smaller models, fewer parameters 312 00:22:04,400 --> 00:22:05,800 and all those types of things. 313 00:22:05,800 --> 00:22:12,760 So I think it goes to show that there is a number of pressures to try, well, we talked 314 00:22:12,760 --> 00:22:16,460 about water use last week, but costs and all these other factors to actually try and figure 315 00:22:16,460 --> 00:22:22,800 out how to get these models, these tools that we're using to operate without quite so much 316 00:22:22,800 --> 00:22:25,280 compute power required. 317 00:22:25,280 --> 00:22:27,560 So I think we'll see that as well. 318 00:22:27,560 --> 00:22:36,160 Yeah, and attached to that compute power kind of restriction at the moment is also something 319 00:22:36,160 --> 00:22:42,240 that was teased or spoke about when GPT-4 was launched, which was the context window. 320 00:22:42,240 --> 00:22:50,640 So GPT-4 was announced and it had an 8K context window and a 32K context window, but the rollout 321 00:22:50,640 --> 00:22:56,240 of the 32K context via the API has been incredibly slow. 322 00:22:56,240 --> 00:23:03,760 I don't know anyone actually that has got access to that, but they are saying that higher 323 00:23:03,760 --> 00:23:09,600 context windows, as high as 1 million tokens, we've spoken at length on this podcast in 324 00:23:09,600 --> 00:23:17,560 recent weeks about Anthropix 100K context window in Claude, a million tokens, that's 325 00:23:17,560 --> 00:23:19,480 750,000 words. 326 00:23:19,480 --> 00:23:26,520 Only 1 million token context windows are plausible in the near future. 327 00:23:26,520 --> 00:23:32,360 That makes me salivate, but also makes me realise that you would absolutely burn through 328 00:23:32,360 --> 00:23:35,000 cash using that API. 329 00:23:35,000 --> 00:23:37,280 Well, yeah, unless they can get it down. 330 00:23:37,280 --> 00:23:42,340 One of the things in this article that they mentioned that Sam focused on is a cheaper 331 00:23:42,340 --> 00:23:49,880 and faster GPT-4 is more important to them right now than say GPT-5 to try and drive 332 00:23:49,880 --> 00:23:57,080 the cost of these things down so that certainly in terms of API use, they keep the price down 333 00:23:57,080 --> 00:24:00,800 because as you mentioned, 1 million tokens would burn cash. 334 00:24:00,800 --> 00:24:04,480 You'd have to charge quite a lot to offer that as a service. 335 00:24:04,480 --> 00:24:08,360 We were speaking off air, what could you do with a million tokens? 336 00:24:08,360 --> 00:24:17,680 We were like, oh, you could throw the text of 10 marketing books in and from there, ask 337 00:24:17,680 --> 00:24:22,920 it what do these books agree about? 338 00:24:22,920 --> 00:24:23,920 What do they disagree about? 339 00:24:23,920 --> 00:24:25,320 What are the key contention points? 340 00:24:25,320 --> 00:24:27,400 What are the key discussion points? 341 00:24:27,400 --> 00:24:31,640 Imagine doing that with religious text, crumbs, that would be an interesting one, right? 342 00:24:31,640 --> 00:24:35,760 Although probably, I don't know, maybe a lot of religious texts might really stretch the 343 00:24:35,760 --> 00:24:39,840 context window a bit but I think there's a lot of interesting stuff that could come out 344 00:24:39,840 --> 00:24:40,840 the back of that. 345 00:24:40,840 --> 00:24:41,840 Yeah, massively. 346 00:24:41,840 --> 00:24:46,360 I mean, like I say, it gets my juices going. 347 00:24:46,360 --> 00:24:48,120 Yeah, that'd be cool. 348 00:24:48,120 --> 00:24:51,560 I think some other things that Sam talked about that they were aiming for for this year 349 00:24:51,560 --> 00:24:58,320 is what's called in this article a stateful API. 350 00:24:58,320 --> 00:25:04,320 In other words, when you call the API today, you have to keep passing through the same 351 00:25:04,320 --> 00:25:08,200 conversation history and pay for those tokens and in the future, they're hoping for a version 352 00:25:08,200 --> 00:25:11,760 of the API that remembers the conversation history. 