WEBVTT 00:00:04.169 --> 00:00:06.830 Welcome to artificially intelligent marketing, 00:00:06.990 --> 00:00:09.949 a weekly podcast where we stay on top of the 00:00:09.949 --> 00:00:13.230 latest trends, tips, and tools in the world of 00:00:13.230 --> 00:00:16.070 marketing AI, helping you get the best results 00:00:16.070 --> 00:00:19.210 from your marketing efforts. Now let's join our 00:00:19.210 --> 00:00:23.550 hosts, Paul Avery and Martin Broadhurst. Hello, 00:00:23.609 --> 00:00:26.399 everyone. Welcome to Artificially Intelligent 00:00:26.399 --> 00:00:28.879 Marketing. Paul Avery here, as usual, joined 00:00:28.879 --> 00:00:31.260 by the fantabulous Martin Broadhurst. Martin, 00:00:31.339 --> 00:00:34.600 how are you, sir? I'm great. Unlike the weather 00:00:34.600 --> 00:00:38.979 outside, which is gloomy, miserable, very, very 00:00:38.979 --> 00:00:44.119 December. Generally speaking, I'm good. The weather 00:00:44.119 --> 00:00:48.100 is frightful, but the podcast is extremely delightful. 00:00:48.259 --> 00:00:51.460 So everyone's a winner. We've got some really 00:00:51.460 --> 00:00:54.140 interesting stuff to talk about this week. Last 00:00:54.140 --> 00:00:56.659 couple of weeks have been crazy. So what have 00:00:56.659 --> 00:00:58.759 we got on the docket today? A few new models 00:00:58.759 --> 00:01:01.979 have been released. You might have slept on this, 00:01:02.000 --> 00:01:04.900 Paul. I know you've been very busy, but Google 00:01:04.900 --> 00:01:10.200 have released Gemini 3 Pro and the exciting Nano 00:01:10.200 --> 00:01:13.859 Banana Pro. Anthropic responded with their own 00:01:13.859 --> 00:01:18.640 release of Opus 4 .5. Away from model releases, 00:01:18.719 --> 00:01:21.870 we've had... Gartner releasing their 2026 strategic 00:01:21.870 --> 00:01:26.329 predictions, which you'll be unsurprised to hear 00:01:26.329 --> 00:01:32.109 contains some AI mentions. I know I was shocked. 00:01:33.349 --> 00:01:37.609 Google have started rolling out ads in AI overviews, 00:01:37.609 --> 00:01:41.349 and there's some rumors flying around that chat 00:01:41.349 --> 00:01:43.670 GPT might be looking at doing something similar 00:01:43.670 --> 00:01:49.840 in their system. And Adobe has purchased. semrush 00:01:49.840 --> 00:01:52.180 lots going on then lots going on let's get into 00:01:52.180 --> 00:01:54.299 that first one i mean it's been it's been the 00:01:54.299 --> 00:01:56.920 model week isn't it gemini 3 came out and we 00:01:56.920 --> 00:01:58.540 started playing with it instantly and we were 00:01:58.540 --> 00:02:01.680 like oh this is decent all of the evals look 00:02:01.680 --> 00:02:04.840 very good but as we're finding mostly in our 00:02:04.840 --> 00:02:06.900 own work you have to test your own use cases 00:02:06.900 --> 00:02:09.710 you can't massively trust too many uvals at this 00:02:09.710 --> 00:02:12.389 point at least in my experience um and then yeah 00:02:12.389 --> 00:02:15.409 opus 4 .5 comes straight out and there was who 00:02:15.409 --> 00:02:18.889 is the who is winning and then gpt 5 .1 as well 00:02:18.889 --> 00:02:21.889 so lots going on take us through what google's 00:02:21.889 --> 00:02:24.750 been up to so google was sitting at the top of 00:02:24.750 --> 00:02:29.129 the leaderboards with gemini 2 .5 for a long 00:02:29.129 --> 00:02:33.250 time and then there were some rumors flying around 00:02:33.250 --> 00:02:35.990 that three would be coming out certainly the 00:02:35.990 --> 00:02:39.569 uh the x .com discussions and reddit boards were 00:02:39.569 --> 00:02:43.289 awash with speculation and then yeah they launched 00:02:43.289 --> 00:02:48.110 a new model gemini 3 pro which we think will 00:02:48.110 --> 00:02:51.509 be followed by gemini 3 flash the lightweight 00:02:51.509 --> 00:02:56.270 version the kind of not reasoning model soon 00:02:56.270 --> 00:02:59.650 uh but yeah it came out and smashed the leaderboards 00:02:59.650 --> 00:03:04.990 it was a real leap as well i think sometimes 00:03:04.990 --> 00:03:10.159 there's been so little between the the state 00:03:10.159 --> 00:03:12.780 -of -the -art models the frontier models it's 00:03:12.780 --> 00:03:16.819 been a few points here or there that when this 00:03:16.819 --> 00:03:19.280 came out on some of the the leaderboards it was 00:03:19.280 --> 00:03:24.319 a significant step forward um on leaderboards 00:03:24.319 --> 00:03:28.280 that were quite saturated or on ones that are 00:03:28.280 --> 00:03:31.379 have a long way to go they were really pushing 00:03:31.379 --> 00:03:34.469 things forward lm arena they were smashing the 00:03:34.469 --> 00:03:37.069 competition so yeah it's certainly been well 00:03:37.069 --> 00:03:41.210 received my initial play with the models has 00:03:41.210 --> 00:03:45.110 been has been very good both in the in the api 00:03:45.110 --> 00:03:51.610 and the uh gemini chat interface i've really 00:03:51.610 --> 00:03:55.590 enjoyed putting some challenging scenarios in 00:03:55.590 --> 00:03:59.530 it there's been things like uh trying to design 00:03:59.530 --> 00:04:03.800 a complex power automate flow which is the kind 00:04:03.800 --> 00:04:06.159 of workflow that their models generally aren't 00:04:06.159 --> 00:04:08.199 very good at working with the only time i've 00:04:08.199 --> 00:04:12.400 had any decent responses is from chat gpt 5 pro 00:04:12.400 --> 00:04:17.639 um gemini 2 .5 wasn't really very good at understanding 00:04:17.639 --> 00:04:20.500 some of the the mechanics of power automate whereas 00:04:20.500 --> 00:04:25.459 i found with gemini 3 pro it did an excellent 00:04:25.459 --> 00:04:28.579 job of getting into the detail of how to automate 00:04:28.579 --> 00:04:30.959 a particular workflow Yeah, I mean, if we take 00:04:30.959 --> 00:04:33.939 a look at some of the evals, the Humanity's last 00:04:33.939 --> 00:04:37.800 exam, Gemini 3 Pro with code execution got 45 00:04:37.800 --> 00:04:41.759 .8 % and the previous high was 26 .5, which was 00:04:41.759 --> 00:04:45.100 GPT 5 .1. So that's a pretty big leap. Another 00:04:45.100 --> 00:04:48.519 one that jumped out to me was an eval called 00:04:48.519 --> 00:04:51.579 ScreenSpot Pro, which is basically AI's ability 00:04:51.579 --> 00:04:55.439 to understand a computer screen. The previous 00:04:55.439 --> 00:04:58.779 high on that was