WEBVTT 00:00:15.679 --> 00:00:18.399 Welcome to artificially intelligent marketing, 00:00:18.539 --> 00:00:21.519 a weekly podcast where we stay on top of the 00:00:21.519 --> 00:00:24.800 latest trends, tips, and tools in the world of 00:00:24.800 --> 00:00:27.620 marketing AI, helping you get the best results 00:00:27.620 --> 00:00:30.780 from your marketing efforts. Now let's join our 00:00:30.780 --> 00:00:34.520 hosts, Paul Avery and Martin Broadhurst. Hello 00:00:34.520 --> 00:00:36.920 and welcome to Artificially Intelligent Marketing. 00:00:37.340 --> 00:00:41.560 You remember us from 1991 or whether it was that 00:00:41.560 --> 00:00:43.740 we last published an episode, but I'm back today 00:00:43.740 --> 00:00:46.960 with my fantastic colleague, Martin Broadhurst. 00:00:46.979 --> 00:00:49.380 Martin, how are you, sir? Sorry, who are you? 00:00:49.799 --> 00:00:53.060 Is this a spam call? I don't even know who I 00:00:53.060 --> 00:00:56.380 am anymore. Hello. Yes. Nice to be back, isn't 00:00:56.380 --> 00:00:58.899 it? It is. What have you been up to? How long 00:00:58.899 --> 00:01:02.299 have we been offline? Maybe not offline. Come 00:01:02.299 --> 00:01:05.659 on, we're all attached to our phones 24 -7. Off 00:01:05.659 --> 00:01:08.579 air. How long have we been off air? It's over 00:01:08.579 --> 00:01:14.900 a year. Oh, I think our not particularly large 00:01:14.900 --> 00:01:17.519 subscriber base have probably fled. They're probably 00:01:17.519 --> 00:01:19.120 somewhere else. But hopefully you've still got 00:01:19.120 --> 00:01:22.000 this. on your podcast app and this lands in and 00:01:22.000 --> 00:01:23.859 you're like oh good i'm glad the gent's back 00:01:23.859 --> 00:01:26.239 i really want to know what they feel about what's 00:01:26.239 --> 00:01:29.180 going on um i have got a pretty good excuse i 00:01:29.180 --> 00:01:32.480 think for not being around martin um those of 00:01:32.480 --> 00:01:34.939 you who listen to the podcast that know me will 00:01:34.939 --> 00:01:37.659 probably already know this but um i was co -founder 00:01:37.659 --> 00:01:42.500 of an agency and we were acquired last year and 00:01:42.500 --> 00:01:44.219 we were going through the acquisition process 00:01:44.219 --> 00:01:47.379 and it takes a huge amount of time and effort 00:01:48.020 --> 00:01:51.620 And then that gobbled up all my time. And yeah, 00:01:51.659 --> 00:01:53.280 I couldn't really invest too much time in anything 00:01:53.280 --> 00:01:55.540 else, to be honest. Hence having to take a hiatus 00:01:55.540 --> 00:01:58.700 that probably should have been two months, but 00:01:58.700 --> 00:02:01.480 then turned into like 18 months. So sorry about 00:02:01.480 --> 00:02:03.280 that, Martin. Martin's been messaging me. He's 00:02:03.280 --> 00:02:04.819 been like, dude, we've got to do the podcast. 00:02:04.980 --> 00:02:06.599 We've got to do the podcast. I'm like, yeah, 00:02:06.719 --> 00:02:09.340 no, I will. I'll get back to it. But well, how 00:02:09.340 --> 00:02:10.759 have you been filling your time, Martin? Because 00:02:10.759 --> 00:02:14.090 obviously I know this is all you live for. Hanging 00:02:14.090 --> 00:02:15.949 out with me on here. So what have you been filling 00:02:15.949 --> 00:02:19.849 your time with? I've been sat here. Sat here 00:02:19.849 --> 00:02:24.909 waiting. Just clicking, going, he'll be back 00:02:24.909 --> 00:02:28.349 soon. I believe in Paul. And now you're back. 00:02:28.669 --> 00:02:30.509 What have I, in all seriousness, what have I 00:02:30.509 --> 00:02:33.689 been up to? Well, there's been a huge amount 00:02:33.689 --> 00:02:37.090 of delivering workshopping and training and implementing 00:02:37.090 --> 00:02:42.110 AI, unsurprisingly, alongside the HubSpot delivery 00:02:42.110 --> 00:02:47.319 as well. uh really wanting to see their systems 00:02:47.319 --> 00:02:49.419 speak to one another and i think that's going 00:02:49.419 --> 00:02:52.080 to be a theme of the episode today to some extent 00:02:52.080 --> 00:02:56.319 but uh yeah it's the adoption and the application 00:02:56.319 --> 00:03:01.439 of ai is is really becoming front and center 00:03:01.439 --> 00:03:05.219 of lots of leadership teams conversations now 00:03:05.219 --> 00:03:09.060 hey man well that feels like quite a nice little 00:03:09.060 --> 00:03:11.219 segue into what we're going to cover today because 00:03:11.219 --> 00:03:13.849 we have been around for a while We were thinking 00:03:13.849 --> 00:03:16.030 that we'll go through all the things that have 00:03:16.030 --> 00:03:18.289 happened in the last 18 months, not in detail, 00:03:18.349 --> 00:03:21.050 because that would be extremely boring and very 00:03:21.050 --> 00:03:22.909 lengthy. But we're going to touch on some of 00:03:22.909 --> 00:03:25.090 the evolutions that have happened in the world 00:03:25.090 --> 00:03:28.990 of AI, especially for marketers, with a focus 00:03:28.990 --> 00:03:31.530 on where are we now? What can we do that we couldn't 00:03:31.530 --> 00:03:34.430 do before? What is working well for marketers? 00:03:34.509 --> 00:03:37.569 What is still over -promising and under -delivering 00:03:37.569 --> 00:03:40.460 on the AI front? just to bring you up to speed 00:03:40.460 --> 00:03:42.960 if you haven't been paying attention to the world 00:03:42.960 --> 00:03:45.219 of AI and marketing, which I'm sure, honestly, 00:03:45.400 --> 00:03:48.060 you have, because since we started the podcast 00:03:48.060 --> 00:03:51.939 two and a half years ago, probably more, you 00:03:51.939 --> 00:03:54.340 can't really get away from AI in the world of 00:03:54.340 --> 00:03:56.120 anything these days, can you, Martin? So I'm 00:03:56.120 --> 00:03:58.240 sure most people are paying attention even if 00:03:58.240 --> 00:04:01.199 they don't want to. yeah certainly uh copilot 00:04:01.199 --> 00:04:04.599 is plastered across every microsoft surface you 00:04:04.599 --> 00:04:07.659 can possibly find gemini's creeped its way into 00:04:07.659 --> 00:04:10.960 every google application you care to use and 00:04:10.960 --> 00:04:13.780 even ones that you don't yeah and then we've 00:04:13.780 --> 00:04:17.860 got tv ads for chat gpt and google gemini and 00:04:17.860 --> 00:04:21.399 smart ai driven features on your mobile phone 00:04:21.399 --> 00:04:23.759 to improve taking your pictures and yeah it's 00:04:23.759 --> 00:04:28.259 everywhere we're covered in it um rather than 00:04:28.259 --> 00:04:31.660 being covered in it let's cover it in some detail 00:04:31.660 --> 00:04:36.379 on the podcast today so um one of the first things 00:04:36.379 --> 00:04:38.360 we should probably touch on is i don't think 00:04:38.360 --> 00:04:41.959 ai reasoning was a thing when we last did our 00:04:41.959 --> 00:04:44.779 episode was it mine it wasn't the nearest we 00:04:44.779 --> 00:04:47.060 had to it at that time was the prompting technique 00:04:47.060 --> 00:04:49.639 chain of thought prompting where you would ask 00:04:49.639 --> 00:04:54.240 the ai to show it's working and think step by 00:04:54.240 --> 00:04:56.860 step before delivering the answer which would 00:04:57.269 --> 00:05:00.490 improve the the outputs but no that has evolved 00:05:00.490 --> 00:05:08.149 into a new model form yeah i mean i think you're 00:05:08.149 --> 00:05:10.930 right it kind of took that chain of thought approach 00:05:10.930 --> 00:05:15.189 and turned it up to 11 and it just vastly improves 00:05:15.189 --> 00:05:18.870 the outputs especially for more strategic questions 00:05:18.870 --> 00:05:21.850 or challenging questions where the ai needs to 00:05:21.850 --> 00:05:25.250 think and plan a bit it also helps massively 00:05:25.250 --> 00:05:28.589 reduce hallucinations and other factors. Don't 00:05:28.589 --> 00:05:31.009 get rid of them, unfortunately. No, sadly not. 00:05:31.470 --> 00:05:35.029 But I think it does a better job of it. So yeah, 00:05:35.110 --> 00:05:37.310 if you haven't been using reasoning models, you 00:05:37.310 --> 00:05:38.709 should probably be giving them a try. You should 00:05:38.709 --> 00:05:40.670 probably get better outputs. It's kind of hard 00:05:40.670 --> 00:05:42.769 to escape them now, given that Google Gemini 00:05:42.769 --> 00:05:46.009 sort of uses its reasoning capabilities automatically 00:05:46.009 --> 00:05:50.610 when it feels it needs them, as does ChatGPT. 00:05:50.670 --> 00:05:52.589 But we'll talk about that a little bit later, 00:05:52.670 --> 00:05:54.329 because I know you've got a bee in your bonnet 00:05:54.329 --> 00:05:58.870 about... auto routing of uh of models yeah i 00:05:58.870 --> 00:06:00.930 mean there is a downside to the to the reasoning 00:06:00.930 --> 00:06:02.870 in terms of speed of response so the latency 00:06:02.870 --> 00:06:05.589 obviously downgrades because before you get your 00:06:05.589 --> 00:06:07.870 response the model thinks through interestingly 00:06:07.870 --> 00:06:09.649 it does give you a description of what it's doing 00:06:09.649 --> 00:06:13.649 which i always enjoy joy reading through but 00:06:13.649 --> 00:06:16.189 yeah i think this is this is a marked improvement 00:06:16.189 --> 00:06:18.829 on the the capabilities the difference in in 00:06:18.829 --> 00:06:23.029 the quality of the output as you say is is vast 00:06:23.029 --> 00:06:26.519 and all of the model providers are now providing 00:06:26.519 --> 00:06:29.980 a whether it's called with extended thinking 00:06:29.980 --> 00:06:34.819 or gemini has got uh gemini 2 .5 pro which is 00:06:34.819 --> 00:06:38.579 reasoning by default as opposed to 2 .5 flat 00:06:38.579 --> 00:06:41.879 which is the kind of quick response model um 00:06:41.879 --> 00:06:45.620 and just to kind of articulate this for the audience 00:06:45.620 --> 00:06:48.699 that might not be fully familiar with kind of 00:06:48.699 --> 00:06:51.959 how or why this this works the example i often 00:06:51.959 --> 00:06:58.730 give is If I ask you to work out 86 times by 00:06:58.730 --> 00:07:03.350 32, you're probably not going to give me an answer 00:07:03.350 --> 00:07:06.709 straight away off the top of your head. And you 00:07:06.709 --> 00:07:09.470 could approximate and give me a ballpark figure, 00:07:09.529 --> 00:07:12.170 but likely most people aren't going to just spit 00:07:12.170 --> 00:07:16.490 out the right answer first time. Whereas pen 00:07:16.490 --> 00:07:19.649 and paper, given a bit of time to write it down 00:07:19.649 --> 00:07:22.819 and plan all the steps and figure it out. you 00:07:22.819 --> 00:07:25.720 can arrive at the right answer and it's the same 00:07:25.720 --> 00:07:28.019 thing with these models the models that are non 00:07:28.019 --> 00:07:30.959 -reasoning non -thinking models they give a response 00:07:30.959 --> 00:07:33.899 kind of quickly they just say the first thing 00:07:33.899 --> 00:07:38.000 that comes to their mind so to speak whereas 00:07:38.000 --> 00:07:41.139 the reasoning it takes the time to figure it 00:07:41.139 --> 00:07:43.959 out yes interesting there's going to be a lot 00:07:43.959 --> 00:07:47.120 of tangents in today's podcast folks so hopefully 00:07:47.120 --> 00:07:49.040 they'll all be valuable but i was listening to 00:07:49.040 --> 00:07:52.540 a podcast yesterday with the ceo of grok this 00:07:52.540 --> 00:07:56.439 is not elon musk's ai model grok this is grok 00:07:56.439 --> 00:07:59.480 the chip technology that's been built specifically 00:07:59.480 --> 00:08:02.819 for inference which is the process that happens 00:08:02.819 --> 00:08:04.800 when you ask the model a question there's ai 00:08:04.800 --> 00:08:07.579 training which has a particular sort of infrastructure 00:08:07.579 --> 00:08:10.639 and sort of the way it works and then there's 00:08:10.639 --> 00:08:13.579 um inference which is you know works a bit differently 00:08:13.579 --> 00:08:17.259 um So he's got skin in the game on this one. 00:08:17.319 --> 00:08:20.500 He's been building chips with his team to be 00:08:20.500 --> 00:08:24.459 really fast at giving you answers. But he's also 00:08:24.459 --> 00:08:27.399 quite credible. He did build the TPUs at Google, 00:08:27.480 --> 00:08:29.639 so he's got a track record. He was saying yesterday, 00:08:29.860 --> 00:08:35.620 when I was listening to this, that speed of answer 00:08:35.620 --> 00:08:39.629 is critical. And at Google, they... I don't know 00:08:39.629 --> 00:08:41.470 if they showed it with data or whether it was 00:08:41.470 --> 00:08:43.110 sort of like a rule of thumb they used, but for 00:08:43.110 --> 00:08:45.330 every hundred milliseconds faster, they could 00:08:45.330 --> 00:08:47.970 get a person an answer translated to 8 % increase 00:08:47.970 --> 00:08:52.929 in market share. Wow. So he's massively betting 00:08:52.929 --> 00:08:56.210 on the fact that as consumers, we won't be willing 00:08:56.210 --> 00:08:58.710 to wait to get our answer. And whilst it's fantastic 00:08:58.710 --> 00:09:01.110 to have access to these reasoning models, like 00:09:01.110 --> 00:09:04.870 if you do a deep research. which might not have 00:09:04.870 --> 00:09:08.370 existed either. I can't remember. So deep research 00:09:08.370 --> 00:09:10.929 for listeners is when you have to specifically 00:09:10.929 --> 00:09:13.029 activate this in ChatGPT, but you can basically 00:09:13.029 --> 00:09:15.009 give it a research project and it will go read 00:09:15.009 --> 00:09:16.970 like 100 sources on the internet and come back 00:09:16.970 --> 00:09:19.929 and give you a 5 ,000 word, 10 ,000 word report 00:09:19.929 --> 00:09:22.769 on a topic. But you set that running, then you 00:09:22.769 --> 00:09:24.230 have to go make yourself a cup of tea because 00:09:24.230 --> 00:09:26.850 that takes like 25, 30 minutes. And I guess what 00:09:26.850 --> 00:09:32.080 he's thinking is we want things quickly. for 00:09:32.080 --> 00:09:34.059 the model companies that can crack speed actually 00:09:34.059 --> 00:09:37.019 and you can use reasoning you can use deep research 00:09:37.019 --> 00:09:40.059 and still get an answer in near real time we're 00:09:40.059 --> 00:09:42.360 going to switch to those models or at least to 00:09:42.360 --> 00:09:44.399 those sort of inference providers which I think 00:09:44.399 --> 00:09:46.299 is going to be quite interesting to see how that 00:09:46.299 --> 00:09:47.840 plays out because I don't know about you Martin 00:09:47.840 --> 00:09:50.960 but as it stands I'm quite tolerant with letting 00:09:50.960 --> 00:09:53.360 the tool work for 20 minutes on a project that 00:09:53.360 --> 00:09:56.159 would have taken me three hours but if a tool 00:09:56.159 --> 00:09:57.840 came out that would deliver it in 30 seconds 00:09:57.840 --> 00:10:00.659 and the results were nearly or as good I would 00:10:00.909 --> 00:10:03.750 probably switch yeah and it's the biggest frustration 00:10:03.750 --> 00:10:08.610 with turning on thinking in chat gpt 5 you want 00:10:08.610 --> 00:10:12.210 it to do a bit of thinking but then you're waiting 00:10:12.210 --> 00:10:14.610 for two minutes and actually if you could compress 00:10:14.610 --> 00:10:17.789 that down to 10 seconds that would be fantastic 00:10:17.789 --> 00:10:21.029 i must say just on the grok thing if anyone is 00:10:21.029 --> 00:10:23.789 kind of vibe coding and looking for an llm integration 00:10:23.789 --> 00:10:29.039 my word grok is fantastic i created a little 00:10:29.039 --> 00:10:32.379 application yesterday that was giving real -time 00:10:32.379 --> 00:10:35.220 responses to a little game that i created so 00:10:35.220 --> 00:10:37.179 it was like running commentary on the outputs 00:10:37.179 --> 00:10:41.179 of the game and because it can deliver 500 tokens 00:10:41.179 --> 00:10:47.179 in a second which is like 450 400 words in a 00:10:47.179 --> 00:10:50.820 second it's amazing like the real time back and 00:10:50.820 --> 00:10:55.659 forth um it's kind of unparalleled Yeah, I think 00:10:55.659 --> 00:10:58.799 real time's critical. As you know, Martin, I 00:10:58.799 --> 00:11:02.500 hate type these days. Listeners don't know, I 00:11:02.500 --> 00:11:05.600 broke my collarbone about a year ago and I couldn't 00:11:05.600 --> 00:11:08.200 type. And then that was when I went all in on 00:11:08.200 --> 00:11:11.279 dictating everything. And I've got a piece of 00:11:11.279 --> 00:11:15.679 software on my computer that uses local models. 