353 00:25:11,760 --> 00:25:15,760 These are all things that will just make it easier for people to build tools on top of 354 00:25:15,760 --> 00:25:22,360 it that make for a much more natural conversation style where you can sort of say to whatever 355 00:25:22,360 --> 00:25:25,520 tool you're using, hey, you remember that thing we were talking about last week? 356 00:25:25,520 --> 00:25:27,160 And it'll go, yeah, yeah, do you mean this thing? 357 00:25:27,160 --> 00:25:28,160 Well, yeah, yeah, cool. 358 00:25:28,160 --> 00:25:33,400 I want to go back to that without having to like scroll through a load of chat histories 359 00:25:33,400 --> 00:25:36,760 and obviously ideally connect those chat histories together, I guess. 360 00:25:36,760 --> 00:25:46,100 When you look at the expense of GPT-4 in terms of the cost per token, I think a thousand 361 00:25:46,100 --> 00:25:52,280 tokens is like three cents as an input prompt and six cents on the output. 362 00:25:52,280 --> 00:25:59,360 That's really expensive compared to what we've seen on GPT-3 and 3.5. 363 00:25:59,360 --> 00:26:04,880 So bringing those costs down, if you imagine that chat window, if you've got a long conversation 364 00:26:04,880 --> 00:26:10,200 and every time you called the API, you're pushing all of that through, it does soon 365 00:26:10,200 --> 00:26:11,200 start to rack up. 366 00:26:11,200 --> 00:26:14,160 So I can see why that's going to be a priority for them. 367 00:26:14,160 --> 00:26:16,400 Yeah, I'd agree. 368 00:26:16,400 --> 00:26:19,840 I think the last thing we'll pull from this article that we found interesting, Martin, 369 00:26:19,840 --> 00:26:27,800 was the concept of plugins within ChatGPT not having product market fit yet. 370 00:26:27,800 --> 00:26:34,960 So the article says that the usage of plugins other than browsing doesn't have product market 371 00:26:34,960 --> 00:26:35,960 fit yet. 372 00:26:35,960 --> 00:26:41,160 And he suggested that a lot of people thought they wanted their apps to be inside ChatGPT, 373 00:26:41,160 --> 00:26:44,280 but what they really wanted was ChatGPT inside their apps. 374 00:26:44,280 --> 00:26:50,560 And when you think about how people are using a lot of apps at the moment, that makes sense. 375 00:26:50,560 --> 00:26:57,760 It's almost like ChatGPT is an enabling technology inside of you trying to do something else. 376 00:26:57,760 --> 00:26:58,760 Right? 377 00:26:58,760 --> 00:27:05,240 Even just the HubSpot platform that you talked about at the beginning of this episode was 378 00:27:05,240 --> 00:27:06,240 around a goal. 379 00:27:06,240 --> 00:27:09,680 Like I want to write a landing page, I want to write an email, I want to interrogate my 380 00:27:09,680 --> 00:27:14,760 database of contacts to know who lives in the Cambridge area because I'm going out and 381 00:27:14,760 --> 00:27:17,440 about and I want to go meet with some of them. 382 00:27:17,440 --> 00:27:18,720 That context is key. 383 00:27:18,720 --> 00:27:23,520 And so I think this fits in with some of our experience, this somewhat frustrating experience 384 00:27:23,520 --> 00:27:30,760 is trying to use the plugin library because some of them don't really work very well or 385 00:27:30,760 --> 00:27:33,920 have very specific use cases and other of these types of things. 386 00:27:33,920 --> 00:27:38,800 So it's not necessarily a surprise to hear Sam Altman say that, but I think maybe it 387 00:27:38,800 --> 00:27:45,600 shows you that the whole OpenAI is going to dominate the world by having like the app 388 00:27:45,600 --> 00:27:50,040 marketplace for ChatGPT apps maybe is not the way it's going to go. 389 00:27:50,040 --> 00:27:53,520 Yeah, I'm not sold on the plugins yet. 390 00:27:53,520 --> 00:27:56,000 I haven't found a killer app. 391 00:27:56,000 --> 00:28:00,800 I think the browsing plugin is kind of slow. 392 00:28:00,800 --> 00:28:03,200 Ask your PDF one is pretty good. 393 00:28:03,200 --> 00:28:07,280 And I think there's a video insights one, which does a pretty decent job of pulling 394 00:28:07,280 --> 00:28:09,760 out some insights from YouTube videos and whatnot. 