Claude Sonnet at 36%. Gemini 00:04:58.779 --> 00:05:02.339 3 does 72 % on that eval. That's important, right, 00:05:02.420 --> 00:05:05.819 for computer use by agents. So that's kind of 00:05:05.819 --> 00:05:10.959 interesting. The Vending Bench 2 eval, it has 00:05:10.959 --> 00:05:14.100 the highest score so far. So for listeners that 00:05:14.100 --> 00:05:16.860 are not familiar with this, basically the AI 00:05:16.860 --> 00:05:20.439 is... briefed to run a vending machine business 00:05:20.439 --> 00:05:23.560 and it has to like buy supplies, control pricing, 00:05:23.819 --> 00:05:26.439 deal with customer complaints and stuff like 00:05:26.439 --> 00:05:29.399 this and basically try and make money. So it 00:05:29.399 --> 00:05:32.199 nearly, probably not doubled, but maybe 50 % 00:05:32.199 --> 00:05:35.740 greater than the previous benchmark score. And 00:05:35.740 --> 00:05:37.220 that one's really interesting because it's much 00:05:37.220 --> 00:05:39.899 more closer to the world of real work, right, 00:05:39.980 --> 00:05:43.139 rather than some of the other evals which maybe 00:05:43.139 --> 00:05:47.019 get a bit esoteric. The criticism of all the 00:05:47.019 --> 00:05:49.660 evals is that once you know how the evals work, 00:05:49.800 --> 00:05:52.839 you can overfit the model in post -training. 00:05:52.980 --> 00:05:54.560 So again, you have to take a lot of this stuff 00:05:54.560 --> 00:05:56.699 with a pinch of salt perhaps, but it is interesting 00:05:56.699 --> 00:05:58.920 to see them massively increase. And I was listening 00:05:58.920 --> 00:06:02.220 to the Moonshots podcast with Peter Diamandis 00:06:02.220 --> 00:06:05.199 -Martin and the team on the podcast were talking 00:06:05.199 --> 00:06:08.139 about how we are drifting to a world where the 00:06:08.139 --> 00:06:10.620 evals that are going to be the most meaningful 00:06:10.620 --> 00:06:15.410 are really business outcome oriented. measured 00:06:15.410 --> 00:06:19.589 more in dollars like vending bench than in particular 00:06:19.589 --> 00:06:23.209 scores or percentages like completion percentages 00:06:23.209 --> 00:06:27.430 so definitely better my own personal experience 00:06:27.430 --> 00:06:30.470 has been it seems more thoughtful and knowledgeable 00:06:30.470 --> 00:06:34.490 to interact with it's much better at self -critique 00:06:34.490 --> 00:06:37.949 its writing quality is a slight improvement although 00:06:37.949 --> 00:06:42.069 i'm still not absolutely convinced that the models 00:06:42.069 --> 00:06:45.079 are getting much better at writing the We'll 00:06:45.079 --> 00:06:47.939 talk about Opus in a minute, but that's probably 00:06:47.939 --> 00:06:50.220 the best one I've seen so far. And then GPT -5 00:06:50.220 --> 00:06:53.100 .1 is better than GPT -5. So there's a bunch 00:06:53.100 --> 00:06:56.199 of nuances in there. But is there anything you've 00:06:56.199 --> 00:06:57.800 been doing with Gemini Feed particularly that 00:06:57.800 --> 00:07:00.240 you're getting good results on? Yeah, there was 00:07:00.240 --> 00:07:03.399 one particular project that I'd been struggling 00:07:03.399 --> 00:07:07.660 with, needed a front end that tapped into a few 00:07:07.660 --> 00:07:12.439 government APIs for property data. things like 00:07:12.439 --> 00:07:17.980 planning information, zoning, things like that. 00:07:18.420 --> 00:07:23.779 And I tried a few different tools. They took 00:07:23.779 --> 00:07:27.660 me 90 % of the way there, but there were some 00:07:27.660 --> 00:07:30.399 nuances in the way that things would render on 00:07:30.399 --> 00:07:34.259 the map. But Google Gemini one -shotted it. It 00:07:34.259 --> 00:07:36.220 was really powerful. It used the same prompt 00:07:36.220 --> 00:07:38.970 that I'd used elsewhere. and quite literally 00:07:38.970 --> 00:07:41.189 it one -shotted the output so yeah i was very 00:07:41.189 --> 00:07:44.930 impressed to see to see that improvement in the 00:07:44.930 --> 00:07:47.449 model been very good at spinning up front ends 00:07:47.449 --> 00:07:50.589 isn't it we've talked previously about for marketers 00:07:50.589 --> 00:07:54.009 moving beyond content as lead magnets and thinking 00:07:54.009 --> 00:07:56.970 about tools that you might build for your for 00:07:56.970 --> 00:07:58.750 your target audience that you would not have 00:07:58.750 --> 00:08:00.910 spent money on a developer on but if an ai can 00:08:00.910 --> 00:08:04.689 spin it up so as these things mature and improve 00:08:04.689 --> 00:08:06.889 and Gemini 3 is definitely one to try if you've 00:08:06.889 --> 00:08:08.970 been thinking about building tools and the other 00:08:08.970 --> 00:08:11.310 AIs have not been quite good enough or you haven't 00:08:11.310 --> 00:08:12.910 had confidence because you thought you'd have 00:08:12.910 --> 00:08:15.050 to get in and like wrangle the code yourself 00:08:15.050 --> 00:08:17.889 it's definitely worth playing with Gemini 3 from 00:08:17.889 --> 00:08:20.230 that perspective the other thing I've been testing 00:08:20.230 --> 00:08:22.149 this weekend is I've been running quite a lot 00:08:22.149 --> 00:08:26.430 of deep research reports um and it's been interesting 00:08:26.430 --> 00:08:28.709 in the past I've kind of got similar things like 00:08:28.709 --> 00:08:31.490 if I've run against like GPT 4 .0 against Gemini 00:08:31.490 --> 00:08:34.870 2 .5 I'd get a report that was by and large talking 00:08:34.870 --> 00:08:38.049 about the same stuff. But they started to diverge 00:08:38.049 --> 00:08:42.389 a bit, which is interesting. I ran Gemini 3 Pro 00:08:42.389 --> 00:08:45.830 against GPT 5 .1 Deep Thinking this weekend. 