00:11:15.879 --> 00:11:19.320 So all my transcription is done on device. So 00:11:19.320 --> 00:11:21.139 I can say what I like. It's not, you know, I'm 00:11:21.139 --> 00:11:22.940 not sharing any sensitive materials into the 00:11:22.940 --> 00:11:26.820 cloud. the provider of this software actually 00:11:26.820 --> 00:11:28.279 i want to i want to give my shout out because 00:11:28.279 --> 00:11:30.240 it's pretty cool it's called voicing it's for 00:11:30.240 --> 00:11:33.000 mac only unfortunately but we now have access 00:11:33.000 --> 00:11:36.919 to the nvidia transcription model which even 00:11:36.919 --> 00:11:40.100 if i speak like two paragraphs they appear instantly 00:11:40.100 --> 00:11:43.220 and that's being done on my device that's not 00:11:43.220 --> 00:11:47.159 being done in the cloud insane um i do think 00:11:47.159 --> 00:11:49.399 we'll get to a point where we can do fairly impressive 00:11:49.399 --> 00:11:54.059 things on a consumer laptop and that you don't 00:11:54.320 --> 00:11:56.840 actually need to be connected to the cloud which 00:11:56.840 --> 00:11:58.360 will be important for me because when i'm on 00:11:58.360 --> 00:12:00.240 a plane and i'm working i can't do any work now 00:12:00.240 --> 00:12:04.080 if there's no ai and wi -fi and i can't think 00:12:04.080 --> 00:12:07.000 for myself i refuse um so i will need that that 00:12:07.000 --> 00:12:10.399 onboard model but um so outside of reasoning 00:12:10.399 --> 00:12:13.740 some of the platforms have matured martin um 00:12:13.740 --> 00:12:16.820 especially copilot that we used to like absolutely 00:12:16.820 --> 00:12:20.440 kick the crap out of back in the day um but you 00:12:20.440 --> 00:12:23.779 quite like it now I do. They've won me over. 00:12:24.059 --> 00:12:27.059 Microsoft, who knew with all of their trillions 00:12:27.059 --> 00:12:30.220 in market cap, they could crack this particularly 00:12:30.220 --> 00:12:35.419 tricky nut. It used to be terrible. The deployment 00:12:35.419 --> 00:12:38.679 of it across the Office suite was just terrible. 00:12:38.759 --> 00:12:41.679 Trying to use it in Excel, you would ask it to 00:12:41.679 --> 00:12:43.200 do something, it would just break or it would 00:12:43.200 --> 00:12:44.279 tell you that it had done it or it would say, 00:12:44.299 --> 00:12:48.500 oh, I can't do that. Whereas now it is really 00:12:48.500 --> 00:12:52.080 quite functional and... quite impressive in some 00:12:52.080 --> 00:12:54.559 areas i think what they've done is they've taken 00:12:54.559 --> 00:12:59.460 the the base models and they understand the business 00:12:59.460 --> 00:13:01.740 applications and what people actually want to 00:13:01.740 --> 00:13:06.379 use them for so you mentioned um deep research 00:13:06.379 --> 00:13:11.720 now deep research is it takes the agentic models 00:13:11.720 --> 00:13:15.659 the reasoning models and applies this agentic 00:13:15.659 --> 00:13:18.750 layer with tool use where deep research can go 00:13:18.750 --> 00:13:21.789 on the web it can read documents it it compiles 00:13:21.789 --> 00:13:25.330 all of this information and then writes the the 00:13:25.330 --> 00:13:29.450 report well researcher when i've done business 00:13:29.450 --> 00:13:33.509 related research as opposed to say holiday research 00:13:33.509 --> 00:13:36.649 or something like that the researcher agent in 00:13:36.649 --> 00:13:41.029 copilot versus deep research in gemini or deep 00:13:41.029 --> 00:13:47.009 research in open ai is markedly better and The 00:13:47.009 --> 00:13:50.230 thing that really differentiates it is it's attached 00:13:50.230 --> 00:13:53.009 to your Microsoft tenant. So everything that's 00:13:53.009 --> 00:13:56.169 in your SharePoint, your OneDrive, it can crawl 00:13:56.169 --> 00:13:58.269 all of that information as well and bring that 00:13:58.269 --> 00:14:00.450 back into the report. I mean, that's just one 00:14:00.450 --> 00:14:02.610 example. And there are many more. There's the 00:14:02.610 --> 00:14:05.509 analyst tool, which is fantastic. I think the 00:14:05.509 --> 00:14:12.230 Copilot Studio capabilities are, they're very 00:14:12.230 --> 00:14:14.269 functional, very easy to use, much like creating 00:14:14.269 --> 00:14:18.299 custom GPTs. If anything, I would say the issue 00:14:18.299 --> 00:14:22.240 that Microsoft have got at the moment is on the 00:14:22.240 --> 00:14:24.620 kind of commercial and the marketing side. I 00:14:24.620 --> 00:14:26.440 think it's a bit confusing for people to understand 00:14:26.440 --> 00:14:29.940 the pricing of it. Understanding, okay, there's 00:14:29.940 --> 00:14:33.039 Copilot Chat, which is free. Then there's Copilot 00:14:33.039 --> 00:14:37.659 Pro, which is 20 odd quid per user per month. 00:14:38.600 --> 00:14:40.679 What's the difference between those? I don't 00:14:40.679 --> 00:14:42.620 think they've been brilliant at communicating 00:14:42.620 --> 00:14:45.059 those differences. But then when you come to 00:14:45.059 --> 00:14:48.139 use agents, some of it's pay as you go. Some 00:14:48.139 --> 00:14:52.340 of it's free. And when I speak to businesses, 00:14:52.779 --> 00:14:55.360 they're currently in a bit of a muddle about 00:14:55.360 --> 00:14:58.820 the value of Copilot and just exactly how much 00:14:58.820 --> 00:15:03.440 it's going to cost and what the ROI is and that 00:15:03.440 --> 00:15:05.980 kind of thing. Cheeky plug for Martin there. 00:15:06.019 --> 00:15:08.639 Then if you are using Copilot and you think, 00:15:08.659 --> 00:15:10.379 no, it sucks, Martin, you're wrong. Maybe Martin 00:15:10.379 --> 00:15:12.679 can help you unlock the value of it. But I've 00:15:12.679 --> 00:15:14.659 got a question for you, Martin. I don't use Copilot 00:15:14.659 --> 00:15:19.799 at all. What's Analyst do? Analyst is, as the 00:15:19.799 --> 00:15:23.500 name suggests, is a data analyst agent. So you 00:15:23.500 --> 00:15:26.240 have to have a pro license to access this. So 00:15:26.240 --> 00:15:28.340 you can't be using this on the free version of 00:15:28.340 --> 00:15:31.340 Copilot. But in the same way that Researcher 00:15:31.340 --> 00:15:35.700 is using the reasoning models, with this agentic 00:15:35.700 --> 00:15:37.659 capability to go away and do a research brief 00:15:37.659 --> 00:15:40.220 analyst does the same thing but with data analysis 00:15:40.220 --> 00:15:44.100 so they've given it tooling and they've trained 00:15:44.100 --> 00:15:48.179 it specifically to be a data analyst and it does 00:15:48.179 --> 00:15:50.279 a very good job again i actually if i'm doing 00:15:50.279 --> 00:15:54.379 any um if i've got some documents or files that 00:15:54.379 --> 00:15:57.419 i want to run some analysis on i will do that 00:15:57.419 --> 00:16:02.929 in the analyst tool or analyst agent rather than 00:16:02.929 --> 00:16:07.889 doing it in ChatGPT with extended thinking turned 00:16:07.889 --> 00:16:09.889 on because the outputs are just better. They 00:16:09.889 --> 00:16:11.769 seem to have fine -tuned it for that analyst 00:16:11.769 --> 00:16:15.549 role, and that's what makes it a bit different. 00:16:15.909 --> 00:16:18.710 So it's using data, Excel spreadsheets and stuff. 00:16:19.830 --> 00:16:22.289 Yeah, you can provide the data source, you can 00:16:22.289 --> 00:16:25.490 upload files to it, or you can tell it to go 00:16:25.490 --> 00:16:28.610 and find resources on SharePoint. You can point 00:16:28.610 --> 00:16:30.980 it to the... to the location that you want and 00:16:30.980 --> 00:16:34.399 it will go away and and kind of plug it all together 00:16:34.399 --> 00:16:36.919 and give you the report that you're after so 00:16:36.919 --> 00:16:38.779 there you go but probably many listeners have 00:16:38.779 --> 00:16:41.820 to use copilot because their organization won't 00:16:41.820 --> 00:16:45.120 allow them to use anything else but um if you 00:16:45.120 --> 00:16:47.639 have not used it for a while sounds like it used 00:16:47.639 --> 00:16:50.860 to be terrible and is now quite good um so go 00:16:50.860 --> 00:16:53.899 and give it a go um one of the other things we 00:16:53.899 --> 00:16:56.799 still get asked a lot about martin is which model 00:16:56.799 --> 00:17:00.159 is best For what? And that continues to be an 00:17:00.159 --> 00:17:02.220 absolute pain in the bottom to answer that question 00:17:02.220 --> 00:17:06.559 because of the model roundabout. Obviously, Copilot 00:17:06.559 --> 00:17:09.339 got better. ChatGPT's improved. Anthropic Claud 00:17:09.339 --> 00:17:11.299 has improved. There's been open source models. 00:17:11.900 --> 00:17:15.480 Talk us through the model roundabout over the 00:17:15.480 --> 00:17:19.400 last couple of months. So we were all waiting 00:17:19.400 --> 00:17:24.500 for so long for GPT -5. But actually, what was 00:17:24.500 --> 00:17:26.759 kind of king of the hill for a long time was... 00:17:27.759 --> 00:17:34.819 flawed three and 3 .5 and then flawed four came 00:17:34.819 --> 00:17:37.119 out and that was that was dominating and everyone's 00:17:37.119 --> 00:17:39.380 going when's gpt5 coming out when's gpt5 coming 00:17:39.380 --> 00:17:43.460 out then it came out in august um and that became 00:17:43.460 --> 00:17:47.859 top of the heap and then google gemini came out 00:17:47.859 --> 00:17:52.480 with their uh they've got their what's that top 00:17:52.480 --> 00:17:56.619 of the range model is google gemini 2 .5 deep 00:17:56.619 --> 00:18:00.799 think yes which is the one that won the um international 00:18:00.799 --> 00:18:06.000 maths olympiad alongside chat gpt's latest pro 00:18:06.000 --> 00:18:09.920 model as well um so yeah they're all much of 00:18:09.920 --> 00:18:11.740 a muchness at this point i think it's really 00:18:11.740 --> 00:18:15.279 hard to say that any of them is far and away 00:18:15.279 --> 00:18:20.220 better than any other if you're a developer you're 00:18:20.220 --> 00:18:24.259 probably more inclined to go with claude But 00:18:24.259 --> 00:18:27.059 then I see lots of chat with people saying chat 00:18:27.059 --> 00:18:32.140 GPT with GPT -5 extended thinking and pro. I 00:18:32.140 --> 00:18:35.599 mean, pro is fantastic. If you haven't used GPT 00:18:35.599 --> 00:18:42.559 -5 pro to one shot a bit of vibe coding magic, 00:18:42.940 --> 00:18:45.980 you haven't lived, Paul, if you've not done that. 00:18:46.519 --> 00:18:48.819 I've dabbled. I think it's really interesting 00:18:48.819 --> 00:18:51.200 as you were saying that, because ultimately for 00:18:51.200 --> 00:18:54.410 most marketers. staying up to date with the best 00:18:54.410 --> 00:18:58.250 models is actually quite hard because every benchmark 00:18:58.250 --> 00:19:01.150 and discussion about model employment improvement 00:19:01.150 --> 00:19:03.990 feels like it's drifted towards coding applications. 00:19:04.329 --> 00:19:07.430 Like, is it better at coding or not? Which for 00:19:07.430 --> 00:19:09.710 your average marketer is like, that's not what 00:19:09.710 --> 00:19:11.670 I spend my day doing. So I don't really care. 00:19:12.349 --> 00:19:14.230 But the reality is, and we'll talk about this 00:19:14.230 --> 00:19:17.650 a bit later, thanks to vibe coding and being 00:19:17.650 --> 00:19:19.670 able to build things, even if you're not a developer, 00:19:19.809 --> 00:19:23.390 actually having a tool, which is good, at coding 00:19:23.390 --> 00:19:26.569 for you can be very beneficial to a marketer 00:19:26.569 --> 00:19:30.430 as we'll speak about later but um yeah it's and 00:19:30.430 --> 00:19:33.730 it's hard to keep up because gpt5 has codex now 00:19:33.730 --> 00:19:37.789 which is basically a you know a specific tool 00:19:37.789 --> 00:19:40.829 for coding obviously claude already had claude 00:19:40.829 --> 00:19:43.829 code as well so yeah it's very hard to keep up 00:19:43.829 --> 00:19:46.289 with but i'm the same as you i mean when gpt5 00:19:46.289 --> 00:19:50.470 came out i was like hey gi a gi and i was like 00:19:50.470 --> 00:19:53.660 no This is not AGI. And it was very interesting 00:19:53.660 --> 00:19:57.220 that my interpretation of the angle that they 00:19:57.220 --> 00:20:01.680 took was powerful AI for all. So rather than 00:20:01.680 --> 00:20:03.799 like power users getting something that's going 00:20:03.799 --> 00:20:07.140 to totally blow them away, they made it so free 00:20:07.140 --> 00:20:10.839 users weren't stuck using crappy models that 00:20:10.839 --> 00:20:13.019 made them feel like AI wasn't very good. I think 00:20:13.019 --> 00:20:15.019 that is extremely important for the adoption 00:20:15.019 --> 00:20:18.880 of AI overall, and I totally get it. I do think 00:20:18.880 --> 00:20:21.220 it has lower hallucination rate and I do think 00:20:21.220 --> 00:20:23.339 it follows instructions more precisely as well. 00:20:23.960 --> 00:20:26.960 So overall, it's pretty good. It completely broke 00:20:26.960 --> 00:20:30.259 a bunch of my custom GPTs because they started 00:20:30.259 --> 00:20:33.059 following the instructions properly. And I'm 00:20:33.059 --> 00:20:34.880 like, oh, crumbs, this is the output I would 00:20:34.880 --> 00:20:39.220 have got if GPT -4 .0 wasn't like basically just 00:20:39.220 --> 00:20:41.440 deciding which instructions it was going to follow 00:20:41.440 --> 00:20:43.680 and which it wasn't. So it's been a hell of a 00:20:43.680 --> 00:20:45.920 lot of fun rewriting the prompts for all of mine. 00:20:47.019 --> 00:20:49.839 custom GPTs, but there you go. So what's your 00:20:49.839 --> 00:20:52.019 preferred model at the moment then? It depends 00:20:52.019 --> 00:20:56.000 entirely on the use case. At the moment, ChatGPT 00:20:56.000 --> 00:21:00.079 is kind of my daily driver for the kind of work 00:21:00.079 --> 00:21:02.880 productivity stuff. So the things that lots of 00:21:02.880 --> 00:21:05.880 people might use it for, summarization of documents, 00:21:06.660 --> 00:21:11.880 redrafting emails, or kind of refining copy editing, 00:21:12.079 --> 00:21:15.000 that kind