395 00:28:09,760 --> 00:28:13,880 But for the most part, there's actually quite a lot of trash. 396 00:28:13,880 --> 00:28:14,880 I'd agree. 397 00:28:14,880 --> 00:28:19,920 I think the best ones, not entirely, but some of the best ones for me are the ones that 398 00:28:19,920 --> 00:28:24,960 they were really good tools before you could just easily access them in ChatGPT. 399 00:28:24,960 --> 00:28:29,120 So we've featured Ask My PDF on one of the early episodes of the podcast. 400 00:28:29,120 --> 00:28:30,400 That's a pretty cool tool. 401 00:28:30,400 --> 00:28:33,480 And what this just means is I don't have to go to that tool. 402 00:28:33,480 --> 00:28:37,200 I can just access it directly within ChatGPT. 403 00:28:37,200 --> 00:28:42,020 I think there's some cool things you can probably do with Zapier, but sometimes it's maybe just 404 00:28:42,020 --> 00:28:44,760 easier to just copy paste stuff yourself. 405 00:28:44,760 --> 00:28:45,760 Right? 406 00:28:45,760 --> 00:28:50,640 So you can ask ChatGPT to write an email for you based on a brief, and then you can push 407 00:28:50,640 --> 00:28:55,160 it through Zapier into Gmail, or you just copy paste it, right? 408 00:28:55,160 --> 00:28:56,800 Which is pretty fast and easy. 409 00:28:56,800 --> 00:29:01,000 So I'm not super surprised to hear this because I think it fits with our experience, doesn't 410 00:29:01,000 --> 00:29:02,000 it? 411 00:29:02,000 --> 00:29:03,000 Yeah, absolutely. 412 00:29:03,000 --> 00:29:06,520 And the Zapier one is a good example, actually, because I really like the Zapier integration 413 00:29:06,520 --> 00:29:12,480 when I'm doing it through Zapier, not when I'm doing it through ChatGPT. 414 00:29:12,480 --> 00:29:13,480 Yeah. 415 00:29:13,480 --> 00:29:19,400 So I'm usually building multi-step processes in Zapier, and that's where the power is. 416 00:29:19,400 --> 00:29:23,720 I'm like, get this piece of data from here and here, interpret it like this, push it 417 00:29:23,720 --> 00:29:29,840 here, and then do that, which is, at least as I understand it, from my tests beyond the 418 00:29:29,840 --> 00:29:32,840 scope of ChatGPT. 419 00:29:32,840 --> 00:29:36,280 Although I'm sure there's some pretty cool stuff that I haven't been able to do with 420 00:29:36,280 --> 00:29:37,280 it. 421 00:29:37,280 --> 00:29:38,280 Right. 422 00:29:38,280 --> 00:29:39,280 We're going to move on in the interest of time, Martin. 423 00:29:39,280 --> 00:29:45,080 We're going to move on to our last story today, which is a research paper that came out this 424 00:29:45,080 --> 00:29:46,080 week. 425 00:29:46,080 --> 00:29:53,440 We'll only touch on this briefly, where OpenAI had trained a model using human reinforcement-based 426 00:29:53,440 --> 00:30:01,200 learning, but rather than training it by feeding back when it got the right outcome, they actually 427 00:30:01,200 --> 00:30:05,000 molded how it worked based on the reasoning that it took. 428 00:30:05,000 --> 00:30:09,600 So here's a quote from the OpenAI website. 429 00:30:09,600 --> 00:30:13,920 We've trained a model to achieve a new state of the art in mathematical problem solving 430 00:30:13,920 --> 00:30:20,200 by rewarding each correct step of reasoning, process supervision, instead of simply rewarding 431 00:30:20,200 --> 00:30:23,920 the correct final answer, outcome supervision. 432 00:30:23,920 --> 00:30:28,000 In addition to boosting performance relative to outcome supervision, process supervision 433 00:30:28,000 --> 00:30:30,340 also has an important alignment benefit. 434 00:30:30,340 --> 00:30:38,240 It directly trains the model to produce a chain of thought that is endorsed by humans. 