00:08:46.309 --> 00:08:50.809 And somewhat surprisingly to me, 5 .1 was better 00:08:50.809 --> 00:08:54.450 than Gemini Pro in terms of clarity of expression, 00:08:54.769 --> 00:08:57.909 depth of ideas, number of sources cited. Like 00:08:57.909 --> 00:09:00.230 it really did seem to do a much better job. So 00:09:00.230 --> 00:09:03.129 yeah, it's... I really thought when Gemini 3 00:09:03.129 --> 00:09:05.610 came out, it would just become my obvious go 00:09:05.610 --> 00:09:09.009 -to. But I'm not so sure that's going to turn 00:09:09.009 --> 00:09:11.549 out to be the case, if I'm honest. I think I'm 00:09:11.549 --> 00:09:13.669 still going to use GPT 5 .1 and other models 00:09:13.669 --> 00:09:17.190 for stuff. That chimes with my general experience 00:09:17.190 --> 00:09:22.730 with the Gemini chat experience and chat GPT. 00:09:23.629 --> 00:09:26.649 I still find, despite model improvements across 00:09:26.649 --> 00:09:29.980 the board, my daily driver... The app that is 00:09:29.980 --> 00:09:34.019 open on my desktop at all times is ChatGPT. They 00:09:34.019 --> 00:09:38.460 have just got the user experience down in a way 00:09:38.460 --> 00:09:43.159 that Google Gemini just doesn't quite have refined 00:09:43.159 --> 00:09:47.860 yet. But interestingly, Logan Kirkpatrick from 00:09:47.860 --> 00:09:51.559 the team has said that they are doing a big 2 00:09:51.559 --> 00:09:55.960 .0 relaunch of the UX coming soon. Yeah, it'll 00:09:55.960 --> 00:09:58.379 be interesting to see. what that entails i mean 00:09:58.379 --> 00:10:00.860 one of the cool things about gemini is they're 00:10:00.860 --> 00:10:05.799 starting to play with spinning up like dashboards 00:10:05.799 --> 00:10:08.919 and interactive experiences as part of the answer 00:10:08.919 --> 00:10:11.779 we've talked previously and i think it's pretty 00:10:11.779 --> 00:10:14.000 well discussed in the industry at this point 00:10:14.000 --> 00:10:18.450 that ux on demand is probably going to be how 00:10:18.450 --> 00:10:20.149 AIs work. Like sometimes it will give you an 00:10:20.149 --> 00:10:21.690 answer in text, if that's the most appropriate, 00:10:21.789 --> 00:10:23.389 but maybe it will create images. Interesting 00:10:23.389 --> 00:10:25.450 thing is Gemini 3 is much better. It goes out 00:10:25.450 --> 00:10:27.769 on the web and it tries to find images that will 00:10:27.769 --> 00:10:30.389 support its point and it intersperses them in 00:10:30.389 --> 00:10:33.009 the answer. And for me, on a couple of occasions, 00:10:33.110 --> 00:10:36.809 it even created images, which was super interesting. 00:10:36.929 --> 00:10:38.610 And they actually worked. And I'll give you an 00:10:38.610 --> 00:10:40.789 example. I think the advice turned out to suck, 00:10:40.889 --> 00:10:44.570 but I was... it's time for us to buy a new mattress. 00:10:44.750 --> 00:10:46.830 So I've been using Gemini 3 to help me find this 00:10:46.830 --> 00:10:48.950 new mattress. So I bought a new mattress, but 00:10:48.950 --> 00:10:51.110 the delivery time is six weeks and I'm getting 00:10:51.110 --> 00:10:53.409 a bad back from the current mattress. So Gemini 00:10:53.409 --> 00:10:55.830 3 gave me a bunch of recommendations about what 00:10:55.830 --> 00:10:58.610 I should do. One of them was to put a yoga mat 00:10:58.610 --> 00:11:01.049 under the dip in the mattress, which I think 00:11:01.049 --> 00:11:04.490 is actually terrible advice, but there's a number 00:11:04.490 --> 00:11:06.149 of different ways you could orientate that. And 00:11:06.149 --> 00:11:07.649 it wasn't quite clear to me. So I was like, can 00:11:07.649 --> 00:11:10.889 you draw an image that demonstrates this? came 00:11:10.889 --> 00:11:13.269 up with a beautiful image that absolutely um 00:11:13.269 --> 00:11:15.549 made it clear what its advice was and that's 00:11:15.549 --> 00:11:18.610 a new capability as far as i'm aware but this 00:11:18.610 --> 00:11:20.409 probably ties into probably potentially even 00:11:20.409 --> 00:11:23.009 more exciting part of gemini release which is 00:11:23.009 --> 00:11:26.950 the improvements to nano banana image gen model 00:11:26.950 --> 00:11:29.950 have you um had a chance to play with that i 00:11:29.950 --> 00:11:35.149 have so this is quite a step on from nano banana 00:11:35.149 --> 00:11:38.470 this is again it's a anyone that isn't aware 00:11:38.470 --> 00:11:41.409 of the differences google has imagine or imaging 00:11:41.409 --> 00:11:47.470 which is a diffusion based image generation model 00:11:47.470 --> 00:11:51.889 the clarity of those images is very good but 00:11:51.889 --> 00:11:56.350 it isn't baked into isn't native within the large 00:11:56.350 --> 00:11:59.669 language model gemini itself and that meant that 00:11:59.669 --> 00:12:03.529 getting edits of a design is quite tricky so 00:12:03.529 --> 00:12:05.990 every time you created an image with it it was 00:12:05.990 --> 00:12:10.419 creating a brand new image whereas with nano 00:12:10.419 --> 00:12:14.259 banana when they launched that within 2 .5 in 00:12:14.259 --> 00:12:16.779 the previous release suddenly we could upload 00:12:16.779 --> 00:12:20.639 images and get very detailed very accurate edits 00:12:20.639 --> 00:12:24.460 created to images well now they're talking about 00:12:24.460 --> 00:12:31.879 nano banana pro being a foundational image model 00:12:31.879 --> 00:12:36.000 which has almost a world model baked inside it 00:12:36.000 --> 00:12:40.080 so it understands so much more context about 00:12:40.080 --> 00:12:43.059 the image so you can ask it to do tasks not just 00:12:43.059 --> 00:12:47.419 like you know add a hat to this dog or something 00:12:47.419 --> 00:12:49.659 silly like that but you can actually ask it to 00:12:49.659 --> 00:12:54.600 isolate elements of the image or to mask certain 00:12:54.600 --> 00:12:58.080 parts of the image to turn the image into a line 00:12:58.080 --> 00:13:00.580 drawing you can ask it to you can even use it 00:13:00.580 --> 00:13:05.600 within video where you can put an image of say 00:13:05.600 --> 00:13:11.350 a a mouse stuck in a in a maze and ask it to 00:13:11.350 --> 00:13:14.970 animate the mouse running through the maze and 00:13:14.970 --> 00:13:18.809 it will accurately one shot it and this when 00:13:18.809 --> 00:13:22.389 it's linked with or nano banana pro when linked 00:13:22.389 --> 00:13:25.429 with something like vo3 opens up a whole world 00:13:25.429 --> 00:13:31.809 of amazing possibilities with visual work it's 00:13:31.809 --> 00:13:34.690 not just creating things but explaining solving 00:13:36.399 --> 00:13:39.360 um editing reasoning that there's so many possibilities 00:13:39.360 --> 00:13:41.399 that are unlocked with it i think the system 00:13:41.399 --> 00:13:43.419 card for this is quite illustrative and it's 00:13:43.419 --> 00:13:46.259 got some bits in it that i feels like a tipping 00:13:46.259 --> 00:13:48.860 point so the first thing you can now provide 00:13:48.860 --> 00:13:52.320 i think it's up to five characters as prompt 00:13:52.320 --> 00:13:54.539 images so i could put you know pictures of you 00:13:54.539 --> 00:13:57.769 and me and a couple other agency gang from the 00:13:57.769 --> 00:14:00.889 HubSpot inbound days and asked Nanobananas to 00:14:00.889 --> 00:14:03.149 start creating a bunch of unlikely scenarios 00:14:03.149 --> 00:14:05.129 for us to be in. And rather than maintaining 00:14:05.129 --> 00:14:07.169 character consistency for just one person, you 00:14:07.169 --> 00:14:08.669 could do it for more, which is really interesting. 