of thing. ChatGPT tends to be my go 00:21:15.000 --> 00:21:19.490 -to. this past week or two i've really started 00:21:19.490 --> 00:21:22.230 using claude code for a lot of projects just 00:21:22.230 --> 00:21:26.190 building little automations and small scripts 00:21:26.190 --> 00:21:27.789 that are just going to make my life easier day 00:21:27.789 --> 00:21:32.589 to day and helping with configuration of hubspot 00:21:32.589 --> 00:21:37.630 portals and crm config i'll use claude code for 00:21:37.630 --> 00:21:44.309 that but then google gemini i've built some gems 00:21:44.309 --> 00:21:48.710 which are the custom gpt version of google gemini 00:21:48.710 --> 00:21:51.089 where you can build your own kind of flavor as 00:21:51.089 --> 00:21:53.829 it were um are you and i think they're really 00:21:53.829 --> 00:21:59.089 good um i've found a lot of value for these are 00:21:59.089 --> 00:22:01.390 more kind of personal things that i wanted to 00:22:01.390 --> 00:22:03.609 develop and but i've got some great gems that 00:22:03.609 --> 00:22:07.009 i'm using for power lifting for nutrition for 00:22:07.009 --> 00:22:10.660 all sorts of things now and uh yeah i'm i'm I'm 00:22:10.660 --> 00:22:12.759 spread a bit thin across all of them, if I'm 00:22:12.759 --> 00:22:16.259 honest. You always loved a little test and a 00:22:16.259 --> 00:22:19.000 dabble. And I have to admit, the scale of work 00:22:19.000 --> 00:22:21.759 in my new role means that we've got ChachiPT 00:22:21.759 --> 00:22:25.420 Enterprise and we're on Google. So we've got 00:22:25.420 --> 00:22:29.640 Gemini. So I usually run the same prompt or request 00:22:29.640 --> 00:22:32.720 through both and then choose which I like more. 00:22:33.000 --> 00:22:37.299 I'm a bit annoyed. chat gpt5 doesn't really like 00:22:37.299 --> 00:22:40.059 to write sentences it likes to like write three 00:22:40.059 --> 00:22:42.619 words then add a plus symbol then add two words 00:22:42.619 --> 00:22:45.839 and then do an arrow and like so much of that 00:22:45.839 --> 00:22:50.759 stuff is like that um and you know i'm asking 00:22:50.759 --> 00:22:54.500 i'm i'm still mostly honestly overcoming the 00:22:54.500 --> 00:22:57.380 blank page problem but then i'm also very conscious 00:22:57.380 --> 00:23:01.460 that i'm pasting the output into a document it's 00:23:01.460 --> 00:23:03.609 still not what i want it to be it's still not 00:23:03.609 --> 00:23:06.430 quite good enough if i'm honest um plus i feel 00:23:06.430 --> 00:23:08.390 the need to edit it all because otherwise my 00:23:08.390 --> 00:23:10.670 colleagues are going to be like what human uses 00:23:10.670 --> 00:23:13.569 all of these pluses and arrows and like it's 00:23:13.569 --> 00:23:16.230 so obviously not they're the new em dash of the 00:23:16.230 --> 00:23:19.650 world um thanks to gpt5's obsession with them 00:23:19.650 --> 00:23:22.410 so annoying it's almost like it's trying to communicate 00:23:22.410 --> 00:23:25.210 as much valuable information to me and as fewer 00:23:25.210 --> 00:23:29.319 output tokens as possible um yeah But that is 00:23:29.319 --> 00:23:32.579 not great if you're doing blog writing or creating 00:23:32.579 --> 00:23:34.359 e -books or scripts or something. Do you know 00:23:34.359 --> 00:23:37.880 what I mean? So, yeah, GPT -5 got a bit of a 00:23:37.880 --> 00:23:42.059 bash when it came out. Yeah. I think there are 00:23:42.059 --> 00:23:43.940 other problems with the user experience as well. 00:23:43.980 --> 00:23:46.759 So they've made it a routing model. And if you 00:23:46.759 --> 00:23:50.680 go into chat GPT and go into the model selector 00:23:50.680 --> 00:23:54.000 at the top, it will probably be on auto. If you 00:23:54.000 --> 00:23:57.869 run the free account, it's definitely on. I don't 00:23:57.869 --> 00:23:59.710 even know that you've got the option, to be honest, 00:23:59.829 --> 00:24:02.630 on the free account. But if you're in the paid 00:24:02.630 --> 00:24:04.430 for account, you'll have the dropdown and it 00:24:04.430 --> 00:24:07.049 will typically be on auto. And what it's doing 00:24:07.049 --> 00:24:09.970 there is it's running it through its own classification 00:24:09.970 --> 00:24:12.289 engine first to look at your query and then say, 00:24:12.490 --> 00:24:15.130 hmm, do I need to think about this one or can 00:24:15.130 --> 00:24:19.329 I just give an answer? And I'm increasingly finding 00:24:19.329 --> 00:24:22.569 people say, this is people with paid for accounts 00:24:22.569 --> 00:24:25.319 that are... quite adept at using it but they're 00:24:25.319 --> 00:24:29.720 saying GPT -5 is worse and I'm asking them about 00:24:29.720 --> 00:24:31.759 kind of what's going on and what their use cases 00:24:31.759 --> 00:24:34.220 are and basically they've not turned on thinking 00:24:34.220 --> 00:24:36.880 mode so it's in instant mode and I think that 00:24:36.880 --> 00:24:40.579 there's a in the auto model I wouldn't be surprised 00:24:40.579 --> 00:24:42.960 if what they're doing is basically running it 00:24:42.960 --> 00:24:49.000 through a GPT -5 mini or GPT -5 nano for cost 00:24:49.000 --> 00:24:51.140 reduction because most of the time it's kind 00:24:51.140 --> 00:24:53.519 of good enough you know if it's someone writes 00:24:53.519 --> 00:24:55.940 this document nano or mini can do that great 00:24:55.940 --> 00:24:59.799 if it's redrafted email it can do that at a level 00:24:59.799 --> 00:25:02.059 that's kind of fine for most people but i think 00:25:02.059 --> 00:25:07.339 it it doesn't apply its intelligence often enough 00:25:07.339 --> 00:25:09.900 or frequently enough so lots of people are going 00:25:09.900 --> 00:25:11.720 well it seems like a bit of a step backwards 00:25:11.720 --> 00:25:13.539 i asked it for something the quality output is 00:25:13.539 --> 00:25:16.819 not quite there i asked people to then turn on 00:25:16.819 --> 00:25:19.660 thinking mode and say give it a go run the same 00:25:19.660 --> 00:25:22.059 prompt And they're like, oh, wow, that's much 00:25:22.059 --> 00:25:25.079 better. There's a cost to that in terms of time. 00:25:25.160 --> 00:25:27.619 It has to think for a couple of minutes. But 00:25:27.619 --> 00:25:30.259 the quality of the output is so greatly improved. 00:25:30.640 --> 00:25:33.019 And I think unless people know that and it's 00:25:33.019 --> 00:25:37.220 well communicated to them, they feel somewhat 00:25:37.220 --> 00:25:40.799 frustrated with the experience. That said, there's 00:25:40.799 --> 00:25:43.900 great benefits to open AI in serving lower cost 00:25:43.900 --> 00:25:47.279 models that don't use so much compute. Yeah, 00:25:47.319 --> 00:25:50.559 the drive for cost reduction is real. mostly 00:25:50.559 --> 00:25:52.839 i think because of the lack of compute in the 00:25:52.839 --> 00:25:56.259 world like we can't produce chips data centers 00:25:56.259 --> 00:26:00.480 energy energy production facilities fast enough 00:26:00.480 --> 00:26:02.700 and i think that serving all these people these 00:26:02.700 --> 00:26:05.079 models i think that is genuinely a fairly big 00:26:05.079 --> 00:26:09.319 driver of it but um i don't find thinking makes 00:26:09.319 --> 00:26:13.359 my written outputs any better if i'm honest i 00:26:13.359 --> 00:26:15.180 think it makes its analysis like if i'm like 00:26:15.180 --> 00:26:17.660 here's a load of context let's create a content 00:26:17.660 --> 00:26:23.099 strategy together or So let's brainstorm out 00:26:23.099 --> 00:26:27.539 future scenario planning. If this happens, what 00:26:27.539 --> 00:26:28.960 should we do? If this happens, what should we 00:26:28.960 --> 00:26:30.740 do? I think his thinking is awesome for that 00:26:30.740 --> 00:26:34.039 type of stuff. But yeah, I need to play with 00:26:34.039 --> 00:26:37.819 Claude this week, really, the new Sonnet 4 .5, 00:26:38.059 --> 00:26:41.759 10 .5, whatever we're up to. Because in general, 00:26:41.799 --> 00:26:44.299 I always found Claude was better at writing, 00:26:44.339 --> 00:26:48.000 but not... so good as in as instruction following 00:26:48.000 --> 00:26:51.119 so 4 .0 was my go -to model for writing not claude 00:26:51.119 --> 00:26:53.819 because claude thought it knew better like it 00:26:53.819 --> 00:26:56.319 would definitely defer to its own internal style 00:26:56.319 --> 00:26:59.140 guide no matter how hard i tried to push it whereas 00:26:59.140 --> 00:27:03.799 4 .0 was more malleable gpt5 follows the instructions 00:27:03.799 --> 00:27:06.779 better but just doesn't seem to be as good at 00:27:06.779 --> 00:27:09.339 writing like we'll talk about it a little bit 00:27:09.339 --> 00:27:12.960 later around the concept of of work slop and 00:27:12.960 --> 00:27:15.970 where we are with technologies for marketing 00:27:15.970 --> 00:27:17.849 applications and where we're at right now but 00:27:17.849 --> 00:27:20.589 i haven't got a good writing model if i'm honest 00:27:20.589 --> 00:27:23.089 at the moment um and i'm hoping that claude's 00:27:23.089 --> 00:27:25.910 sonnet the new model is going to be it for a 00:27:25.910 --> 00:27:30.750 while on the the model development generally 00:27:30.750 --> 00:27:33.289 just to kind of conclude on that i think we have 00:27:33.289 --> 00:27:35.569 definitely reached a point that we talked about 00:27:35.569 --> 00:27:40.289 over a year ago which is we've kind of saturated 00:27:40.289 --> 00:27:43.450 the intelligence requirements of the average 00:27:43.450 --> 00:27:48.980 user They are plenty smart enough for the vast 00:27:48.980 --> 00:27:53.099 majority of people. And we are now in a world 00:27:53.099 --> 00:27:56.619 where we need to extend their capabilities beyond 00:27:56.619 --> 00:28:00.059 just being smart and being able to think things 00:28:00.059 --> 00:28:04.900 through into a new paradigm. I completely agree. 00:28:05.019 --> 00:28:08.220 And if you look at the chat that's coming out 00:28:08.220 --> 00:28:13.019 of all the major AI labs and companies, they're 00:28:13.019 --> 00:28:16.630 looking at. solving impossible math problems 00:28:16.630 --> 00:28:21.569 and contributing to scientific research. And 00:28:21.569 --> 00:28:24.349 that pushing the intelligence frontier is not 00:28:24.349 --> 00:28:27.150 to help marketers improve their click -through 00:28:27.150 --> 00:28:30.990 rate, right? It's to cure cancer and what have 00:28:30.990 --> 00:28:33.730 you. And what that does tell me is I think it's 00:28:33.730 --> 00:28:35.430 another good example of what you just said, which 00:28:35.430 --> 00:28:39.089 is the frontier is intelligence levels that your 00:28:39.089 --> 00:28:43.930 average user doesn't need. Which is why I guess 00:28:43.930 --> 00:28:46.190 it's still somewhat frustrating that I have a 00:28:46.190 --> 00:28:48.630 bunch of marketing use cases that just don't 00:28:48.630 --> 00:28:52.589 work very well. This is another tangent, but, 00:28:52.690 --> 00:28:55.630 and I'm going to probably not do it justice. 00:28:55.730 --> 00:28:57.289 So if you want the details, you should Google 00:28:57.289 --> 00:29:01.569 this. Claude is basically building, Claude Anthropic 00:29:01.569 --> 00:29:05.910 are building a team that hires human experts 00:29:05.910 --> 00:29:09.410 to train and provide reinforcement learning to. 00:29:10.349 --> 00:29:16.009 ai models for specific work tasks so they want 00:29:16.009 --> 00:29:19.430 computer use and they want to record what people 00:29:19.430 --> 00:29:22.009 do when they're using the crm when they're sending 00:29:22.009 --> 00:29:26.630 an email when they're analyzing um a look a studio 00:29:26.630 --> 00:29:29.529 data report to figure out you know which channel 00:29:29.529 --> 00:29:32.450 which marketing channels working the best the 00:29:32.450 --> 00:29:36.700 so that they can actually take humans that are 00:29:36.700 --> 00:29:39.259 really good at this and try and apply some of 00:29:39.259 --> 00:29:41.420 that knowledge and train their models that way 00:29:41.420 --> 00:29:44.200 i think there's two parts to that the first is 00:29:44.200 --> 00:29:47.680 having the computer use agent models which we'll 00:29:47.680 --> 00:29:51.500 talk about a bit later get better at using your 00:29:51.500 --> 00:29:54.900 standard crm software and what have you effectively 00:29:54.900 --> 00:29:57.160 without clicking on the wrong things and basically 00:29:57.160 --> 00:29:59.079 just breaking and not being able to do anything 00:30:00.670 --> 00:30:03.210 but also just trying to get them to better understand 00:30:03.210 --> 00:30:07.049 what the very best humans do to solve real world 00:30:07.049 --> 00:30:12.609 business problems and work problems, which is 00:30:12.609 --> 00:30:15.809 very different from trying to like meta -analyse 00:30:15.809 --> 00:30:18.930 a hundred papers about this receptor related 00:30:18.930 --> 00:30:21.569 to cancer to try and find some unexpected connection 00:30:21.569 --> 00:30:24.650 and a new pathway that no one had imagined before. 00:30:24.710 --> 00:30:26.029 That's a completely different problem, right? 