435 00:30:38,240 --> 00:30:45,440 I saw this story and we're enthusiasts, we're not machine learning experts, but a lot of 436 00:30:45,440 --> 00:30:52,360 the problems that come up in some of the discussions that we've had over time is, well, ChatGPT 437 00:30:52,360 --> 00:30:56,480 and large language models, they don't think, they just get to an outcome, but in a very 438 00:30:56,480 --> 00:31:02,760 different way from how humans do because it's all based on complex neural architecture, 439 00:31:02,760 --> 00:31:08,120 based on probability matrices and determination parameters and all these other things, and 440 00:31:08,120 --> 00:31:11,480 they don't, in inverted commas, think like us. 441 00:31:11,480 --> 00:31:17,520 But this sounds to me like they're trying to help the models think like us. 442 00:31:17,520 --> 00:31:21,560 They're trying to teach them logical reasoning and things like that, which I can imagine 443 00:31:21,560 --> 00:31:27,640 would be really beneficial in some instances in terms of getting better outcomes, but also, 444 00:31:27,640 --> 00:31:33,840 we've also talked about OpenAI and other large language model developers and providers can't 445 00:31:33,840 --> 00:31:39,400 often describe or explain how the tool got the results. 446 00:31:39,400 --> 00:31:43,760 But that might be easier if you've actually trained things for things to think in a certain 447 00:31:43,760 --> 00:31:46,400 way or to process information in a certain way. 448 00:31:46,400 --> 00:31:48,800 So I just thought that was a cool paper. 449 00:31:48,800 --> 00:31:50,280 Any thoughts on this, Martin? 450 00:31:50,280 --> 00:31:57,080 Yeah, well, chain of thought prompting has been shown to improve outputs for a while 451 00:31:57,080 --> 00:31:58,080 now. 452 00:31:58,080 --> 00:32:02,160 And I think it's really interesting that they're doubling down on the reinforcement learning 453 00:32:02,160 --> 00:32:03,160 of that. 454 00:32:03,160 --> 00:32:08,840 So to improve that as a methodology for coming up with outputs as well. 455 00:32:08,840 --> 00:32:14,160 So yeah, I saw this, they were big on it, OpenAI published it, and Greg Bruckman was 456 00:32:14,160 --> 00:32:16,040 tweeting all about it. 457 00:32:16,040 --> 00:32:21,720 And like you say, this is kind of beyond our area of expertise in terms of machine learning 458 00:32:21,720 --> 00:32:23,800 development. 459 00:32:23,800 --> 00:32:30,480 But yeah, cool to see them doing reinforcement learning in different domains or different 460 00:32:30,480 --> 00:32:31,480 areas. 461 00:32:31,480 --> 00:32:32,480 Absolutely. 462 00:32:32,480 --> 00:32:39,280 And I think as a marketer, if there are particular use cases that these tools are not good at 463 00:32:39,280 --> 00:32:45,680 or untrustworthy in, this might help to see some improvements in terms of improving their 464 00:32:45,680 --> 00:32:52,520 logical reasoning, rather than just potentially regurgitating a remix of information that is 465 00:32:52,520 --> 00:32:57,920 in the training data that helped train the weights in the model to produce a specific 466 00:32:57,920 --> 00:32:58,920 output. 467 00:32:58,920 --> 00:33:04,280 Just as a general note as well, listeners, if they haven't tried chain of thought prompting, 468 00:33:04,280 --> 00:33:06,160 add that into the prompt. 469 00:33:06,160 --> 00:33:11,440 So ask it, if you're asking it to kind of do some planning or come up with a strategy 470 00:33:11,440 --> 00:33:18,640 or do a piece of maths, try saying to it, show your working and show the steps to get 471 00:33:18,640 --> 00:33:23,080 to the output and it will improve the response that you get. 472 00:33:23,080 --> 00:33:24,080 Interesting. 473 00:33:24,080 --> 00:33:25,080 Thanks for sharing that, Martin. 474 00:33:25,080 --> 00:33:27,080 I'm going to have a little play with that. 475 00:33:27,080 --> 00:33:28,080 Right. 476 00:33:28,080 --> 00:33:31,000 Last but not least, let's look at our tool of the week. 477 00:33:31,000 --> 00:33:36,080 It's a quickie because we've not dug into using it too much, but it's been a question 478 00:33:36,080 --> 00:33:41,840 that I've been asked a number of times recently and so it might just be worth flagging in 479 00:33:41,840 --> 00:33:43,640 case people want to go and play with it. 480 00:33:43,640 --> 00:33:51,440 So the tool itself is called recraft.ai and it's another image generation tool, but it's 481 00:33:51,440 --> 00:33:54,360 an image generation tool with a difference. 