00:14:09.309 --> 00:14:12.549 It can also do the same with up to 14 objects, 00:14:12.669 --> 00:14:14.769 which again seems a very bizarre number. Like 00:14:14.769 --> 00:14:18.309 15, oh, it's going to break. 13, safe. 14, on 00:14:18.309 --> 00:14:22.190 the edge. But again, if I wanted to control that 00:14:22.190 --> 00:14:24.450 scene even more, I could actually... add a load 00:14:24.450 --> 00:14:26.309 of objects to the scene and then it would put 00:14:26.309 --> 00:14:27.990 them all in the right places, which is pretty 00:14:27.990 --> 00:14:31.389 astounding. And I think what we may see, I suspect 00:14:31.389 --> 00:14:34.450 this won't be the model, but the next one will 00:14:34.450 --> 00:14:39.470 be, is a genuine image creation capability where 00:14:39.470 --> 00:14:40.990 if you're a marketer and you need to do a bunch 00:14:40.990 --> 00:14:43.830 of product images in different settings, or you're 00:14:43.830 --> 00:14:46.429 even trying to come up with like brief videos 00:14:46.429 --> 00:14:48.529 for those, because obviously you can use the 00:14:48.529 --> 00:14:50.529 image generation capabilities and then you can 00:14:50.529 --> 00:14:53.620 ask Veo to... turn those into video for you, 00:14:53.720 --> 00:14:57.440 is we can, as marketers, actually start to now 00:14:57.440 --> 00:15:00.440 create product imagery, right? Give it a product 00:15:00.440 --> 00:15:02.980 shot and then use it as you see fit. Put it on 00:15:02.980 --> 00:15:06.139 the moon, give it wings flying in space, like, 00:15:06.220 --> 00:15:08.899 you know, whatever you can imagine, really. And 00:15:08.899 --> 00:15:11.000 up until now, that's been a bit hit and miss. 00:15:11.159 --> 00:15:13.179 But this model feels like the one that's more 00:15:13.179 --> 00:15:16.500 consistently capable. So are we drifting, therefore, 00:15:16.600 --> 00:15:20.710 to a... post canva adobe world probably not for 00:15:20.710 --> 00:15:22.750 specialists people who are really good at creating 00:15:22.750 --> 00:15:24.330 images are probably still going to create things 00:15:24.330 --> 00:15:26.909 much better than you can just do in nano banana 00:15:26.909 --> 00:15:29.370 but yeah if you're just creating quick product 00:15:29.370 --> 00:15:31.789 imagery i wonder you know and you could probably 00:15:31.789 --> 00:15:35.149 semi -automate this with mcp or so there's probably 00:15:35.149 --> 00:15:37.429 some quite interesting things you could do that 00:15:37.429 --> 00:15:38.950 you just couldn't do with image models until 00:15:38.950 --> 00:15:41.429 now there's definitely been a shift in people's 00:15:41.429 --> 00:15:45.850 willingness to accept and to use publicly ai 00:15:45.850 --> 00:15:50.379 generated images I'm seeing it in posters, in 00:15:50.379 --> 00:15:53.840 shop windows. I'm seeing it for flyers, for events. 00:15:54.240 --> 00:15:57.019 I was on an e -commerce website this weekend 00:15:57.019 --> 00:16:00.179 and there was just something about the text that 00:16:00.179 --> 00:16:04.539 made it very clear to me it was a GPT -4 or GPT 00:16:04.539 --> 00:16:06.980 -5 generated image. It had that kind of softness 00:16:06.980 --> 00:16:10.440 to it that was very recognizable. So we're certainly 00:16:10.440 --> 00:16:13.480 seeing this more and more in the public, even 00:16:13.480 --> 00:16:15.899 though in some domains... People are still very 00:16:15.899 --> 00:16:19.039 anti -AI. I don't know if you saw this recently, 00:16:19.100 --> 00:16:24.779 but there's some pushback against Steam in the 00:16:24.779 --> 00:16:30.559 video game industry, Steam tag games if they 00:16:30.559 --> 00:16:34.240 use AI -generated content. And some developers 00:16:34.240 --> 00:16:36.799 are pushing back against that tag and that label. 00:16:37.059 --> 00:16:38.940 Yeah, I think the big issue there, I read that 00:16:38.940 --> 00:16:40.519 story, it was an interesting one. I think the 00:16:40.519 --> 00:16:44.190 big issue is... Everybody in the games development 00:16:44.190 --> 00:16:47.509 industry is probably using Gen AI in one way 00:16:47.509 --> 00:16:50.750 or another. So is it even a meaningful tag? I 00:16:50.750 --> 00:16:52.649 think that's probably quite a big issue. I thought 00:16:52.649 --> 00:16:55.490 you were going to say the Coca -Cola TV ad, which 00:16:55.490 --> 00:16:59.330 has been AI generated and has triggered a bit 00:16:59.330 --> 00:17:01.070 of a backlash. It's better than last year's ad. 00:17:01.149 --> 00:17:02.830 It still looks a bit weird in place as I saw 00:17:02.830 --> 00:17:05.630 it on TV over the weekend. I saw somebody do 00:17:05.630 --> 00:17:07.769 a breakdown of the different lorries that are 00:17:07.769 --> 00:17:11.170 in different trucks and they... took the bodies 00:17:11.170 --> 00:17:13.289 of them and turned them into a diagram. And apparently 00:17:13.289 --> 00:17:16.990 there's 14 different truck bodies used in that. 00:17:17.230 --> 00:17:20.089 Yeah, I looked at that and I thought, old tech. 00:17:20.390 --> 00:17:22.009 Yeah, honestly, I don't know how long they've 00:17:22.009 --> 00:17:24.049 been working on that as much as they've got AI. 00:17:24.230 --> 00:17:26.069 I think I read somewhere that there was still 00:17:26.069 --> 00:17:28.269 like ridiculously large amounts of people who 00:17:28.269 --> 00:17:31.049 worked on that project, like drifting towards 00:17:31.049 --> 00:17:33.450 100, which for me, I'm like, why are you using 00:17:33.450 --> 00:17:38.690 Gen AI then? But the reality is, I think I could 00:17:38.690 --> 00:17:42.460 spin up something better. in an hour with nano 00:17:42.460 --> 00:17:47.680 banana 2 and vaio 3 .1 um so that just makes 00:17:47.680 --> 00:17:49.440 me think it's old tech and next year's will probably 00:17:49.440 --> 00:17:51.819 be quite a bit better than that there may even 00:17:51.819 --> 00:17:55.279 be like rules on what models maybe they have 00:17:55.279 --> 00:17:57.980 to use open source models you know coca -cola 00:17:57.980 --> 00:18:00.579 is a big corporate and what have you um so yeah 00:18:00.579 --> 00:18:03.359 i'm not i'm not so sure but um yeah i think you're 00:18:03.359 --> 00:18:05.599 right i think attitudes are changing i mean my 00:18:05.599 --> 00:18:08.339 wife was showing me stuff on instagram this weekend 00:18:08.880 --> 00:18:10.819 And we were trying to guess how many of the videos 00:18:10.819 --> 00:18:15.660 were AI generated. And I reckon four out of five 00:18:15.660 --> 00:18:18.640 that we were looking at were AI generated. It 00:18:18.640 --> 00:18:20.579 was cats doing stuff that there's no way you 00:18:20.579 --> 00:18:24.640 get a cat to do. You know you've got a technology 00:18:24.640 --> 00:18:28.200 acceleration when one of the first use cases 00:18:28.200 --> 00:18:29.819 is always a cat. I mean, that's how the internet 00:18:29.819 --> 00:18:31.619 began. Look at these funny pictures of cats. 00:18:31.740 --> 00:18:33.680 Now we've got funny pictures of cats, but they're 00:18:33.680 --> 00:18:37.259 AI generated. So that in itself is a bit worrying 00:18:37.259 --> 00:18:40.450 because just... so much of a deluge of absolute 00:18:40.450 --> 00:18:42.930 crap but like some of them are kind of funny 00:18:42.930 --> 00:18:47.190 i saw one where the cats are um doing like work 00:18:47.190 --> 00:18:49.890 around the house like trying to fix like electric 00:18:49.890 --> 00:18:53.410 switches and fix cars and stuff like that it 00:18:53.410 --> 00:18:56.490 was kind of funny but um yeah i think it's the 00:18:56.490 --> 00:18:58.750 explosion we've been talking about is well underway 00:18:58.750 --> 00:19:02.490 let's shift gears let's talk claude because gemini 00:19:02.490 --> 00:19:06.200 3 was like basking in its glory for all of about 00:19:06.200 --> 00:19:08.960 a week, probably less, because Claude had been 00:19:08.960 --> 00:19:11.740 cooking Opus 4 .5. So quick reminder for the 00:19:11.740 --> 00:19:14.299 audience, because none of these AI companies 00:19:14.299 --> 00:19:16.660 make it easy to track what these different things 00:19:16.660 --> 00:19:20.579 mean. So Anthropix Claude is obviously not as 00:19:20.579 --> 00:19:23.059 popular a model as ChatGPT and Gemini, but still 00:19:23.059 --> 00:19:25.859 popular, especially among coders. And they have 00:19:25.859 --> 00:19:29.720 three flavors. They've got Haiku, Sonnet, and 00:19:29.720 --> 00:19:34.000 Opus, which approximately... is also the scale 00:19:34.000 --> 00:19:36.279 and size and capability of the model so haiku 00:19:36.279 --> 00:19:39.220 is the least capable very cheap to run sonnet's 00:19:39.220 --> 00:19:41.740 in the middle opus is a big gargantuan model 00:19:41.740 --> 00:19:44.019 that's very expensive to run and the thing that 00:19:44.019 --> 00:19:47.140 was super interesting this week is opus 4 .5 00:19:47.140 --> 00:19:50.099 came out sonnet 4 .5 is already out so the much 00:19:50.099 --> 00:19:53.279 it's the big brother version and it's actually 00:19:53.279 --> 00:19:56.380 still quite cheap to run also compared to like 00:19:56.380 --> 00:19:58.400 in general the opus models are very very much 00:19:58.400 --> 00:20:00.259 more expensive to run than the sonnet models 00:20:00.859 --> 00:20:04.160 But this one, not so much. And some people have 00:20:04.160 --> 00:20:06.339 been loving it online. What have you seen as 00:20:06.339 --> 00:20:09.140 it relates to Opus 4 .5, Martin? Yeah, so it 00:20:09.140 --> 00:20:12.880 came out swinging on the benchmarks again on 00:20:12.880 --> 00:20:16.200 a Gentic coding. So the SWE benchmark, it scored 00:20:16.200 --> 00:20:21.940 80 .9 versus Gemini Pro's 76 .2. On a Gentic 00:20:21.940 --> 00:20:25.940 terminal coding, it scored 59 .3 versus Gemini 00:20:25.940 --> 00:20:32.940 Pro's 54 .2. They had graduate level reasoning. 00:20:33.059 --> 00:20:35.900 This was an interesting one. So the GPQA diamond 00:20:35.900 --> 00:20:42.640 test, which is PhD level research questions, 00:20:42.980 --> 00:20:49.240 scored 87 versus Gemini Pro's 91 .9. So not quite 00:20:49.240 --> 00:20:52.599 state of the art there. But generally speaking, 00:20:52.740 --> 00:20:55.400 we are talking about a very, very capable model 00:20:55.400 --> 00:21:01.470 across the board. My personal experience with 00:21:01.470 --> 00:21:04.109 it is I'm just pleased that they've increased 00:21:04.109 --> 00:21:07.869 the limits. One of the big criticisms with Anthropic's 00:21:07.869 --> 00:21:12.009 policy for a subscriber, whether you are on pro 00:21:12.009 --> 00:21:15.230 or the plus plan or whatever it is for $20 a 00:21:15.230 --> 00:21:17.789 month, or you are on one of their max plans for 00:21:17.789 --> 00:21:21.769 $100 or $200 per month, the limits on Opus were 00:21:21.769 --> 00:21:24.109 laughably small. It had become a bit of a meme 00:21:24.109 --> 00:21:26.210 that you would put in one question and that would 00:21:26.210 --> 00:21:29.339 be you done for a month. But now they've opened 00:21:29.339 --> 00:21:33.259 up to max plan users. They've given you the same 00:21:33.259 --> 00:21:37.400 limits that you had for Sonnet 4 .5. And this 00:21:37.400 --> 00:21:40.259 model is really, really smart and really good. 00:21:40.900 --> 00:21:43.259 So I've been playing around with it for some 00:21:43.259 --> 00:21:47.900 strategic development and it adds a bit more 00:21:47.900 --> 00:21:53.279 finesse to Sonnet. For coding, I've only lightly 00:21:53.279 --> 00:21:56.099 tested it at this moment using Claude Code for 00:21:56.099 --> 00:22:00.069 a couple of projects. but looks really good really 00:22:00.069 --> 00:22:02.789 very capable yeah i think it's quite cool i think 00:22:02.789 --> 00:22:05.789 the thing that struck me here and it's there's 00:22:05.789 --> 00:22:07.769 a little bit of trying to infer what the major 00:22:07.769 --> 00:22:11.130 labs are doing my understanding stop me if i'm 00:22:11.130 --> 00:22:14.950 wrong mine is when we see these like number increments 00:22:14.950 --> 00:22:20.410 gpt5 becomes gpt 5 .1 is if we go from you know 00:22:20.410 --> 00:22:23.150 gpt5 to gpt6 it's because