00:30:26.150 --> 00:30:28.920 But in business... I'm sure that could be applied, 00:30:28.980 --> 00:30:30.579 but we just don't care as much about that. So 00:30:30.579 --> 00:30:33.700 I am quite excited for the next year from that 00:30:33.700 --> 00:30:37.000 perspective. There were a bunch of anthropic 00:30:37.000 --> 00:30:38.960 people on another podcast I was listening to, 00:30:39.019 --> 00:30:41.460 in essence, where they were saying, I think it 00:30:41.460 --> 00:30:44.059 might have been the Dwarkesh podcast, that computers 00:30:44.059 --> 00:30:47.319 can reinforcement learn off of humans doing knowledge 00:30:47.319 --> 00:30:50.440 work because it's all digital and they will get 00:30:50.440 --> 00:30:53.750 better at it than humans. Right. So it's going 00:30:53.750 --> 00:30:55.789 to be interesting to see how that how that plays 00:30:55.789 --> 00:30:58.829 out over time as well. But should we talk a little 00:30:58.829 --> 00:31:00.869 bit about open versus closed models? We used 00:31:00.869 --> 00:31:03.369 to talk about it quite a lot. I think it's less 00:31:03.369 --> 00:31:06.349 of a thing now. What are your thoughts on where 00:31:06.349 --> 00:31:09.150 we are in the open versus closed debate, especially 00:31:09.150 --> 00:31:11.690 over while we've been off air? Yeah. So I had 00:31:11.690 --> 00:31:15.849 a bit of a moment at the start of the year around 00:31:15.849 --> 00:31:18.029 Christmas, in fact, wasn't it? It was Christmas 00:31:18.029 --> 00:31:22.009 and New Year. deep seek from china came out as 00:31:22.009 --> 00:31:25.210 a reasoning model that was cheap to run it was 00:31:25.210 --> 00:31:30.490 cheap to train and this announcement and model 00:31:30.490 --> 00:31:35.450 debut effectively wiped off billions from the 00:31:35.450 --> 00:31:37.549 us stock market but it all recovered they're 00:31:37.549 --> 00:31:40.349 doing okay no no need to panic nvidia will be 00:31:40.349 --> 00:31:45.589 fine um but the the model itself was interesting 00:31:45.589 --> 00:31:48.750 and people started to learn a lot from it because 00:31:48.750 --> 00:31:52.390 the paper that was written alongside it introduced 00:31:52.390 --> 00:31:55.369 some new techniques for model optimization and 00:31:55.369 --> 00:31:58.670 hence why they were able to to do it so so cheaply 00:31:58.670 --> 00:32:02.430 but then things kind of stalled and then the 00:32:02.430 --> 00:32:05.130 world as far as open source models turned towards 00:32:05.130 --> 00:32:09.509 meta who were the the kind of shining light for 00:32:09.509 --> 00:32:12.509 research and development of open source models 00:32:12.509 --> 00:32:21.430 and the llama 4 model was what released and the 00:32:21.430 --> 00:32:26.009 reception was oh is that it was something of 00:32:26.009 --> 00:32:30.829 a damp squib and subsequent to that meta have 00:32:30.829 --> 00:32:34.750 kind of shifted their approach and they're talking 00:32:34.750 --> 00:32:39.089 more about closed proprietary models and going 00:32:39.089 --> 00:32:43.289 after the agi model development that they're 00:32:43.289 --> 00:32:46.809 hiring at a very aggressive rate taking um that 00:32:46.809 --> 00:32:49.660 some of the best talent from across all the ai 00:32:49.660 --> 00:32:57.359 labs but it signaled to me at least that open 00:32:57.359 --> 00:33:04.559 source at the the very frontier is not where 00:33:04.559 --> 00:33:06.640 things are going to happen right that's not where 00:33:06.640 --> 00:33:08.240 we should be looking there's going to be lots 00:33:08.240 --> 00:33:10.680 of interesting development in the open source 00:33:10.680 --> 00:33:14.859 world i'm sure but if you're looking for the 00:33:14.859 --> 00:33:18.619 biggest baddest most exciting intelligent models 00:33:18.619 --> 00:33:23.019 that's not where the action's going to be yeah 00:33:23.019 --> 00:33:25.420 it's going to be quite interesting quite a lot 00:33:25.420 --> 00:33:27.819 of people who speak about this a lot in our space 00:33:27.819 --> 00:33:33.319 are really of the belief that models are going 00:33:33.319 --> 00:33:36.160 to be a commodity like the amount of money that 00:33:36.160 --> 00:33:39.119 you'll make with any particular model they'll 00:33:39.119 --> 00:33:41.039 all like there'll be there might be one model 00:33:41.039 --> 00:33:42.660 to rule them all there might be like a handful 00:33:42.660 --> 00:33:45.259 of models there might be Loads of model niche 00:33:45.259 --> 00:33:48.480 specialists, like there's a GPT Finance for all 00:33:48.480 --> 00:33:50.279 the finance people to use. Like all of these 00:33:50.279 --> 00:33:54.019 things are still possible. But as the amount 00:33:54.019 --> 00:33:56.480 of compute build out goes and the amount of need 00:33:56.480 --> 00:33:59.819 for AI intelligence on demand, it's just this 00:33:59.819 --> 00:34:02.400 idea that the underlying model won't matter as 00:34:02.400 --> 00:34:04.839 much as whether or not you have brand affinity. 00:34:04.940 --> 00:34:07.539 You're a ChatGPT user. You're a Gemini user. 00:34:07.859 --> 00:34:09.460 You know, this is why I still think there's a 00:34:09.460 --> 00:34:12.139 massive chance that Google wins. They have great 00:34:12.139 --> 00:34:14.679 models. And they have distribution and brand 00:34:14.679 --> 00:34:19.000 equity. I don't think Meta is going to win in 00:34:19.000 --> 00:34:21.300 anything because no one trusts them. Like no 00:34:21.300 --> 00:34:23.619 one wants to give them data. We were talking, 00:34:23.639 --> 00:34:26.599 not on a podcast recording, about the new Meta 00:34:26.599 --> 00:34:30.119 glasses, which are really cool and seem really 00:34:30.119 --> 00:34:33.440 interesting. But do I want to go around recording 00:34:33.440 --> 00:34:36.420 my life? No, I know that Meta is, there must 00:34:36.420 --> 00:34:38.039 be somewhere in the T's and C's that they're 00:34:38.039 --> 00:34:40.760 able to use the real world data from my glasses 00:34:40.760 --> 00:34:43.840 to figure out what. human paul does and where 00:34:43.840 --> 00:34:46.980 he goes and what he looks like and how he opens 00:34:46.980 --> 00:34:48.940 doors and all these other things that could probably 00:34:48.940 --> 00:34:51.119 be used to i don't know train robots to do stuff 00:34:51.119 --> 00:34:55.500 i just don't trust a mine yeah that is a real 00:34:55.500 --> 00:34:58.900 reputational risk for them and if you follow 00:34:58.900 --> 00:35:02.760 any of the development team on threads it seems 00:35:02.760 --> 00:35:05.380 like they're aware of that well they would be 00:35:05.380 --> 00:35:07.420 aware of that because lots of people tell them 00:35:07.420 --> 00:35:09.900 directly in the comments all the time that we 00:35:09.900 --> 00:35:13.340 don't trust you So it is something of an uphill 00:35:13.340 --> 00:35:16.619 battle for them. I have a pair of the Meta Ray 00:35:16.619 --> 00:35:19.800 -Bans and they're great. I really like them. 00:35:19.840 --> 00:35:23.340 They can do some cool things. But interestingly, 00:35:23.539 --> 00:35:27.179 the reception, even if you as a wearer do trust 00:35:27.179 --> 00:35:31.340 them, there is definitely mistrust amongst people 00:35:31.340 --> 00:35:33.719 that you're with when you're wearing them. Not 00:35:33.719 --> 00:35:36.300 everyone, but some people will be like, is that 00:35:36.300 --> 00:35:40.460 a camera on your face? Why, yes, yes, it is. 00:35:41.469 --> 00:35:43.750 I don't like that. Do you know, I'm going to 00:35:43.750 --> 00:35:47.170 make a dystopian prediction here. I don't think 00:35:47.170 --> 00:35:49.690 any of us like that. I don't think we want the 00:35:49.690 --> 00:35:51.369 fact that someone else is walking around with 00:35:51.369 --> 00:35:54.269 a camera recording what we say and do. Like when 00:35:54.269 --> 00:35:55.909 you have an argument with someone, you're like, 00:35:55.969 --> 00:35:58.010 I never said that. And they're like, well, I'm 00:35:58.010 --> 00:36:02.630 just going to refer you to Exhibit A here. This 00:36:02.630 --> 00:36:05.090 recording of you that you may or may not have 00:36:05.090 --> 00:36:08.809 known that I was taking. But I think, unfortunately... 00:36:09.480 --> 00:36:11.340 we're going to allow it and the reason we're 00:36:11.340 --> 00:36:14.519 going to allow it is because the new version 00:36:14.519 --> 00:36:19.199 of the meta ray bands with the wrist um controllers 00:36:19.199 --> 00:36:21.119 where basically you can just move your hands 00:36:21.119 --> 00:36:23.480 and some of those videos of people being able 00:36:23.480 --> 00:36:26.699 to like scroll their instagram feed in the little 00:36:26.699 --> 00:36:30.360 um screen that's in the glasses that no one else 00:36:30.360 --> 00:36:32.260 can see and these are just normal looking glasses 00:36:32.260 --> 00:36:35.219 with your hands in your pocket scrolling through 00:36:35.219 --> 00:36:38.329 the feed so no one can even necessarily see I 00:36:38.329 --> 00:36:41.210 think we're so addicted to consuming short form 00:36:41.210 --> 00:36:45.389 dopamine charged crap on our phones that the 00:36:45.389 --> 00:36:48.670 fact that we could do that on a train or even 00:36:48.670 --> 00:36:51.250 at dinner with a bunch of other people and no 00:36:51.250 --> 00:36:53.690 one know that we're actually watching a TikTok 00:36:53.690 --> 00:36:55.769 video and we're not really paying attention to 00:36:55.769 --> 00:36:58.429 them. I think that the allure of that is going 00:36:58.429 --> 00:37:01.070 to be too great and that we're all going to accept. 00:37:01.920 --> 00:37:03.780 The fact that all these glasses are in the world 00:37:03.780 --> 00:37:05.920 because of the benefits that we personally get, 00:37:06.039 --> 00:37:08.179 even though it probably creates a less private, 00:37:08.380 --> 00:37:12.579 always on recorded state for everybody. Yeah, 00:37:12.599 --> 00:37:14.920 I can't disagree. I think people will lean into 00:37:14.920 --> 00:37:20.619 it. And it's that world view of it is somewhat 00:37:20.619 --> 00:37:24.260 tragic, but probably where we're heading. Yeah, 00:37:24.320 --> 00:37:26.300 as we talked about, I have a use case. I do want 00:37:26.300 --> 00:37:29.980 to be able to more easily respond to emails and 00:37:29.980 --> 00:37:33.199 Slack messages. walking the dog. And if I don't 00:37:33.199 --> 00:37:35.199 have to have my phone in one hand and the dog 00:37:35.199 --> 00:37:37.860 in the other, it is going to increase my life 00:37:37.860 --> 00:37:41.079 expectancy. So I'm all for that use case. But 00:37:41.079 --> 00:37:43.639 yeah, it's the fact they'll be on our faces all 00:37:43.639 --> 00:37:47.300 day. Meta's entire industry is the attention 00:37:47.300 --> 00:37:51.000 economy. So what's their incentive? It isn't 00:37:51.000 --> 00:37:53.820 to make you more productive. It's to show you 00:37:53.820 --> 00:37:59.079 more content and monetize that as aggressively 00:37:59.079 --> 00:38:03.139 as possible. Yeah, like, you you need you you 00:38:03.139 --> 00:38:05.219 put your phone down during the day for various 00:38:05.219 --> 00:38:08.079 reasons you never need to put your phone down 00:38:08.079 --> 00:38:12.320 again it's on your face just imagine constant 00:38:12.320 --> 00:38:17.019 notifications in your eyeline oh my god it's 00:38:17.019 --> 00:38:19.000 going to be super interesting to see if it's 00:38:19.000 --> 00:38:22.739 really annoying or if it actually is good but 00:38:22.739 --> 00:38:24.780 it's here and it's coming and i don't think it's 00:38:24.780 --> 00:38:28.059 going to stop for the reason you just said um 00:38:28.059 --> 00:38:30.239 right let's go on to the next section i think 00:38:30.239 --> 00:38:32.320 this is my favorite part of the podcast today 00:38:32.320 --> 00:38:35.639 let's talk about agentic progress because i know 00:38:35.639 --> 00:38:38.340 this is something that you play with a lot perhaps 00:38:38.340 --> 00:38:40.699 just there surely can't be a listener in the 00:38:40.699 --> 00:38:43.659 world who hasn't heard agentic and the term agent 00:38:43.659 --> 00:38:47.719 now but why don't you just outline what the emergence 00:38:47.719 --> 00:38:50.639 of agents is and then start to talk us through 00:38:50.639 --> 00:38:54.280 what that means for people at work people who 00:38:54.280 --> 00:38:56.619 work in marketing etc and maybe even talk about 00:38:56.619 --> 00:38:58.360 some of the things you've been playing with so 00:38:58.360 --> 00:39:03.340 agentic ai this is ai models that have some form 00:39:03.340 --> 00:39:07.400 of agency the ability to take actions or make 00:39:07.400 --> 00:39:10.940 decisions plan and then interact with the world 00:39:10.940 --> 00:39:13.519 now in some instances this might just be a very 00:39:13.519 --> 00:39:17.920 simple customized version of chat gpt where you've 00:39:17.920 --> 00:39:21.860 given it some access to documents and you ask 00:39:21.860 --> 00:39:24.179 it a question and it can go away research those 00:39:24.179 --> 00:39:26.139 documents and come back and present you the answer 00:39:26.139 --> 00:39:31.159 that's a very simple ai agentic assistant you 00:39:31.159 --> 00:39:34.400 might connect it to some tools so it can go and 00:39:34.400 --> 00:39:37.360 plug into your crm and it can crawl your crm 00:39:37.360 --> 00:39:39.639 and bring back some data from there it's a kind 00:39:39.639 --> 00:39:44.219 of extended ai assistant agent then we've got 00:39:44.219 --> 00:39:46.940 kind of workflow agents and workflow agents are 00:39:46.940 --> 00:39:50.760 the more trigger based interactions these are 00:39:50.760 --> 00:39:53.059 very similar to the sorts of workflows that we've 00:39:53.059 --> 00:39:54.929 talked about for a long time in this podcast 00:39:54.929 --> 00:39:57.869 things that you might build with make .com or 00:39:57.869 --> 00:40:01.369 zapier or na10 which has become a big workflow 00:40:01.369 --> 00:40:05.050 tool of choice now these systems that they're 00:40:05.050 --> 00:40:08.869 using triggers so somebody fills out a form on 00:40:08.869 --> 00:40:11.869 your website and then you build out a sequence 00:40:11.869 --> 00:40:18.010 of steps at which point you might use ai to do 00:40:18.010 --> 00:40:20.630 one of those steps so it might be a prompt that 00:40:20.630 --> 00:40:25.500 formats some data or It might be AI -generated 00:40:25.500 --> 00:40:28.400 video gets created to respond to somebody who's 00:40:28.400 --> 00:40:30.019 come on the website and filled out a form and 00:40:30.019 --> 00:40:33.619 they get this HeyGen video. And they are agentic 00:40:33.619 --> 00:40:39.179 workflows. They are created by humans and built 00:40:39.179 --> 00:40:40.639 by humans. And what we're doing is effectively 00:40:40.639 --> 00:40:43.719 telling the AI how to respond within certain 00:40:43.719 --> 00:40:48.340 steps. And these are being posited by or positioned, 00:40:48.619 --> 00:40:53.780 should I say, as agents. quite a lot and and 00:40:53.780 --> 00:40:56.739 i can see the argument for that and i just think 00:40:56.739 --> 00:41:00.760 they're workflows with an ai step but we don't 00:41:00.760 --> 00:41:04.019 need to get into the semantics of that um but 00:41:04.019 --> 00:41:06.760 then we've got the kind of uh i don't really 00:41:06.760 --> 00:41:09.699 know the word for it but the kind of true agents 00:41:09.699 --> 00:41:14.619 which are goal oriented and you give them a task 00:41:14.619 --> 00:41:18.559 and they go off and execute it they do the planning 00:41:18.559 --> 00:41:23.219 they can overcome obstacles so if they hit a 00:41:23.219 --> 00:41:25.719 problem they can see what the problem is and 00:41:25.719 --> 00:41:28.579 then try to find a way around that problem and 00:41:28.579 --> 00:41:31.239 they can just run in the background come back 00:41:31.239 --> 00:41:35.699 and deliver you the final output clawed code 00:41:35.699 --> 00:41:39.219 is a fantastic example of this from anthropic 00:41:39.219 --> 00:41:42.179 this is primarily for developers but it runs 00:41:42.179 --> 00:41:46.639 in terminal or in your ide of choice it can work 00:41:46.639 --> 00:41:49.960 in vs code and others and you can just ask it 00:41:49.960 --> 00:41:52.119 to go and create something and there's a plan 00:41:52.119 --> 00:41:55.659 mode where it writes out its plan for your brief 00:41:55.659 --> 00:41:59.840 and then you tell it to go and do it and watching 00:41:59.840 --> 00:42:01.659 it work is fascinating because you can see it 00:42:01.659 --> 00:42:04.659 type out the the things that it's doing and then 00:42:04.659 --> 00:42:08.099 when it hits a an error code it looks at the 00:42:08.099 --> 00:42:10.579 error code and it says hmm it looks like the 00:42:10.579 --> 00:42:13.079 server isn't responding i'm now going to try 00:42:13.079 --> 00:42:16.559 and update this and running through all of the 00:42:16.559 --> 00:42:19.769 steps in the background and You know, at the 00:42:19.769 --> 00:42:22.949 moment, these kind of models can run for limited 00:42:22.949 --> 00:42:27.670 periods of time. But I think the exciting thing 00:42:27.670 --> 00:42:31.269 is when we start getting into a world where he's 00:42:31.269 --> 00:42:35.369 on 24 hours a day, seven days a week, you're 00:42:35.369 --> 00:42:42.090 always on