482 00:33:54,360 --> 00:34:00,280 And the reason it's different is because it produces vector-based images. 483 00:34:00,280 --> 00:34:06,600 So far, when you're using stable diffusion-based models, you get gorgeous images, but they 484 00:34:06,600 --> 00:34:08,880 are JPEGs, PNGs. 485 00:34:08,880 --> 00:34:13,480 They are like your photographs that you would take. 486 00:34:13,480 --> 00:34:17,560 As such, they are limited in terms of their resolution and a number of other factors. 487 00:34:17,560 --> 00:34:22,040 You can't easily change color of specific elements and all those good bits. 488 00:34:22,040 --> 00:34:26,680 What you really want, if you want to be able to scale images infinitely or to be able to 489 00:34:26,680 --> 00:34:32,440 easily change colors and move bits of images around, is to have them as vectors. 490 00:34:32,440 --> 00:34:37,960 And for most marketers, their exposure to vectors is mostly going to be things like 491 00:34:37,960 --> 00:34:45,240 logos, icons, some sort of fairly simplistic, I would have thought compared to photographs, 492 00:34:45,240 --> 00:34:46,240 graphics. 493 00:34:46,240 --> 00:34:51,760 And what recraft.ai allows you to do is it allows you to provide a natural language prompt 494 00:34:51,760 --> 00:34:54,840 like you would for mid-journey or any other similar tool. 495 00:34:54,840 --> 00:35:00,600 But what you can do is produce vectors like icons, colored icons, just like text-based 496 00:35:00,600 --> 00:35:02,920 outlines. 497 00:35:02,920 --> 00:35:09,280 You can ask it to do all of the things that you would imagine, like I was playing with 498 00:35:09,280 --> 00:35:12,960 it trying to create different icons that you might use on a website. 499 00:35:12,960 --> 00:35:17,240 And once it produces them, you can actually go in and edit bits of that vector. 500 00:35:17,240 --> 00:35:23,100 So you can change the colors and things like that, which was pretty cool. 501 00:35:23,100 --> 00:35:31,880 When I tried to get creative and I really tried to get outside of what would be a standard 502 00:35:31,880 --> 00:35:34,240 icon, it did struggle. 503 00:35:34,240 --> 00:35:40,400 So for example, I asked it, six people holding hands in a circle icon to form the shape of 504 00:35:40,400 --> 00:35:41,400 a light bulb. 505 00:35:41,400 --> 00:35:44,000 And it just had no chance. 506 00:35:44,000 --> 00:35:48,900 It did not do well at all, but it did try. 507 00:35:48,900 --> 00:35:57,040 And so I think that for those people who had been trying to get those icon style outputs, 508 00:35:57,040 --> 00:36:01,320 this might be a tool to play with and go and give a try to. 509 00:36:01,320 --> 00:36:04,040 Yeah, I'll definitely check that one out. 510 00:36:04,040 --> 00:36:08,360 It's also one that we're expecting to see as a piece of functionality in Firefly. 511 00:36:08,360 --> 00:36:16,360 If you look at Firefly coming soon, it's one of the tools there, text to vector. 512 00:36:16,360 --> 00:36:20,760 Yeah, because when we were having a play with Firefly, you could do some cool stuff with 513 00:36:20,760 --> 00:36:23,720 their vector tool in the beta, but it was kind of limited. 514 00:36:23,720 --> 00:36:29,160 And I do think when they roll that out into Photoshop, I'm expecting we'll see a huge 515 00:36:29,160 --> 00:36:33,720 amount of extra power, because when you can actually edit those vectors in Adobe Illustrator 516 00:36:33,720 --> 00:36:38,720 and do some of the things I'm talking about here, I think it will be cool. 517 00:36:38,720 --> 00:36:45,200 So if I was going to be brutally honest, I'd say Recraft AI is free at the moment. 518 00:36:45,200 --> 00:36:48,720 You can just create a login and go and have a play with it. 519 00:36:48,720 --> 00:36:51,440 Maybe get some of the stuff that you want. 520 00:36:51,440 --> 00:36:57,640 When this comes to Adobe Suite and you can do this in Illustrator, I think it's probably 521 00:36:57,640 --> 00:37:01,480 going to fit most people's workflows a bit better. 