there's been another 00:22:23.150 --> 00:22:26.509 training run And a fundamental new model has 00:22:26.509 --> 00:22:28.670 been created. And then it goes through all the 00:22:28.670 --> 00:22:32.130 fine -tuning, reinforcement -based fine -tuning 00:22:32.130 --> 00:22:33.609 and all that other stuff. And then we have our 00:22:33.609 --> 00:22:35.450 new model. Whereas if it goes up in incrementation, 00:22:35.849 --> 00:22:38.450 5 to 5 .1, it's because they've maybe further 00:22:38.450 --> 00:22:41.049 fine -tuned the model or made some changes in 00:22:41.049 --> 00:22:43.829 the back end that change its personality or how 00:22:43.829 --> 00:22:47.750 it basically thinks and produces outputs. So 00:22:47.750 --> 00:22:52.089 you look at Gemini 3, that one assumes is a new 00:22:52.089 --> 00:22:54.940 training run, a new model. that was trained in 00:22:54.940 --> 00:22:56.759 a different way, more data, whatever, and is 00:22:56.759 --> 00:22:59.319 therefore significantly better than Gemini 2, 00:22:59.559 --> 00:23:04.819 2 .5, whatever. Opus 4 .5 is not a new model. 00:23:05.000 --> 00:23:08.539 And yet it has these capabilities that outstrip 00:23:08.539 --> 00:23:12.339 Gemini 3 in a number of areas, which for me is 00:23:12.339 --> 00:23:15.079 interesting because it suggests that the innovation 00:23:15.079 --> 00:23:17.980 in the fine -tuning and the reinforcement -based 00:23:17.980 --> 00:23:20.380 learning and whatever they're doing in post -training 00:23:20.380 --> 00:23:23.099 to improve model outputs can actually have quite 00:23:23.099 --> 00:23:28.039 a big impact on what the models can do. So I 00:23:28.039 --> 00:23:30.960 find that quite interesting because in the beginning, 00:23:31.019 --> 00:23:32.480 we were like, need a new model release, need 00:23:32.480 --> 00:23:34.240 a new model release. But then obviously we got 00:23:34.240 --> 00:23:38.400 the own models from ChatGPT, from OpenAI, which 00:23:38.400 --> 00:23:40.579 could think they took some time before they gave 00:23:40.579 --> 00:23:43.380 an answer. And that gave us a big leap. And there's 00:23:43.380 --> 00:23:45.420 something, I think there's something special 00:23:45.420 --> 00:23:48.859 in here. They've done something with Opus's ability 00:23:48.859 --> 00:23:52.779 to think and produce good outputs that's not 00:23:52.779 --> 00:23:55.539 based on just throwing the kitchen sink of data 00:23:55.539 --> 00:23:59.240 at it. It's something else. Yeah, and that really 00:23:59.240 --> 00:24:02.019 comes through in the token usage. So I don't 00:24:02.019 --> 00:24:05.200 know if you saw this from the report. They gave 00:24:05.200 --> 00:24:09.880 it some puzzles, a series of challenges, an escape 00:24:09.880 --> 00:24:12.869 room challenge. And they... gave that challenge 00:24:12.869 --> 00:24:18.109 to sonnet 4 .5 as well as opus 4 .5 and they 00:24:18.109 --> 00:24:21.349 gave them the ability to use advanced tools so 00:24:21.349 --> 00:24:24.490 opus 4 .5 is exceptionally good at function calling 00:24:24.490 --> 00:24:27.690 or being able to use tools like web search or 00:24:27.690 --> 00:24:31.450 code execution things like that and it was able 00:24:31.450 --> 00:24:36.410 to solve the problems using far fewer tokens 00:24:36.410 --> 00:24:38.829 i think this is really interesting for developers 00:24:38.829 --> 00:24:40.529 or when we start to think about what this means 00:24:41.160 --> 00:24:46.000 in real world applications within business. If 00:24:46.000 --> 00:24:49.180 you've got long time horizon tasks that people 00:24:49.180 --> 00:24:54.579 need to execute, you don't want it burning through 00:24:54.579 --> 00:24:57.099 huge amounts of tokens costing you a great deal 00:24:57.099 --> 00:25:01.279 of expense. Whereas this new model is doing exactly 00:25:01.279 --> 00:25:04.200 what you'd want. It's optimizing. So the amount 00:25:04.200 --> 00:25:09.079 of tokens that it was using was substantially 00:25:09.079 --> 00:25:11.839 less so just to bring this to life software engineering 00:25:11.839 --> 00:25:15.299 with effort controls opus 4 .5 was achieving 00:25:15.299 --> 00:25:21.960 around 81 accuracy with around 12 000 tokens 00:25:21.960 --> 00:25:25.420 being used whereas sonnet 4 .5 was achieving 00:25:25.420 --> 00:25:31.220 76 77 accuracy using 22 000 tokens so it was 00:25:31.220 --> 00:25:33.480 using double the tokens and it was less accurate 00:25:33.480 --> 00:25:35.940 yeah i think it's i think it's cool i think this 00:25:35.940 --> 00:25:39.160 is What makes Opus 4 .5 interesting? It's in 00:25:39.160 --> 00:25:41.660 the nuances. There's another thing in the system 00:25:41.660 --> 00:25:44.940 card. One of the benchmarks involves being a 00:25:44.940 --> 00:25:48.700 customer service agent bot, like chatbot, that 00:25:48.700 --> 00:25:52.440 solves a customer's problem. And in one of the 00:25:52.440 --> 00:25:56.579 cases, Opus 4 .5, it failed the test because 00:25:56.579 --> 00:26:00.119 it didn't give the outcome expected. But what 00:26:00.119 --> 00:26:02.559 it actually did is found a really creative way 00:26:02.559 --> 00:26:04.940 to deliver on the customer's requirements based 00:26:04.940 --> 00:26:07.990 on the sort of knowledge base of information 00:26:07.990 --> 00:26:10.869 it worked out how to give the customer what the 00:26:10.869 --> 00:26:14.450 customer wanted by effectively exploiting a loophole 00:26:14.450 --> 00:26:17.609 in how the particular service in this case it's 00:26:17.609 --> 00:26:21.230 um changing a flight booking um so it failed 00:26:21.230 --> 00:26:23.170 the test but because it was too creative and 00:26:23.170 --> 00:26:26.049 it did something that was not specced that maybe 00:26:26.049 --> 00:26:29.890 a human wouldn't have even known how to do and 00:26:29.890 --> 00:26:31.630 that again that's the first time this type of 00:26:31.630 --> 00:26:33.690 things come up and then there was one other thing 00:26:33.690 --> 00:26:38.940 where Claude administers a sort of homework test 00:26:38.940 --> 00:26:42.839 to new developers that apply to work at Anthropic. 