agentic co -workers. Yeah, I completely 00:42:42.090 --> 00:42:47.079 agree. The sort of Zapier make style. agents 00:42:47.079 --> 00:42:49.840 are not really agents the humans planning all 00:42:49.840 --> 00:42:54.800 of the logic there and leveraging ai to for its 00:42:54.800 --> 00:42:57.000 language abilities really to write something 00:42:57.000 --> 00:43:01.480 or to analyze some text um i still think there 00:43:01.480 --> 00:43:05.500 is the opportunity to build in routing options 00:43:05.500 --> 00:43:09.780 where if the llm says so let's say you get customer 00:43:09.780 --> 00:43:12.860 reviews and it processes customer reviews and 00:43:12.860 --> 00:43:16.050 analyzes the customer reviews for sentiment and 00:43:16.050 --> 00:43:18.150 then gives a report back on whether it's positive 00:43:18.150 --> 00:43:21.429 or negative. And then there's separate routing 00:43:21.429 --> 00:43:24.530 in the agent that then will, the email back to 00:43:24.530 --> 00:43:26.889 that person will have certain text in it based 00:43:26.889 --> 00:43:29.130 on that classification. But the human's coming 00:43:29.130 --> 00:43:31.190 up with the plan. And these things look like 00:43:31.190 --> 00:43:35.590 crazy spider webs, right, of like nodes, like 00:43:35.590 --> 00:43:38.289 line, decision point, action, line, decision 00:43:38.289 --> 00:43:42.059 point, action. etc etc um and the impressive 00:43:42.059 --> 00:43:43.940 ones are incredible we build a lot of them at 00:43:43.940 --> 00:43:46.480 supreme group and some of ours look like the 00:43:46.480 --> 00:43:51.679 zoomed out view is like um chaos theory in front 00:43:51.679 --> 00:43:53.599 like it's like oh if i zoom out far enough it 00:43:53.599 --> 00:43:57.320 starts to turn into this beautiful swirl um and 00:43:57.320 --> 00:44:00.579 it's really cool and i think it's the closest 00:44:00.579 --> 00:44:04.139 that we've got to building in some deterministic 00:44:04.139 --> 00:44:07.139 nature into these agents right where you like 00:44:07.139 --> 00:44:10.969 i don't want to trust the ai here to not hallucinate 00:44:10.969 --> 00:44:13.389 i'm going to make it make a decision based on 00:44:13.389 --> 00:44:18.110 the framework that i've built but um being able 00:44:18.110 --> 00:44:20.469 to use agents that browse the web click on websites 00:44:20.469 --> 00:44:23.130 fill in forms gather information put them in 00:44:23.130 --> 00:44:26.130 documents i mean you use manus a lot which is 00:44:26.130 --> 00:44:29.369 a like a computer using agent what sort of stuff 00:44:29.369 --> 00:44:31.190 have you been doing that marketers might find 00:44:31.190 --> 00:44:35.050 interesting in manus so minus is really interesting 00:44:35.050 --> 00:44:38.860 it started out as it was a like you say computer 00:44:38.860 --> 00:44:41.360 use it could browse the web but they keep adding 00:44:41.360 --> 00:44:43.420 new features so if you're a marketer and you're 00:44:43.420 --> 00:44:45.099 thinking well i've got to put that slide deck 00:44:45.099 --> 00:44:48.039 together this will go away and research it will 00:44:48.039 --> 00:44:51.079 do external research on the web or you can upload 00:44:51.079 --> 00:44:53.699 a transcript you can upload your own files whatever 00:44:53.699 --> 00:44:56.380 it may be and ask it to put a slide deck together 00:44:56.380 --> 00:44:59.619 it will just do that on its own or you can tell 00:44:59.619 --> 00:45:01.639 it quite specifically these are the slides that 00:45:01.639 --> 00:45:04.099 i want and outline that and it will build it 00:45:04.099 --> 00:45:07.019 for you and make that fully downloadable and 00:45:07.019 --> 00:45:10.019 you can view it on the web it does videos it 00:45:10.019 --> 00:45:12.519 does image generation it creates little micro 00:45:12.519 --> 00:45:17.119 apps and micro sites and will deploy them on 00:45:17.119 --> 00:45:21.340 the live web so i've started using this for proposal 00:45:21.340 --> 00:45:24.260 development when i've drafted the the proposal 00:45:24.260 --> 00:45:28.480 and i've got the document together i'm not a 00:45:28.480 --> 00:45:32.019 great designer and everything i create looks 00:45:32.019 --> 00:45:36.730 a bit boring in google docs So I've just started 00:45:36.730 --> 00:45:40.550 taking that Google Doc and that's the main document. 00:45:40.670 --> 00:45:43.590 That's the thing where all the detail is. But 00:45:43.590 --> 00:45:45.769 the first impression that I want the client to 00:45:45.769 --> 00:45:48.489 have is something that's a bit more user friendly. 00:45:48.650 --> 00:45:51.269 It looks a bit smart, a bit polished, maybe some 00:45:51.269 --> 00:45:54.050 interactive elements. So I give the proposal 00:45:54.050 --> 00:45:56.190 to Manus and I say, turn this into an interactive 00:45:56.190 --> 00:45:59.690 web proposal. It's got all of my brand guidelines. 00:45:59.889 --> 00:46:02.739 It's got. links to where it can source relevant 00:46:02.739 --> 00:46:05.719 images that I've told it about. And it builds 00:46:05.719 --> 00:46:08.900 me a little microsite that I then send as the 00:46:08.900 --> 00:46:12.019 proposal. So when I email the client, I'll give 00:46:12.019 --> 00:46:14.340 them a link to that. And then I attach the PDF, 00:46:14.460 --> 00:46:18.619 which has all of the proposal details outlined 00:46:18.619 --> 00:46:22.260 in full. And it just does it. It just works. 00:46:22.320 --> 00:46:25.039 It's like using ChatGPT. It's a chat based interface 00:46:25.039 --> 00:46:29.829 and it has lots of connectors. so you can connect 00:46:29.829 --> 00:46:32.849 it with your you can connect it with hubspot 00:46:32.849 --> 00:46:34.630 you can connect it with zapier you can connect 00:46:34.630 --> 00:46:38.309 it with lots of different tools right these sas 00:46:38.309 --> 00:46:40.769 products that you're using every day you can 00:46:40.769 --> 00:46:43.650 integrate it with those and interestingly you 00:46:43.650 --> 00:46:46.710 can do scheduled tasks this is a big one every 00:46:46.710 --> 00:46:51.409 morning at 9 00 am i want you to go online and 00:46:51.409 --> 00:46:54.170 find me the top news stories in my industry and 00:46:54.170 --> 00:46:58.449 then deliver that to me in a google doc it will 00:46:58.449 --> 00:47:03.889 do that quite happily um they're incredibly useful 00:47:03.889 --> 00:47:06.869 and i think when people start to think of them 00:47:06.869 --> 00:47:09.690 as having like an office junior that you can 00:47:09.690 --> 00:47:12.389 just go away and i need these things doing regularly 00:47:12.389 --> 00:47:15.409 you go away and do it but then adding in all 00:47:15.409 --> 00:47:19.349 of those connectors and tool use um gives them 00:47:19.349 --> 00:47:24.050 so much more value yeah i i haven't played with 00:47:24.050 --> 00:47:26.230 manas as much as i really want to i mean chachi 00:47:26.230 --> 00:47:30.159 bt has its own agent but and in fact they released 00:47:30.159 --> 00:47:33.760 an updated version recently um so i do need to 00:47:33.760 --> 00:47:35.480 play with that because the old agent mode was 00:47:35.480 --> 00:47:37.599 basically only on pro so you had to be spending 00:47:37.599 --> 00:47:39.579 200 odd dollars a month to be able to access 00:47:39.579 --> 00:47:44.159 it but um it's quite interesting there is a is 00:47:44.159 --> 00:47:47.440 it the is it the metr benchmark for agents like 00:47:47.440 --> 00:47:50.639 how long they can work autonomously and still 00:47:50.639 --> 00:47:53.780 get a high level of success and it's on this 00:47:53.780 --> 00:47:58.280 really interesting exponential curve where it's 00:47:58.280 --> 00:48:00.460 doubling something like every seven months so 00:48:00.460 --> 00:48:02.619 it's like they could operate for 15 minutes autonomously 00:48:02.619 --> 00:48:05.300 now that's 30 minutes now it's an hour now it's 00:48:05.300 --> 00:48:07.960 two hours and currently i think we're sort of 00:48:07.960 --> 00:48:11.900 around is it eight hours that we're predicted 00:48:11.900 --> 00:48:13.980 to be at i think at about this time and i think 00:48:13.980 --> 00:48:16.579 if you follow the trend out the ai agents should 00:48:16.579 --> 00:48:20.599 be able to work for a week by mid to end of 2026 00:48:20.599 --> 00:48:25.360 so you know there's a huge amount of improvement 00:48:25.360 --> 00:48:30.150 here in a ai agents being able to conduct multi 00:48:30.150 --> 00:48:35.349 -step processes and produce a high quality outcome 00:48:35.349 --> 00:48:37.909 at the end so i definitely want to spend a bit 00:48:37.909 --> 00:48:40.409 more time playing with manas because i think 00:48:40.409 --> 00:48:44.420 this is like the one of the first forays that 00:48:44.420 --> 00:48:46.139 we'll get into this and it'll probably be baked 00:48:46.139 --> 00:48:49.880 in to Gemini and ChatGPT and as these things 00:48:49.880 --> 00:48:52.340 mature you know in two years ago we'll be like 00:48:52.340 --> 00:48:53.980 would you remember that tool Manus unfortunately 00:48:53.980 --> 00:48:56.679 for the folks at Manus because of course if you've 00:48:56.679 --> 00:48:58.539 got a Google account you're just going to want 00:48:58.539 --> 00:49:01.519 to rely on the one tool you've got yeah yeah 00:49:01.519 --> 00:49:04.059 the interesting thing on the subscriptions I 00:49:04.059 --> 00:49:07.760 remember when ChatGPT came out with you know 00:49:07.760 --> 00:49:11.059 they've got Plus and Teams and then Pro at $200 00:49:11.059 --> 00:49:13.849 a month And there was talk, do you remember that 00:49:13.849 --> 00:49:17.469 report came out and it said they're internally 00:49:17.469 --> 00:49:21.329 talking about subscriptions that cost $2 ,000 00:49:21.329 --> 00:49:24.110 a month or even up to $20 ,000 a month. And, 00:49:24.110 --> 00:49:27.230 you know, a year ago when you'd hear that, you'd 00:49:27.230 --> 00:49:30.710 think, wow, that's ambitious. But you suddenly 00:49:30.710 --> 00:49:34.050 start to see the value. When you have a chat 00:49:34.050 --> 00:49:37.829 GPT agent that's goal -driven, it can plan, it 00:49:37.829 --> 00:49:41.500 can overcome obstacles. give you final outputs 00:49:41.500 --> 00:49:45.820 if it works around the clock and it's two thousand 00:49:45.820 --> 00:49:48.900 dollars a month a lot of companies are going 00:49:48.900 --> 00:49:53.079 to go for that it is interesting that we were 00:49:53.079 --> 00:49:56.260 told about that over a year ago and we haven't 00:49:56.260 --> 00:49:59.019 seen any of those types of subscriptions emerge 00:49:59.019 --> 00:50:01.320 which does tell me that they haven't been able 00:50:01.320 --> 00:50:03.500 to get the agents to a point where they can reliably 00:50:03.500 --> 00:50:08.039 deliver economically viable work like high quality 00:50:08.039 --> 00:50:10.440 work like you say like delegating to a junior 00:50:10.440 --> 00:50:13.820 consistently enough that somebody would pay two 00:50:13.820 --> 00:50:15.900 thousand dollars a month and go yeah i'm getting 00:50:15.900 --> 00:50:20.480 massive roi on this um probably again because 00:50:20.480 --> 00:50:23.960 we've seen so much progress in coding for them 00:50:23.960 --> 00:50:28.039 as software developers you know creating your 00:50:28.039 --> 00:50:31.119 own recursive improvement engine for your ai 00:50:31.119 --> 00:50:34.199 infrastructure and your models and your products 00:50:34.199 --> 00:50:38.039 is probably much more important than, you know, 00:50:38.039 --> 00:50:40.699 a reliable tool that can write and send emails 00:50:40.699 --> 00:50:44.159 out to segments of your database. But obviously 00:50:44.159 --> 00:50:46.679 they will come for that work eventually. But 00:50:46.679 --> 00:50:48.659 the fact that, yeah, we don't have a $2 ,000 00:50:48.659 --> 00:50:51.420 a month subscription tells me it's because they 00:50:51.420 --> 00:50:53.500 can't do it yet. Or maybe it's a compute thing 00:50:53.500 --> 00:50:56.340 because there's not enough compute as well for 00:50:56.340 --> 00:50:58.559 a lot of this stuff. So maybe when these massive 00:50:58.559 --> 00:51:01.039 data centers come online, they'll unleash all 00:51:01.039 --> 00:51:03.239 these agents because they'll be like, oh yeah, 00:51:03.320 --> 00:51:05.219 they always work really well. But the amount 00:51:05.219 --> 00:51:07.679 of compute they gobbled up was just made them 00:51:07.679 --> 00:51:11.420 completely, you know, they're just not feasible 00:51:11.420 --> 00:51:14.460 to use. So, yeah, interesting. I want to talk 00:51:14.460 --> 00:51:17.280 a little bit as part of this section about MCP, 00:51:17.380 --> 00:51:19.760 Model Context Protocol, because I know this is 00:51:19.760 --> 00:51:21.960 something else that you've been playing with 00:51:21.960 --> 00:51:25.179 as part of some of your experiments with Claude. 00:51:25.199 --> 00:51:26.820 And there's another podcast I listen to where 00:51:26.820 --> 00:51:29.119 they do some cool stuff with MCP. So can you 00:51:29.119 --> 00:51:32.519 briefly explain, Martin, to people what MCP is 00:51:32.519 --> 00:51:36.579 and how you've been able to? sort of access it 00:51:36.579 --> 00:51:39.440 and start using bits of the sort of infrastructure 00:51:39.440 --> 00:51:44.659 associated with mcps in the simplest form mcp 00:51:44.659 --> 00:51:52.579 is a way of giving access to a tool such as hubspot 00:51:52.579 --> 00:51:58.079 your crm and saying to the model here's all the 00:51:58.079 --> 00:52:01.349 things you can do on this platform and you can 00:52:01.349 --> 00:52:04.929 it's basically opening up the api so saying you 00:52:04.929 --> 00:52:07.210 can read the information on this platform you 00:52:07.210 --> 00:52:09.489 can edit the information you can delete you can 00:52:09.489 --> 00:52:15.090 create new things and lots of different software 00:52:15.090 --> 00:52:17.269 providers have opened up their tooling through 00:52:17.269 --> 00:52:21.289 mcps and mcps can connect as a standard industry 00:52:21.289 --> 00:52:25.070 -wide protocol now you can connect them to different 00:52:26.159 --> 00:52:28.440 different large language models i use them through 00:52:28.440 --> 00:52:32.659 the claude desktop app mostly i've also connected 00:52:32.659 --> 00:52:36.219 claude code as well but primarily and where i 00:52:36.219 --> 00:52:40.440 saw the biggest gains was creating an mcp that 00:52:40.440 --> 00:52:44.320 connects my hubspot account and client hubspot 00:52:44.320 --> 00:52:48.739 accounts to claude desktop which can then read 00:52:48.739 --> 00:52:54.139 write and can edit according to the api rules 00:52:54.670 --> 00:52:57.570 that i allow it to and the scopes that i define 00:52:57.570 --> 00:53:01.849 and the first win that i had with it was when 00:53:01.849 --> 00:53:04.869 i had claude being able to access all of my google 00:53:04.869 --> 00:53:08.429 suite so google drive gmail everywhere that i 00:53:08.429 --> 00:53:12.489 do my my work it can read all of that content 00:53:12.489 --> 00:53:17.409 and then in hubspot i gave it full scopes so 00:53:17.409 --> 00:53:23.219 it could do everything and i hadn't really been 00:53:23.219 --> 00:53:26.199 staying on top of my you know my my sales administration 00:53:26.199 --> 00:53:29.019 pipeline management wasn't really where it should 00:53:29.019 --> 00:53:32.380 have been so staying on top of deals keeping 00:53:32.380 --> 00:53:34.760 notes up to date stuff like that would have fallen 00:53:34.760 --> 00:53:39.559 behind and i just asked chat gpt to sorry claude 00:53:39.559 --> 00:53:45.739 to read all of my pipeline identified deals that 00:53:45.739 --> 00:53:48.739 had kind of been sitting in pipeline stages for 00:53:48.739 --> 00:53:52.019 a long time and then to go and read my emails 00:53:52.989 --> 00:53:55.389 and find if there was any updates. Go and check 00:53:55.389 --> 00:53:58.090 Google Drive, see if there was any updates. And 00:53:58.090 --> 00:54:02.170 I had seven deals that had gone a bit stale as 00:54:02.170 --> 00:54:05.150 far as HubSpot was concerned. And it went through, 00:54:05.250 --> 00:54:06.829 it read all of my emails, it read the Google 00:54:06.829 --> 00:54:08.949 Drive, came back and said, it looks like these 00:54:08.949 --> 00:54:10.849 things need updating, and it outlined the seven 00:54:10.849 --> 00:54:14.369 deals. And I said, great, make those updates. 