522 00:37:01,480 --> 00:37:04,720 Yeah, that sounds like a fair shout. 523 00:37:04,720 --> 00:37:06,160 But definitely worth a play. 524 00:37:06,160 --> 00:37:10,840 And I was impressed at how easy to use it is. 525 00:37:10,840 --> 00:37:13,920 When you ask for your icon, it's a little bit like Mid Journey. 526 00:37:13,920 --> 00:37:18,560 It gives you four or five options and then you can recraft, which is basically, I didn't 527 00:37:18,560 --> 00:37:22,120 like the ones I got, go and have another go. 528 00:37:22,120 --> 00:37:26,720 And yeah, I think certainly for graphic designers, I'd be having a play and see if it can speed 529 00:37:26,720 --> 00:37:28,600 up your workflow. 530 00:37:28,600 --> 00:37:33,160 And for small marketing teams, solo marketers that want to see if they can just produce 531 00:37:33,160 --> 00:37:37,720 some novel icons quickly without having to go out to a designer, maybe it will help you 532 00:37:37,720 --> 00:37:39,680 get some of the way that you need to go. 533 00:37:39,680 --> 00:37:44,240 I haven't played with it enough to say how reliable the output is, but I think it's worth 534 00:37:44,240 --> 00:37:45,240 having a play with. 535 00:37:45,240 --> 00:37:50,120 Right, I think we'll leave it there, shall we, Martin, for another week? 536 00:37:50,120 --> 00:37:51,880 Yeah, I think so. 537 00:37:51,880 --> 00:37:59,960 I can just leave now and go and be sad about the series succession having come to an end. 538 00:37:59,960 --> 00:38:04,480 I've been really enjoying that in recent weeks, but I have to find something new to watch 539 00:38:04,480 --> 00:38:05,480 now. 540 00:38:05,480 --> 00:38:09,640 Do you know, if it's not Derby County bringing you down now as its succession, have you seen 541 00:38:09,640 --> 00:38:13,040 For All Mankind on the Apple TVs? 542 00:38:13,040 --> 00:38:15,560 I have not. 543 00:38:15,560 --> 00:38:18,040 I didn't think I was going to like it. 544 00:38:18,040 --> 00:38:23,680 This is not what this podcast is about, apologies everyone, but it's an alternate history as 545 00:38:23,680 --> 00:38:28,280 if the Russians got to the moon first and the space race that kicked off. 546 00:38:28,280 --> 00:38:32,760 So it starts off at a point in time that's the same, if you like, as our history and 547 00:38:32,760 --> 00:38:35,640 then it quickly diverges and then goes through time. 548 00:38:35,640 --> 00:38:39,920 I'm only like six episodes in, but it's been really fascinating. 549 00:38:39,920 --> 00:38:40,920 That's worth having a little look at. 550 00:38:40,920 --> 00:38:42,200 And what's that called? 551 00:38:42,200 --> 00:38:43,880 It's called For All Mankind, I think. 552 00:38:43,880 --> 00:38:44,880 For All Mankind. 553 00:38:44,880 --> 00:38:46,560 Cool, I'll check it out. 554 00:38:46,560 --> 00:38:49,160 Right then, thanks always, Martin, for your time. 555 00:38:49,160 --> 00:38:52,040 Always lovely and I will catch you next week. 556 00:38:52,040 --> 00:38:54,080 Yeah, see you next week. 557 00:38:54,080 --> 00:38:56,080 Thanks for listening, everyone. 558 00:38:56,080 --> 00:39:05,000 Follow us, aim marketing pod on Twitter and head to the website artificiallyintelligentmarketing.com 559 00:39:05,000 --> 00:39:08,560 and share your comments, give us feedback, send us your voice notes on WhatsApp, which 560 00:39:08,560 --> 00:39:11,400 we haven't given you a number for, but we will do one day. 561 00:39:11,400 --> 00:39:15,360 Yeah, that about sums it up for me. 562 00:39:15,360 --> 00:39:17,400 Right, have a lovely weekend, everyone. 563 00:39:17,400 --> 00:39:20,120 Cheers, bye. 564 00:39:20,120 --> 00:39:23,560 Thank you for listening to artificially intelligent marketing. 565 00:39:23,560 --> 00:39:29,640 To stay on top of the latest trends, tips and tools in the world of marketing AI, be 566 00:39:29,640 --> 00:39:31,380 sure to subscribe. 567 00:39:31,380 --> 00:39:34,960 We look forward to seeing you again next week.