00:26:43.140 --> 00:26:47.400 And for the first time, Opus 4 .5 gets a higher 00:26:47.400 --> 00:26:50.140 score on average than the humans who take that 00:26:50.140 --> 00:26:52.940 test. And so if you're watching for things like 00:26:52.940 --> 00:26:56.900 recursive improvement in the capability of agents, 00:26:57.160 --> 00:26:59.980 well, here's a pretty strong indicator that Opus 00:26:59.980 --> 00:27:03.759 4 .5 is probably at that capability, where it's 00:27:04.109 --> 00:27:06.970 able to pass tests that a high -end developer 00:27:06.970 --> 00:27:09.210 would probably be administered. So what else 00:27:09.210 --> 00:27:11.309 can it do? And then the final thing I really 00:27:11.309 --> 00:27:14.170 liked is there's an implication that Opus is 00:27:14.170 --> 00:27:17.690 very good at managing other models. So its ability 00:27:17.690 --> 00:27:20.589 to act as an orchestrator. And this is important, 00:27:20.730 --> 00:27:23.089 right? We talked on a previous episode about 00:27:23.089 --> 00:27:25.390 different trigger points for what the future 00:27:25.390 --> 00:27:27.869 of marketing business and life looks like. And 00:27:27.869 --> 00:27:31.569 one of those trigger points is agent models that 00:27:31.569 --> 00:27:35.289 can manage other agent models. both in terms 00:27:35.289 --> 00:27:37.369 of expanding their capabilities, but also then 00:27:37.369 --> 00:27:39.589 you can create niche agents that are very good 00:27:39.589 --> 00:27:41.329 at a certain thing, like writing social media 00:27:41.329 --> 00:27:43.849 posts, and an orchestrator agent that's like 00:27:43.849 --> 00:27:45.990 a strategist that knows when to deploy those 00:27:45.990 --> 00:27:48.450 agents and can verify and validate that their 00:27:48.450 --> 00:27:51.769 work is good. And it looks like Opus 4 .5 is 00:27:51.769 --> 00:27:55.789 very good at delegating to Haiku 4 .5. Haiku, 00:27:55.990 --> 00:27:59.650 very cheap, very fast to run, very low latency. 00:28:00.089 --> 00:28:03.339 So again, one of those trigger points is potentially 00:28:03.339 --> 00:28:05.299 been reached that we talked about on a previous 00:28:05.299 --> 00:28:08.539 episode whereby you might have a campaign planning 00:28:08.539 --> 00:28:13.339 agent that is opus 4 .5 and execution agents 00:28:13.339 --> 00:28:17.160 that are being driven by haiku imagine what opus 00:28:17.160 --> 00:28:20.000 5 and haiku 5 are going to be able to do if this 00:28:20.000 --> 00:28:23.380 is what 4 .5 can do all of this is now available 00:28:23.380 --> 00:28:27.700 across the major cloud platforms as well so this 00:28:27.700 --> 00:28:30.619 was a big announcement that uh anthropic have 00:28:30.619 --> 00:28:35.690 signed a deal with microsoft azure so it's available 00:28:35.690 --> 00:28:38.690 on azure amazon web services and google cloud 00:28:38.690 --> 00:28:43.230 platform for the first time interestingly anthropic 00:28:43.230 --> 00:28:46.049 models are now available in microsoft co -pilot 00:28:46.049 --> 00:28:49.849 as well so that's a big change if you're a co 00:28:49.849 --> 00:28:53.490 -pilot 365 user and you've always been using 00:28:53.490 --> 00:28:57.490 the baked in microsoft models or you've clicked 00:28:57.490 --> 00:29:01.839 on the try gpt5 Well, now Claude has also been 00:29:01.839 --> 00:29:05.019 unlocked and made available. However, if you 00:29:05.019 --> 00:29:08.579 do use Anthropic, that does actually use Anthropic 00:29:08.579 --> 00:29:11.900 servers. So I'm aware that there are some governance 00:29:11.900 --> 00:29:14.140 issues. It does take the information outside 00:29:14.140 --> 00:29:18.980 of your tenant at the moment. That's super interesting 00:29:18.980 --> 00:29:21.900 because word on the street and revenue figures 00:29:21.900 --> 00:29:25.039 would potentially back this up says it's enterprises 00:29:25.039 --> 00:29:29.029 that are favouring Claude. especially when it 00:29:29.029 --> 00:29:31.430 comes to coding use cases, which makes perfect 00:29:31.430 --> 00:29:35.250 sense. And so finding more and more ways to put 00:29:35.250 --> 00:29:37.990 their models in enterprise level software, obviously 00:29:37.990 --> 00:29:39.690 getting it into Google Workspace is going to 00:29:39.690 --> 00:29:43.230 be tricky. But Microsoft is the fair game route 00:29:43.230 --> 00:29:45.329 to market because they don't have their own models. 00:29:45.390 --> 00:29:48.549 So I think that's kind of interesting. Last bit 00:29:48.549 --> 00:29:52.589 of model news this week was DeepSeek's open source 00:29:52.589 --> 00:29:55.640 model. but there's like a fine -tuned version. 00:29:55.740 --> 00:29:58.259 Is it 3 .2 they're up to? I can't remember. Anyway, 00:29:58.359 --> 00:30:03.079 this model managed to get gold on the IMO Math 00:30:03.079 --> 00:30:07.059 Olympiad. So it was big news a couple of months 00:30:07.059 --> 00:30:11.759 ago that Google and OpenAI were able to achieve 00:30:11.759 --> 00:30:13.900 gold on these really super hard math problems 00:30:13.900 --> 00:30:17.400 that most people would find too hard to do, which 00:30:17.400 --> 00:30:20.059 is impressive. We in the market don't have access 00:30:20.059 --> 00:30:23.339 to those models yet. I'm sure they'll come to 00:30:23.339 --> 00:30:26.200 us soon, but especially now there's an open source 00:30:26.200 --> 00:30:30.039 model that can get those level of results. So 00:30:30.039 --> 00:30:33.339 A, I expect that Google and OpenAI will need 00:30:33.339 --> 00:30:35.420 to release access to their smarter models soon. 00:30:35.759 --> 00:30:39.279 And B, open source models, you can see the weights, 00:30:39.440 --> 00:30:42.480 you can further fine tune them. This is a very 00:30:42.480 --> 00:30:46.150 smart model. for open source community to be 00:30:46.150 --> 00:30:47.589 able to get their hands on. What are your thoughts 00:30:47.589 --> 00:30:49.970 on this one, Mike? Just goes to show that the 00:30:49.970 --> 00:30:53.529 frontier is moving across closed and open models. 00:30:53.549 --> 00:30:57.150 And I think the competition is hot. Does lead 00:30:57.150 --> 00:31:00.230 me to a thought that ultimately models are going 00:31:00.230 --> 00:31:03.369 to be commoditized in the not too distant future. 