00:54:15.070 --> 00:54:17.730 And in one shot with zero errors, it updated 00:54:17.730 --> 00:54:21.130 my whole deal pipeline, moving deals to closed 00:54:21.130 --> 00:54:24.829 one, closed last, moving it to quoted, updating 00:54:24.829 --> 00:54:28.750 the notes, changing the amounts. It did the whole 00:54:28.750 --> 00:54:33.230 thing in, well, from start to finish, it took 00:54:33.230 --> 00:54:37.250 minutes. And there were a few bits where I had 00:54:37.250 --> 00:54:39.590 to give it a bit of extra input because it would 00:54:39.590 --> 00:54:41.630 say I couldn't find anything about this deal 00:54:41.630 --> 00:54:43.860 and I would just... write the notes about what 00:54:43.860 --> 00:54:47.559 needed to be updated but it did my whole pipeline 00:54:47.559 --> 00:54:52.219 and that now is my my kind of go -to way of managing 00:54:52.219 --> 00:54:55.719 my sales pipeline i just use claw to kind of 00:54:55.719 --> 00:54:57.980 manage that and stay on top of that um day to 00:54:57.980 --> 00:55:01.280 day so you're using you are genuinely using that 00:55:01.280 --> 00:55:03.900 not as a fun thing that you set up because you 00:55:03.900 --> 00:55:05.599 were like oh i wonder how well this will work 00:55:05.599 --> 00:55:07.360 but that is actually part of your workflow now 00:55:07.360 --> 00:55:11.460 100 absolutely it's a that regular but it's the 00:55:11.460 --> 00:55:13.579 only way that i really update my because it makes 00:55:13.579 --> 00:55:16.480 better notes that's the thing you know how you 00:55:16.480 --> 00:55:19.360 go into a crm and my crm is basically just me 00:55:19.360 --> 00:55:22.019 uh that needs to access it but the amount of 00:55:22.019 --> 00:55:25.179 times that i go to a client and i'm working on 00:55:25.179 --> 00:55:27.820 their their pipeline and the deal structure and 00:55:27.820 --> 00:55:31.199 we click into a deal and you see some notes and 00:55:31.199 --> 00:55:35.179 the notes are terrible the call notes are incomplete 00:55:35.179 --> 00:55:36.920 it's like what was discussed there's nothing 00:55:36.920 --> 00:55:40.489 there well with claude it can read the emails 00:55:40.489 --> 00:55:42.789 it can read the transcript because i recorded 00:55:42.789 --> 00:55:45.610 the meeting in google meet which was linked to 00:55:45.610 --> 00:55:48.550 my calendar and it can now go off and read the 00:55:48.550 --> 00:55:52.110 transcript so it it captures all of that leaves 00:55:52.110 --> 00:55:55.610 really robust notes and writes really good next 00:55:55.610 --> 00:55:58.369 steps do you know what martin we won't do it 00:55:58.369 --> 00:56:00.070 today because we're trying to do a whistle stop 00:56:00.070 --> 00:56:02.710 tour but maybe in the next episode that's right 00:56:02.710 --> 00:56:04.710 listeners this isn't a one -off we're bringing 00:56:04.710 --> 00:56:07.949 it back properly uh in the next episode Maybe 00:56:07.949 --> 00:56:10.289 we can do a bit of a deep dive on MCP because 00:56:10.289 --> 00:56:13.230 I think it's one of those things that you've 00:56:13.230 --> 00:56:16.110 got to download code and you have to like give 00:56:16.110 --> 00:56:17.690 it access to your computer and you're like, oh 00:56:17.690 --> 00:56:19.710 no, I'm scared of that. And how do I set up the 00:56:19.710 --> 00:56:23.289 config files? Like it's not simple to do, but 00:56:23.289 --> 00:56:26.010 a little bit of knowledge goes a long way. So 00:56:26.010 --> 00:56:27.730 maybe you can just talk us through on the next 00:56:27.730 --> 00:56:30.670 episode, how you set it up. We talk a little 00:56:30.670 --> 00:56:32.489 bit more about how to get connectors working 00:56:32.489 --> 00:56:34.690 because there are some fantastic things you could 00:56:34.690 --> 00:56:36.369 do. I mean, if you're listening to this, you're 00:56:36.369 --> 00:56:39.539 probably thinking, how easy it would be to semi 00:56:39.539 --> 00:56:42.179 -automate elements of your customer support workflow 00:56:42.179 --> 00:56:44.380 like this, right? You get a support ticket in, 00:56:44.480 --> 00:56:47.280 analyze the support ticket, check in with the 00:56:47.280 --> 00:56:49.679 human. This is the approach I'm going to take. 00:56:49.920 --> 00:56:53.539 Send the email, make an update in the CRM as 00:56:53.539 --> 00:56:55.440 to what message was sent. Like there's so many 00:56:55.440 --> 00:56:59.539 interesting things you could do here. And my 00:56:59.539 --> 00:57:05.320 sense is this is probably not wasted time. Because 00:57:05.320 --> 00:57:11.340 as MCP becomes more prevalent and you want to 00:57:11.340 --> 00:57:15.559 give agents access to more and more tools, what's 00:57:15.559 --> 00:57:18.559 going to be important is the use cases that you've 00:57:18.559 --> 00:57:23.239 defined and how you want the tools to interact 00:57:23.239 --> 00:57:26.119 to get the results that you want. And I'm guessing 00:57:26.119 --> 00:57:28.260 the prompts, some of this is still very prompt 00:57:28.260 --> 00:57:30.440 driven in terms of being quite explicit about 00:57:30.440 --> 00:57:34.579 what the goal is. how information should be used, 00:57:34.679 --> 00:57:37.960 etc. So yeah, maybe we make that part of a future 00:57:37.960 --> 00:57:41.780 episode. Yeah, I'll start off by saying my experience 00:57:41.780 --> 00:57:45.460 of configuring MCP is one of constant frustration. 00:57:46.139 --> 00:57:49.159 But when it works and when it is set up, it's 00:57:49.159 --> 00:57:52.280 great. I was excited there, Martin. Then you're 00:57:52.280 --> 00:57:54.059 like, oh, there's a but. There's always a but. 00:57:54.340 --> 00:57:58.159 There is. That's a nice segue to the last part 00:57:58.159 --> 00:58:01.880 of today's podcast, which is... We wanted to 00:58:01.880 --> 00:58:05.699 discuss a bit the state of play for AI for marketers, 00:58:05.860 --> 00:58:08.780 right? If you've listened to our previous episodes, 00:58:08.840 --> 00:58:11.219 you'll know that we have been always testing 00:58:11.219 --> 00:58:13.679 tools to try and be at the cutting edge of what 00:58:13.679 --> 00:58:16.019 they can do to speed up work, make it better 00:58:16.019 --> 00:58:18.880 quality for us, but also where they break and 00:58:18.880 --> 00:58:21.780 can't be relied upon yet. We heard from Martin 00:58:21.780 --> 00:58:24.000 there that NCPs are cool when they work, but 00:58:24.000 --> 00:58:26.860 they're really hard to set up. And so your expectations 00:58:26.860 --> 00:58:29.420 should be managed, I think, in a huge amount 00:58:29.420 --> 00:58:32.400 of areas. it's still an expectations need to 00:58:32.400 --> 00:58:37.039 be managed um scenario which is i think maybe 00:58:37.039 --> 00:58:39.679 a a reflection on us as humans really martin 00:58:39.679 --> 00:58:42.900 we're so like if you'd have told me two years 00:58:42.900 --> 00:58:44.980 ago you were going to have this app on your computer 00:58:44.980 --> 00:58:47.539 that would allow you to like engage in natural 00:58:47.539 --> 00:58:50.159 language with all your other apps and then it 00:58:50.159 --> 00:58:51.900 would go do a load of work for you i'd have been 00:58:51.900 --> 00:58:54.179 like science fiction and then chat gpt came out 00:58:54.179 --> 00:58:57.800 and everything gets normalized very quickly but 00:58:57.800 --> 00:59:00.599 um let's start with content creation and writing 00:59:00.599 --> 00:59:05.440 this is something that i do a lot of in my role 00:59:05.440 --> 00:59:08.239 at supreme group and i'm testing tools constantly 00:59:08.239 --> 00:59:11.599 and we talked about it a little bit earlier martin 00:59:11.599 --> 00:59:16.659 how good is ai at writing stuff and i'm still 00:59:16.659 --> 00:59:19.440 a bit disappointed if i'm honest and there was 00:59:19.440 --> 00:59:21.659 a report out recently i think it was in the harford 00:59:21.659 --> 00:59:24.639 business review about the concept of work slop 00:59:24.639 --> 00:59:29.789 so People using AI to create documents, memos, 00:59:29.789 --> 00:59:32.769 maybe blog posts, drafts, presentations, and 00:59:32.769 --> 00:59:35.469 sending it to colleagues, but for whatever reason, 00:59:35.489 --> 00:59:38.369 not really paying enough attention to the outputs 00:59:38.369 --> 00:59:40.489 and improving the outputs. And then the poor 00:59:40.489 --> 00:59:44.050 colleague has to spend hours fixing what they 00:59:44.050 --> 00:59:49.889 call the AI work slop. And I think this is very 00:59:49.889 --> 00:59:52.449 indicative of where the AI tools still are when 00:59:52.449 --> 00:59:54.769 it comes to writing and content creation. They 00:59:54.769 --> 00:59:59.409 are amazing at writing sensible sounding sentences 00:59:59.409 --> 01:00:04.809 that mean nothing. Like genuinely. Even with 01:00:04.809 --> 01:00:07.730 a very prescriptive brief, the style guide we 01:00:07.730 --> 01:00:10.929 use in our custom GPTs is like 30 bullets long. 01:00:11.050 --> 01:00:14.389 Like it does not leave much to chance. It's incredibly 01:00:14.389 --> 01:00:18.690 prescriptive about the importance of the logical 01:00:18.690 --> 01:00:25.289 progression of one statement to the next. how, 01:00:25.289 --> 01:00:27.849 even how a paragraph should be structured to 01:00:27.849 --> 01:00:30.730 build a logical argument, even from within a 01:00:30.730 --> 01:00:33.150 paragraph and from paragraph to paragraph. And 01:00:33.150 --> 01:00:36.590 it is very good at saying, if we look at this 01:00:36.590 --> 01:00:39.010 thing A and this thing B, then we obviously get 01:00:39.010 --> 01:00:41.489 C and it's like A and B are not related. And 01:00:41.489 --> 01:00:43.530 even if they were, they would not generate C. 01:00:43.690 --> 01:00:45.929 But the way it's written just sounds so darn 01:00:45.929 --> 01:00:47.889 plausible that if you're not paying attention, 01:00:48.309 --> 01:00:51.860 you just let it go. So I think. That's where 01:00:51.860 --> 01:00:53.420 we are with writing in the moment. I have two 01:00:53.420 --> 01:00:56.500 custom GPTs. I have a writing GPT and I have 01:00:56.500 --> 01:00:59.079 an editor GPT, which has even more bullet points 01:00:59.079 --> 01:01:01.780 that it has to follow, that basically massively 01:01:01.780 --> 01:01:05.639 pulls to pieces everything the writing GPT produced 01:01:05.639 --> 01:01:08.820 so that we can then go back and improve the output, 01:01:09.019 --> 01:01:12.599 mostly to improve the logical flow of the arguments 01:01:12.599 --> 01:01:15.599 that it's making and to stop it making these 01:01:15.599 --> 01:01:19.019 fluffy, sexy sounding statements that don't mean 01:01:19.019 --> 01:01:24.030 anything. And that kind of works, but still needs 01:01:24.030 --> 01:01:26.769 a lot of human intervention. So that's where 01:01:26.769 --> 01:01:29.590 we are on writing. And if you've suffered with 01:01:29.590 --> 01:01:33.170 work slop, like either you've accidentally produced 01:01:33.170 --> 01:01:36.610 some or you've been sent some, I think it's because 01:01:36.610 --> 01:01:40.530 for very busy people, this stuff sounds so plausible 01:01:40.530 --> 01:01:43.070 that it's very easy to copy and paste into a 01:01:43.070 --> 01:01:45.309 Slack message and email a document and send it 01:01:45.309 --> 01:01:49.789 on. And it is probably more important than ever. 01:01:50.199 --> 01:01:53.059 to critically evaluate AI outputs, probably more 01:01:53.059 --> 01:01:55.880 so than it even was a year ago, because AI has 01:01:55.880 --> 01:01:58.900 got that much better at sounding plausible. But 01:01:58.900 --> 01:02:00.719 you have to interrogate every sentence and every 01:02:00.719 --> 01:02:03.980 paragraph, honestly. So you have to be willing 01:02:03.980 --> 01:02:06.420 to invest the time and energy to do that. The 01:02:06.420 --> 01:02:08.539 other thing is, I still believe you very much 01:02:08.539 --> 01:02:11.139 have to have the expertise and experience to 01:02:11.139 --> 01:02:14.699 know what good looks like. And I mean that from 01:02:14.699 --> 01:02:17.920 a writing quality perspective. Like if you... 01:02:18.809 --> 01:02:21.590 I'm lucky. I've spent the last 15 years training 01:02:21.590 --> 01:02:26.389 marketing copywriters and I was very well trained 01:02:26.389 --> 01:02:29.030 myself, fortunately, by the people I worked with 01:02:29.030 --> 01:02:30.949 early in my career. So I feel like I've got a 01:02:30.949 --> 01:02:33.349 good sense for what good looks like. But if I 01:02:33.349 --> 01:02:36.050 didn't, how would I know that this stuff didn't 01:02:36.050 --> 01:02:38.110 make sense, didn't build a logical argument, 01:02:38.230 --> 01:02:40.309 wasn't a persuasive, compelling story, didn't 01:02:40.309 --> 01:02:42.250 keep the audience at the forefront, was connecting 01:02:42.250 --> 01:02:44.110 points together that should never be connected 01:02:44.110 --> 01:02:47.880 because it doesn't make any sense. It's one of 01:02:47.880 --> 01:02:50.420 those things that I still feel like it's probably 01:02:50.420 --> 01:02:54.440 best applied in the hands of an expert who can 01:02:54.440 --> 01:02:57.019 tell whether or not the outputs are any good. 01:02:57.440 --> 01:03:00.239 And that restricts its usability because you 01:03:00.239 --> 01:03:04.039 can't just, you know, hire a 21 year old, give 01:03:04.039 --> 01:03:06.320 them chat GPT and say, right, now you can do 01:03:06.320 --> 01:03:07.820 all my content marketing because you're just 01:03:07.820 --> 01:03:10.019 not going to get good stuff out of it. No, it 01:03:10.019 --> 01:03:15.219 can make experts faster and help them work more 01:03:15.219 --> 01:03:18.400 efficiently and improve. sort of their outputs 01:03:18.400 --> 01:03:22.019 because they can spot the errors and fix them 01:03:22.019 --> 01:03:24.380 quicker and what have you but also if you put 01:03:24.380 --> 01:03:25.760 it in the hands of someone that doesn't know 01:03:25.760 --> 01:03:31.440 particularly around high technical topics because 01:03:31.440 --> 01:03:33.820 and there is still the risk of hallucination 01:03:33.820 --> 01:03:37.920 as well i had a deep research brief that i asked 01:03:37.920 --> 01:03:41.019 open ai to to put together related to the sport 01:03:41.019 --> 01:03:43.820 of power lifting recently and it came back with 01:03:43.820 --> 01:03:48.340 a number which was a related to a competition 01:03:48.340 --> 01:03:51.420 score that it said was relevant to a particular 01:03:51.420 --> 01:03:55.460 age category in a particular group and at first 01:03:55.460 --> 01:03:57.280 glance the whole document read really well and 01:03:57.280 --> 01:04:00.739 i came away feeling quite informed but upon further 01:04:00.739 --> 01:04:03.159 inspection these there were two numbers in it 01:04:03.159 --> 01:04:07.980 and they were wrong it had misread the pdf everything 01:04:07.980 --> 01:04:10.539 else in it was more or less true as far as i 01:04:10.539 --> 01:04:14.420 could see um I didn't scrutinize it by line by 01:04:14.420 --> 01:04:16.500 line, but I did spot these two things that were 01:04:16.500 --> 01:04:20.260 wrong and thus made the whole thing irrelevant 01:04:20.260 --> 01:04:24.579 or incorrect at quite a significant level. If 01:04:24.579 --> 01:04:26.039 you were to hand this off to someone and say, 01:04:26.139 --> 01:04:29.880 here, read this, that would completely have shifted 01:04:29.880 --> 01:04:34.219 their perspective on what was required in competition, 01:04:34.519 --> 01:04:38.039 which was the main thrust of the document, right? 01:04:41.110 --> 01:04:43.050 it's right you're right in what you say i think 01:04:43.050 --> 01:04:48.590 the it's very good for doing a lot of the legwork 01:04:48.590 --> 01:04:51.710 but it's you cannot hand off entirely and outsource 01:04:51.710 --> 01:04:54.610 all of your thinking all of your scrutiny and 01:04:54.610 --> 01:04:58.030 you've still got to be on it yeah it's paul reuter 01:04:58.030 --> 01:05:01.349 on the um i think their podcast is called the 01:05:01.349 --> 01:05:04.849 ai the artificial intelligence show now because 01:05:04.849 --> 01:05:06.469 i think they changed the name of it but they 01:05:06.469 --> 01:05:09.739 talk about this massive missing verification 01:05:09.739 --> 01:05:13.679 step like ai can help us produce stuff at scale 01:05:13.679 --> 01:05:17.559 like massive volumes of it but the human time 01:05:17.559 --> 01:05:20.239 and effort and expertise required to verify the 01:05:20.239 --> 01:05:23.619 outputs is still very real yeah in my experience 01:05:23.619 --> 01:05:27.139 and we'll maybe we talk a little bit about some 01:05:27.139 --> 01:05:29.079 of the stuff we've been playing with with um 01:05:29.079 --> 01:05:33.400 vibe coding and stuff later but um yeah finding 01:05:33.400 --> 01:05:36.179 a way to verify the outputs is the new bottleneck 01:05:36.179 --> 01:05:39.849 because I don't think the AI tools have got meaningfully 01:05:39.849 --> 01:05:42.650 better. Like, I really don't. Even for GPT -5 01:05:42.650 --> 01:05:46.030 being better at not hallucinating and not making 01:05:46.030 --> 01:05:49.730 stuff up, it still strings arguments together 01:05:49.730 --> 01:05:54.889 that are not logical. And so, yeah, we need to 01:05:54.889 --> 01:05:58.730 do something there. And, yeah, if you're producing 01:05:58.730 --> 01:06:02.769 three blog posts, five blog posts a week, and 01:06:02.769 --> 01:06:06.170 you're a small team, I can almost guarantee you 01:06:06.170 --> 01:06:09.719 that... there's crap in those because the humans 01:06:09.719 --> 01:06:13.179 you'd max out your human capability to verify 01:06:13.179 --> 01:06:15.579 and and edit them and make sure they're good 01:06:15.579 --> 01:06:17.300 because that's still the current state of play 01:06:17.300 --> 01:06:20.639 for content in my mind but um imaging's an area 01:06:20.639 --> 01:06:22.800 where things have perhaps been a bit more interesting 01:06:22.800 --> 01:06:24.760 and got a bit better what your thoughts on the 01:06:24.760 --> 01:06:27.239 state of play and imaging mine well the fingers 01:06:27.239 --> 01:06:32.420 are fixed so that's nice we now have five fingers 01:06:32.420 --> 01:06:36.739 fairly consistently imaging has come along incredibly 01:06:36.739 --> 01:06:41.920 i think when dali 3 came out in chat gpt text 01:06:41.920 --> 01:06:45.480 as a as a you know written form on on images 01:06:45.480 --> 01:06:48.159 was was more or less solved for the most part 01:06:48.159 --> 01:06:50.579 and that was good but then there was a weird 01:06:50.579 --> 01:06:59.519 stylistic thing um then 4o had image um kind 01:06:59.519 --> 01:07:02.059 of native image generation and you could upload 01:07:02.059 --> 01:07:05.440 images and get it to to change aspects of that 01:07:05.440 --> 01:07:10.159 image and an example i've always liked in kind 01:07:10.159 --> 01:07:12.400 of personal uses is taking a photo of a room 01:07:12.400 --> 01:07:15.000 and then asking it to recreate that room with 01:07:15.000 --> 01:07:18.280 different uh paint or wallpaper or you know different 01:07:18.280 --> 01:07:20.980 sofa or whatever it may be and it was it was 01:07:20.980 --> 01:07:22.300 very good for that so you've got this kind of 01:07:22.300 --> 01:07:26.559 consistency across the models then you've got 01:07:26.559 --> 01:07:29.360 model providers like mid journey whose aesthetic 01:07:29.360 --> 01:07:32.610 styling is just i still think that's unmatched 01:07:32.610 --> 01:07:36.630 if i'm if i'm honest i think their image quality 01:07:36.630 --> 01:07:40.610 from midgen is still in very very good but struggles 01:07:40.610 --> 01:07:44.550 with things like image editing and text in image 01:07:44.550 --> 01:07:47.849 generation and then more recently we've had the 01:07:47.849 --> 01:07:53.170 launch of nano banana from google which is just 01:07:53.170 --> 01:07:56.989 superb i think that model is fantastic you can 01:07:56.989 --> 01:07:59.889 upload your image and ask it natural language 01:07:59.889 --> 01:08:03.670 just change this element or make the character 01:08:03.670 --> 01:08:06.030 face a different direction and it keeps this 01:08:06.030 --> 01:08:09.130 character consistency incredibly accurately some 01:08:09.130 --> 01:08:11.349 of the the outputs that i've seen from that have 01:08:11.349 --> 01:08:15.690 been really i mean if you haven't played with 01:08:15.690 --> 01:08:19.550 it yet go to aistudio .google .com and just have 01:08:19.550 --> 01:08:23.090 a play but from a use case perspective from a 01:08:23.090 --> 01:08:26.250 marketing use case being able to create mock 01:08:26.250 --> 01:08:29.720 -ups or even final outputs, if you're so inclined, 01:08:29.920 --> 01:08:33.899 has never been easier. I've got one client that 01:08:33.899 --> 01:08:40.800 they provide festive lighting for events and 01:08:40.800 --> 01:08:44.079 decorative lighting for events. And people will 01:08:44.079 --> 01:08:48.520 send them photos of a space. It might be a shopping 01:08:48.520 --> 01:08:50.079 center or something like that. And they'll say, 01:08:50.180 --> 01:08:52.500 this is the space we want. We want this with 01:08:52.500 --> 01:08:55.239 some Christmas lighting. What could it look like? 01:08:55.659 --> 01:08:57.079 Now, it used to be that they would send this 01:08:57.079 --> 01:09:01.060 off to a designer who, you know, the design team 01:09:01.060 --> 01:09:02.680 was typically stretched. They would create their 01:09:02.680 --> 01:09:04.939 mock -ups. There might be a seven -day turnaround 01:09:04.939 --> 01:09:07.300 in creating that mock -up to get it to send back 01:09:07.300 --> 01:09:12.659 to the client for a quote alongside, oh, here's 01:09:12.659 --> 01:09:15.460 what we think it could look like. They tried 01:09:15.460 --> 01:09:19.640 it with GPT -4 .0 with the native image generation, 01:09:19.880 --> 01:09:23.859 and it did a decent job, but they might have 01:09:23.859 --> 01:09:26.619 to spend an hour going back and forth iterating 01:09:26.619 --> 01:09:28.880 restarting the process because the model gets 01:09:28.880 --> 01:09:33.000 a bit lost now with nano banana they can often 01:09:33.000 --> 01:09:35.840 one shot it which means from writing the prompt 01:09:35.840 --> 01:09:41.520 to hitting enter it takes about 13 seconds and 01:09:41.520 --> 01:09:45.560 they the the difference in quality and speed 01:09:45.560 --> 01:09:48.939 of production in service delivery for the client 01:09:48.939 --> 01:09:54.060 well unmatched really it's incomparable yeah 01:09:54.060 --> 01:09:57.300 i think for your average marketer the the major 01:09:57.300 --> 01:09:59.880 improvement with nano banana it's not great at 01:09:59.880 --> 01:10:01.819 image generation per se i think you're right 01:10:01.819 --> 01:10:03.500 like if i was starting from scratch i'd still 01:10:03.500 --> 01:10:08.319 go for mid journey um but or maybe check gpt 01:10:08.319 --> 01:10:12.239 4o but um not 4o but now gpt 5 which is probably 01:10:12.239 --> 01:10:14.460 still using the underlying 4o image model but 01:10:14.460 --> 01:10:17.039 um it's the stealability of nano banana it's 01:10:17.039 --> 01:10:19.260 being able to give it a picture of yourself and 01:10:19.260 --> 01:10:22.300 and asking for you to be in different clothes 01:10:22.300 --> 01:10:27.029 um and they're looking almost perfect um which 01:10:27.029 --> 01:10:29.630 is unbelievable it's being able to give it a 01:10:29.630 --> 01:10:31.569 picture of a product and then ask it to show 01:10:31.569 --> 01:10:33.510 that product in different settings and the product 01:10:33.510 --> 01:10:36.729 image that you gave it stays the same right this 01:10:36.729 --> 01:10:38.869 was critical like maintaining that consistency 01:10:38.869 --> 01:10:41.649 was nigh on impossible and without it you can't 01:10:41.649 --> 01:10:45.369 really do anything interesting um i think these 01:10:45.369 --> 01:10:47.029 things will only get better and ultimately you 01:10:47.029 --> 01:10:50.029 and i are not designers martin we're we're marketers 01:10:50.029 --> 01:10:53.279 who want to do quick things with images that 01:10:53.279 --> 01:10:56.699 usually we'd have to outsource to a design freelancer 01:10:56.699 --> 01:10:58.720 or user and internal design team but then there's 01:10:58.720 --> 01:11:00.800 back and forth and maybe the design team's busy 01:11:00.800 --> 01:11:02.859 and what have you so being able to just create 01:11:02.859 --> 01:11:05.520 those images and have them be high quality and 01:11:05.520 --> 01:11:07.880 not have to go through that process that might 01:11:07.880 --> 01:11:11.800 mean cost and time is extremely valuable but 01:11:11.800 --> 01:11:15.720 we can't speak to you know how useful these things 01:11:15.720 --> 01:11:19.029 are in adobe photoshop for example i mean a lot 01:11:19.029 --> 01:11:21.369 of people calling nano banana the photoshop killer 01:11:21.369 --> 01:11:23.329 which is nonsense right like if you know how 01:11:23.329 --> 01:11:25.430 to use photoshop well there are incredible things 01:11:25.430 --> 01:11:28.350 that you can do with fine precision that you 01:11:28.350 --> 01:11:31.569 just cannot do in nano banana or any other ai 01:11:31.569 --> 01:11:34.430 model yet and at the moment they still can't 01:11:34.430 --> 01:11:36.810 output in layers which is the critical piece 01:11:36.810 --> 01:11:39.229 for most designers to be able to break that image 01:11:39.229 --> 01:11:41.390 down and so that you can manipulate specific 01:11:41.390 --> 01:11:45.630 elements in the image however you want to do 01:11:45.630 --> 01:11:48.409 but But for your average marketer, it's great. 01:11:49.149 --> 01:11:52.329 So much of this is about the application and 01:11:52.329 --> 01:11:54.369 the use case of the image at the end of the day. 01:11:54.470 --> 01:11:57.369 If you've got a social media post which has an 01:11:57.369 --> 01:12:01.229 expected lifespan of about three hours, you don't 01:12:01.229 --> 01:12:03.710 want to give it to a designer who's going to 01:12:03.710 --> 01:12:08.189 spend a day putting it together. I had a client 01:12:08.189 --> 01:12:10.970 that took an image of a warehouse. They work 01:12:10.970 --> 01:12:14.609 in logistics industry. They wanted to do a cutaway. 01:12:14.989 --> 01:12:18.569 now normally this would have taken ad modelers 01:12:18.569 --> 01:12:21.930 3d designers that have had to have renders but 01:12:21.930 --> 01:12:24.770 instead they took a drone photo of a warehouse 01:12:24.770 --> 01:12:26.989 that they had in the library put it into nano 01:12:26.989 --> 01:12:30.550 banana and said create an image of this with 01:12:30.550 --> 01:12:35.050 a cutaway showing the racking inside the warehouse 01:12:35.050 --> 01:12:37.890 with a forklift truck and change the weather 01:12:37.890 --> 01:12:41.029 so that it's um i think it was like a gloomy 01:12:41.029 --> 01:12:43.289 day so they wanted it a bit brighter a bit sunnier 01:12:43.289 --> 01:12:46.489 it did it in 13 seconds and they said that was 01:12:46.489 --> 01:12:50.369 going to be used on their next e -shot and it 01:12:50.369 --> 01:12:53.149 displayed what they needed you know it wasn't 01:12:53.149 --> 01:12:56.229 showing their particular product they hadn't 01:12:56.229 --> 01:12:59.310 done it for that but it was just showing a cutaway 01:12:59.310 --> 01:13:01.310 of a warehouse and the logistics industry and 01:13:01.310 --> 01:13:03.350 it did a really nice example and it was good 01:13:03.350 --> 01:13:06.810 to go in in under a minute yeah i think this 01:13:06.810 --> 01:13:09.819 leads us on to You know, images have got better 01:13:09.819 --> 01:13:11.800 and therefore, as you might expect, video has 01:13:11.800 --> 01:13:15.960 as well. You know, lots of images shown, 24 images 01:13:15.960 --> 01:13:19.819 a second. But there's been some really impressive 01:13:19.819 --> 01:13:23.659 leaps with Google's Veo 3 model and now Sora 01:13:23.659 --> 01:13:29.079 2 from OpenAI released this week. Kling has got 01:13:29.079 --> 01:13:32.470 a new video model as well. a lot of these video 01:13:32.470 --> 01:13:35.510 models now generate the sound to go with the 01:13:35.510 --> 01:13:37.369 videos and i've started to see people with saura 01:13:37.369 --> 01:13:41.689 2 trying to create like anime cartoons that could 01:13:41.689 --> 01:13:44.850 be like five minutes long you only get 15 second 01:13:44.850 --> 01:13:47.390 outputs so you're certainly going to burn some 01:13:47.390 --> 01:13:50.289 some credits or uh you're gonna have to have 01:13:50.289 --> 01:13:52.229 multiple accounts probably to be able to do that 01:13:52.229 --> 01:13:55.369 in any reasonable time but um what's what's your 01:13:55.369 --> 01:13:59.220 thought on the progress in video models fantastic 01:13:59.220 --> 01:14:02.399 and specifically the well the launch of sora 01:14:02.399 --> 01:14:06.000 this week was interesting because of one feature 01:14:06.000 --> 01:14:07.899 in particular i think and that's the cameo feature 01:14:07.899 --> 01:14:12.579 and that's the ability to insert yourself or 01:14:12.579 --> 01:14:16.060 an object could be you it could be your dog into 01:14:16.060 --> 01:14:18.920 a scene that it has generated and this has created 01:14:18.920 --> 01:14:22.420 a bunch of viral videos of sam altman in various 01:14:22.420 --> 01:14:24.420 scenarios i don't know if you saw the mock -up 01:14:24.420 --> 01:14:26.930 of them Someone had done Sam Altman stealing 01:14:26.930 --> 01:14:32.470 GPUs in Target and a bunch of other silly scenarios. 01:14:32.909 --> 01:14:35.810 But that's a really powerful thing. That's leading 01:14:35.810 --> 01:14:38.810 to them creating their own Zora social network, 01:14:39.010 --> 01:14:42.470 which is currently available only in the US and 01:14:42.470 --> 01:14:45.430 Canada. But that's what they're looking at this 01:14:45.430 --> 01:14:49.909 being a content generator. Again, going back 01:14:49.909 --> 01:14:53.630 to the attention economy. How can you as a brand? 01:14:54.510 --> 01:14:58.289 think about inserting your products into ai generated 01:14:58.289 --> 01:15:01.810 content that people are consuming on a on a feed 01:15:01.810 --> 01:15:04.930 much like instagram what are the opportunities 01:15:04.930 --> 01:15:06.689 there i think that's going to be something that 01:15:06.689 --> 01:15:10.310 we can explore in a future episode for sure um 01:15:10.310 --> 01:15:12.569 but yeah the quality of the outputs are amazing 01:15:12.569 --> 01:15:17.270 vo3 or i think that's the sound the first time 01:15:17.270 --> 01:15:19.529 i heard or generated something with the sound 01:15:19.529 --> 01:15:26.079 that is one of those holy where you just go, 01:15:26.220 --> 01:15:29.619 oh, wow, this is pretty good. And again, as with 01:15:29.619 --> 01:15:31.180 all of these things, it's the worst it's ever 01:15:31.180 --> 01:15:34.939 going to be from this point on. So fast forward 01:15:34.939 --> 01:15:40.340 18 months and wow, we're in a whole new world. 01:15:40.500 --> 01:15:43.180 Put that into the hands of people that think 01:15:43.180 --> 01:15:47.039 in video, video production people. I don't, you 01:15:47.039 --> 01:15:49.619 know, I'm more of a text -based person. But if 01:15:49.619 --> 01:15:54.800 that's the way that you think creatively, It's 01:15:54.800 --> 01:15:57.800 a huge unlock and it's an amazing tool for you 01:15:57.800 --> 01:16:03.800 to be able to express and create a new dimension, 01:16:04.079 --> 01:16:06.439 so to speak. I think that's the big part of it, 01:16:06.479 --> 01:16:08.899 isn't it? It's the creativity that it unleashes. 01:16:09.220 --> 01:16:12.239 I mean, there's a bunch of aspects to this. So 01:16:12.239 --> 01:16:16.159 the first one is OpenAI were a bit criticized 01:16:16.159 --> 01:16:18.840 for taking the powerful models they've developed 01:16:18.840 --> 01:16:24.239 and then launching a social media app. so that 01:16:24.239 --> 01:16:27.300 people can create ai slop and send it to each 01:16:27.300 --> 01:16:29.840 other and that that would be a good use of humans 01:16:29.840 --> 01:16:33.640 ingenuity and time on this planet um which i 01:16:33.640 --> 01:16:35.779 understand as a criticism it's a bit meta -like 01:16:35.779 --> 01:16:40.260 in terms of like doing it but he in fairness 01:16:40.260 --> 01:16:42.300 sam altman posted and i'm not going to quote 01:16:42.300 --> 01:16:43.739 this directly because i don't have it in front 01:16:43.739 --> 01:16:46.779 me but in essence he put on twitter stroke x 01:16:46.779 --> 01:16:51.260 that he completely understands the point that 01:16:51.260 --> 01:16:52.439 people are making because i think one of the 01:16:52.439 --> 01:16:55.640 posts was like You promised us AGI, you promised 01:16:55.640 --> 01:16:59.300 us an end to disease, and you gave us an AI slop 01:16:59.300 --> 01:17:02.680 social app. And his response was basically, yeah, 01:17:02.720 --> 01:17:07.020 because building those systems, those data centers 01:17:07.020 --> 01:17:11.060 that can cure cancer costs lots of money. And 01:17:11.060 --> 01:17:12.939 we have to have a mechanism of making money. 01:17:13.260 --> 01:17:17.140 And ultimately, I think Gemini went to number 01:17:17.140 --> 01:17:22.039 one on the app charts. on ios and it was because 01:17:22.039 --> 01:17:24.020 everybody wanted down a banana because it was 01:17:24.020 --> 01:17:26.420 fun to be able to like take a picture of your 01:17:26.420 --> 01:17:29.260 friend and then put them in strange scenarios 01:17:29.260 --> 01:17:31.939 and it would and it would produce that output 01:17:31.939 --> 01:17:34.760 for you and basically sora 2 is the video version 01:17:34.760 --> 01:17:37.460 of that with its cameo feature and you can see 01:17:37.460 --> 01:17:39.680 why that is probably going to be really popular 01:17:39.680 --> 01:17:42.020 and i doubt they'll be able to there'll be nowhere 01:17:42.020 --> 01:17:44.180 near able to roll this out fast enough for the 01:17:44.180 --> 01:17:47.619 masses that want to get in on the act so there's 01:17:47.619 --> 01:17:51.250 sort of a you know social media dopamine hits 01:17:51.250 --> 01:17:52.989 we talked about it a bit at the beginning i'm 01:17:52.989 --> 01:17:55.130 a bit sad about all of that but it's kind of 01:17:55.130 --> 01:17:59.350 how humans operate and our phones and these tech 01:17:59.350 --> 01:18:01.390 companies have certainly learned how to keep 01:18:01.390 --> 01:18:03.670 our attentions on these platforms there's a reason 01:18:03.670 --> 01:18:06.310 that people doom scroll tiktok like clearly there's 01:18:06.310 --> 01:18:08.989 something about this stuff that humans like we 01:18:08.989 --> 01:18:11.050 might not like it and we may feel that we've 01:18:11.050 --> 01:18:13.569 been hacked um our brains have been hacked to 01:18:13.569 --> 01:18:16.010 make us do it but it works um and so people are 01:18:16.010 --> 01:18:17.649 going to do it because it because it works but 01:18:18.140 --> 01:18:21.380 When Veo 3 came out, and maybe this is going 01:18:21.380 --> 01:18:24.680 to be easier in like B2C than B2B, but some people 01:18:24.680 --> 01:18:26.119 did some really interesting things. Did you, 01:18:26.239 --> 01:18:31.279 there were quite a lot of sort of fake, like 01:18:31.279 --> 01:18:35.020 influencer type point of view type videos. And 01:18:35.020 --> 01:18:37.779 there was, there's one of a, I think it's of 01:18:37.779 --> 01:18:41.100 a gorilla and it was created by a dentist to 01:18:41.100 --> 01:18:44.260 promote his private dentistry in the States. 01:18:44.520 --> 01:18:47.380 And he's got this gorilla like. jumping out of 01:18:47.380 --> 01:18:50.560 airplanes and doing all of this quite funny stuff 01:18:50.560 --> 01:18:53.260 and it you know it's ai obviously but it but 01:18:53.260 --> 01:18:55.640 it looks really good and the delivery is mostly 01:18:55.640 --> 01:18:57.939 pretty good of the scripts that were that it 01:18:57.939 --> 01:19:01.020 was fed um and you have to take your hat off 01:19:01.020 --> 01:19:03.619 to it and say this is great creativity this is 01:19:03.619 --> 01:19:06.079 how you go viral this is how you you know i suspect 01:19:06.079 --> 01:19:09.239 um they've got more business than they can handle 01:19:09.239 --> 01:19:13.920 now and so to your point as a brand as these 01:19:13.920 --> 01:19:17.239 tools get better and better What creative, interesting 01:19:17.239 --> 01:19:20.899 ideas can you use them for to promote your products 01:19:20.899 --> 01:19:23.920 and brand in a way that's in line with your brand 01:19:23.920 --> 01:19:26.199 values, right? Like not everybody can do a funny 01:19:26.199 --> 01:19:31.479 gorilla video that you would never have bothered 01:19:31.479 --> 01:19:33.520 to do if it was going to take three months and 01:19:33.520 --> 01:19:37.319 cost $200 ,000. But if it takes a day and costs 01:19:37.319 --> 01:19:42.180 $10 in Sora to API credits, assuming maybe we 01:19:42.180 --> 01:19:43.939 get access to models like this, and you can access 01:19:43.939 --> 01:19:48.109 VO3. feo3 via the api um what can you do now 01:19:48.109 --> 01:19:50.670 right i think that's the question we should all 01:19:50.670 --> 01:19:53.609 be asking ourselves because there's so much text 01:19:53.609 --> 01:19:57.590 -based slop out there um and we are visual auditory 01:19:57.590 --> 01:19:59.250 creatures and we're going to engage much more 01:19:59.250 --> 01:20:01.890 with video than we are text anyway in my opinion 01:20:01.890 --> 01:20:04.430 especially on on social if you're looking to 01:20:04.430 --> 01:20:06.970 get attention so one of the things i really hope 01:20:06.970 --> 01:20:09.430 that we're able to do in our team is to spend 01:20:09.430 --> 01:20:11.649 some time thinking of what those creative ideas 01:20:11.649 --> 01:20:15.590 are because that's the hard part I think prompting 01:20:15.590 --> 01:20:17.109 the tools to get the output you want is probably 01:20:17.109 --> 01:20:18.869 pretty difficult and takes a bit of trial and 01:20:18.869 --> 01:20:20.550 error as well. But it's the creativity, it's 01:20:20.550 --> 01:20:24.630 the ideation that's the bottleneck now. People 01:20:24.630 --> 01:20:27.729 haven't played with any of these tools yet. Another 01:20:27.729 --> 01:20:31.550 one to go check out is Blow. And this is from 01:20:31.550 --> 01:20:35.829 Google. It uses their image and video generation 01:20:35.829 --> 01:20:38.649 models, but it allows you to combine together 01:20:38.649 --> 01:20:42.829 into a full... video clip using the kind of eight 01:20:42.829 --> 01:20:46.609 to 15 second videos you generate into a storyboard 01:20:46.609 --> 01:20:50.090 and combine them together it's a lot of fun to 01:20:50.090 --> 01:20:53.590 play with and i've really enjoyed that for creative 01:20:53.590 --> 01:20:56.930 concepting yeah i think i haven't played with 01:20:56.930 --> 01:20:58.470 that but i'm definitely going to go and have 01:20:58.470 --> 01:21:00.869 a go with it um another thing that's sort of 01:21:00.869 --> 01:21:02.890 tangentially video related that we've been using 01:21:02.890 --> 01:21:07.399 a fair bit in supreme group for some tests Not 01:21:07.399 --> 01:21:09.420 only client work, I work in the marketing team 01:21:09.420 --> 01:21:11.220 for Supreme Group. So all of the applications 01:21:11.220 --> 01:21:14.340 I'm talking about are within our business, not 01:21:14.340 --> 01:21:19.000 for clients. But it's the power of video avatars 01:21:19.000 --> 01:21:21.800 and just how realistic they're getting. A colleague 01:21:21.800 --> 01:21:24.079 of mine, Jake, was playing with a potential use 01:21:24.079 --> 01:21:29.199 case that cloned a video likeness of him analyzing 01:21:29.199 --> 01:21:34.680 a company's website. And it was basically a... 01:21:35.519 --> 01:21:39.399 fairly realistic video of him explaining issues 01:21:39.399 --> 01:21:44.039 with a website based on a automation built in 01:21:44.039 --> 01:21:48.239 Zapier. And we're nearly there. I think most 01:21:48.239 --> 01:21:50.079 people would still know that it wasn't the real 01:21:50.079 --> 01:21:53.279 Jake, but they're now able to better clone your 01:21:53.279 --> 01:21:55.920 mannerisms in terms of how you move and speak. 01:21:56.239 --> 01:21:58.279 They're better at delivering the scripts in a 01:21:58.279 --> 01:22:00.760 more natural way. And I think most people would 01:22:00.760 --> 01:22:03.439 see it was AI, but... We're probably only one 01:22:03.439 --> 01:22:06.899 or two more versions away, I think, before using 01:22:06.899 --> 01:22:10.479 AI avatars at scale for things like prospecting 01:22:10.479 --> 01:22:14.640 email outreach is very viable in a way that people 01:22:14.640 --> 01:22:16.579 might not be able to tell the difference between 01:22:16.579 --> 01:22:20.539 a person and the AI version of them. So that's 01:22:20.539 --> 01:22:23.279 pretty cool. Do you want an AI version of you, 01:22:23.340 --> 01:22:28.359 Martin? We've got several in various guises, 01:22:28.359 --> 01:22:32.520 mostly through HeyGen, I must admit. there's 01:22:32.520 --> 01:22:34.159 a great unlock i think from a sales perspective 01:22:34.159 --> 01:22:37.180 but it leads to this ai slop debate in my head 01:22:37.180 --> 01:22:39.880 i can't i can't get past that as cool as these 01:22:39.880 --> 01:22:42.619 things are any application that just allows you 01:22:42.619 --> 01:22:47.500 to kind of mass spread more content with very 01:22:47.500 --> 01:22:49.979 little discernment as to each individual item 01:22:49.979 --> 01:22:53.699 produced makes me kind of wince a little bit 01:22:53.699 --> 01:22:55.819 you're absolutely right we've talked about this 01:22:55.819 --> 01:22:58.039 a hundred times but that's still why in the future 01:22:58.590 --> 01:23:01.529 of AI marketers in the world that we live in, 01:23:01.609 --> 01:23:06.949 having a trusted brand, not just a brand in terms 01:23:06.949 --> 01:23:10.010 of your products and services, but your content 01:23:10.010 --> 01:23:12.529 and your knowledge and the thought leadership 01:23:12.529 --> 01:23:15.949 that you can share. Because people are not going 01:23:15.949 --> 01:23:20.250 to be able to scroll through and curate good 01:23:20.250 --> 01:23:23.010 content out of all of the crap that we're all 01:23:23.010 --> 01:23:24.890 going to be drowning in. They're going to be 01:23:24.890 --> 01:23:27.369 like, I just need three or four sources that 01:23:27.369 --> 01:23:29.960 I trust. because they just can't be bothered 01:23:29.960 --> 01:23:33.359 to go fishing through through everything else 01:23:33.359 --> 01:23:36.939 but um that we had one more section we were going 01:23:36.939 --> 01:23:38.880 to do today but it's already a pretty long podcast 01:23:38.880 --> 01:23:41.420 so i think we tease folks mine and we let them 01:23:41.420 --> 01:23:43.899 know this is what we're going to go into next 01:23:43.899 --> 01:23:47.060 week um which is we're going to talk about all 01:23:47.060 --> 01:23:49.939 of the trends affecting marketers the most today 01:23:49.939 --> 01:23:52.500 like the rise of generative search the release 01:23:52.500 --> 01:23:55.560 of chat gpt shopping the impact of vibe coding 01:23:55.560 --> 01:23:58.279 for marketers but Let's do that next week, Martin, 01:23:58.279 --> 01:24:02.340 otherwise we'll have kept people too long. Sounds 01:24:02.340 --> 01:24:06.439 great. I've got plenty to say on all of those 01:24:06.439 --> 01:24:10.819 topics. So, you know, do tune in. We're back. 01:24:11.359 --> 01:24:13.380 Well, as always, thank you for your time and 01:24:13.380 --> 01:24:14.779 your thoughts and your energy, Martin. Really 01:24:14.779 --> 01:24:16.739 appreciate it. I hope you've, dear listeners, 01:24:16.800 --> 01:24:19.560 enjoyed the return for the podcast and you've 01:24:19.560 --> 01:24:21.380 already had a little teaser of what we'll be 01:24:21.380 --> 01:24:24.920 bringing you next week. Have a great week, Martin, 01:24:24.939 --> 01:24:27.729 and I'll speak to you soon. Thank you for listening 01:24:27.729 --> 01:24:30.810 to artificially intelligent marketing to stay 01:24:30.810 --> 01:24:33.869 on top of the latest trends, tips, and tools 01:24:33.869 --> 01:24:37.430 in the world of marketing AI. Be sure to subscribe. 01:24:37.770 --> 01:24:40.489 We look forward to seeing you again next week.