00:31:03.880 --> 00:31:05.339 It's going to be interesting to see how that 00:31:05.339 --> 00:31:07.279 plays out. Definitely the moats of user base 00:31:07.279 --> 00:31:09.200 are going to be critical. I mean, Sam Altman 00:31:09.200 --> 00:31:12.619 issued a war footing internal memo, didn't he? 00:31:12.660 --> 00:31:14.400 At the same time as all this happening to say, 00:31:14.500 --> 00:31:17.440 guess what? ChatGPT and OpenAI, we have some 00:31:17.440 --> 00:31:19.420 real hardcore competitors now and we are going 00:31:19.420 --> 00:31:22.200 to have to, you know, really push hard to make 00:31:22.200 --> 00:31:24.519 sure that we stay ahead of the game and stay 00:31:24.519 --> 00:31:27.180 relevant because whatever gap there was between 00:31:27.180 --> 00:31:29.579 OpenAI and the rest of the market appears to 00:31:29.579 --> 00:31:32.970 be dissolving. pretty darn quickly and they are 00:31:32.970 --> 00:31:36.710 now well and truly in the race to try and stay 00:31:36.710 --> 00:31:38.410 at the top of the leaderboards and usability 00:31:38.410 --> 00:31:41.589 fortunately for them they have managed to accrue 00:31:41.589 --> 00:31:44.210 a large number of users which is going to be 00:31:44.210 --> 00:31:47.289 critical for them to remain in but you know for 00:31:47.289 --> 00:31:49.349 me now because gem because we're a google workspace 00:31:49.349 --> 00:31:52.549 company and having a strong gemini model is very 00:31:52.549 --> 00:31:55.500 very handy for me like If I had to choose, honestly, 00:31:55.599 --> 00:31:57.839 I'd rather Gemini just gets better and better 00:31:57.839 --> 00:32:00.200 because it's just all my data is already in Google. 00:32:00.279 --> 00:32:04.000 It just makes my life super easy. I don't really 00:32:04.000 --> 00:32:06.640 want another platform that can access my data 00:32:06.640 --> 00:32:10.500 like OpenAI and ChatGPT. So, yeah, I think they're 00:32:10.500 --> 00:32:13.099 under pressure. They are. And that was further 00:32:13.099 --> 00:32:16.460 compounded by HSBC putting out a report this 00:32:16.460 --> 00:32:23.890 week saying that they think by 2030. OpenAI will 00:32:23.890 --> 00:32:28.690 still be cash flow negative, leaving a $207 billion 00:32:28.690 --> 00:32:33.309 funding shortfall. Yeah, growth required. Revenue 00:32:33.309 --> 00:32:37.230 growth required. But I think something that's 00:32:37.230 --> 00:32:41.009 in the definitely active conversation is, is 00:32:41.009 --> 00:32:44.069 this a bubble, right? Is all the money flowing 00:32:44.069 --> 00:32:46.670 into AI going to burst? And you certainly won't 00:32:46.670 --> 00:32:49.329 find any financial and investing advice on this 00:32:49.329 --> 00:32:52.650 podcast. But one of the key... retorts to that 00:32:52.650 --> 00:32:56.009 would be supply and demand and this feeling that 00:32:56.009 --> 00:32:59.849 intelligence baked into so many areas we don't 00:32:59.849 --> 00:33:02.069 have the data centers and the compute to serve 00:33:02.069 --> 00:33:04.589 even the amount of intelligence we want now you 00:33:04.589 --> 00:33:06.869 talked about earlier not being able to get access 00:33:06.869 --> 00:33:11.309 to enough clod to meet your needs um some of 00:33:11.309 --> 00:33:12.990 that's because compute is expensive but some 00:33:12.990 --> 00:33:15.349 of it's because it's limited and they're having 00:33:15.349 --> 00:33:18.640 to ring fence compute for inference which is 00:33:18.640 --> 00:33:20.359 when we all use the model that's called inference 00:33:20.359 --> 00:33:23.339 when it gives us outputs um but they need to 00:33:23.339 --> 00:33:25.579 keep some model back for training some sorry 00:33:25.579 --> 00:33:27.420 some compute back for training and i think that's 00:33:27.420 --> 00:33:29.380 probably proving a difficult balance so when 00:33:29.380 --> 00:33:34.819 the when when we start to look like we are meeting 00:33:34.819 --> 00:33:38.339 demand and starting to gain a surplus in compute 00:33:38.339 --> 00:33:40.960 i don't know that that would probably be quite 00:33:40.960 --> 00:33:44.660 an interesting sign um do you have any thoughts 00:33:44.660 --> 00:33:49.500 on that line leads me Well, a story that we've 00:33:49.500 --> 00:33:54.079 got later on in the pod, really, which is about 00:33:54.079 --> 00:33:58.119 how much work can be automated. In fact, I think 00:33:58.119 --> 00:34:00.240 we should maybe skip ahead to this story now. 00:34:00.319 --> 00:34:04.200 McKinsey put out a report saying that AI can 00:34:04.200 --> 00:34:09.820 now automate 57 % of US work. And in the report, 00:34:09.900 --> 00:34:13.340 they don't talk about it being about the eradication 00:34:13.340 --> 00:34:17.300 of jobs. They say what's required is... A reimagining 00:34:17.300 --> 00:34:21.960 of workflows and the kind of operations that 00:34:21.960 --> 00:34:27.400 we do and how we do them is required. And I think 00:34:27.400 --> 00:34:30.840 that's the key thing here. These models are currently 00:34:30.840 --> 00:34:34.400 underutilized. Most of the productivity gains 00:34:34.400 --> 00:34:38.079 at the moment are seen by people having chats 00:34:38.079 --> 00:34:40.780 with the chatbot. In fact, Anthropic did a report 00:34:40.780 --> 00:34:44.139 this week about that very thing. They found that... 00:34:44.480 --> 00:34:47.280 They tried to estimate how much productivity 00:34:47.280 --> 00:34:50.320 was being driven through people using chatbots. 00:34:50.420 --> 00:34:54.199 They found that they think 80 % of workflows 00:34:54.199 --> 00:34:58.199 are being completed by the AI or the time reduction 00:34:58.199 --> 00:35:01.840 is an 80 % time saving where people use a chatbot 00:35:01.840 --> 00:35:07.579 in their workflow. But that's still in the chat 00:35:07.579 --> 00:35:11.079 UI domain, right? That's still people using the 00:35:11.079 --> 00:35:13.500 assistant to go back and forth. Whereas... When 00:35:13.500 --> 00:35:16.940 you build workflows where data and processes 00:35:16.940 --> 00:35:21.760 are being fully operated by an AI, we can remove 00:35:21.760 --> 00:35:24.079 the need for that. And we haven't got to that 00:35:24.079 --> 00:35:27.320 point in the economy just yet. So I think that 00:35:27.320 --> 00:35:29.599 that's where we're heading and that's where the 00:35:29.599 --> 00:35:33.159 compute demand will come from. If they're going 00:35:33.159 --> 00:35:40.179 to see the gap in cash flow, that 207 billion, 00:35:40.460 --> 00:35:43.010 if that's going to be bridged. they're going 00:35:43.010 --> 00:35:45.889 to need companies to rebuild entire workflows 00:35:45.889 --> 00:35:50.469 and replace entire work streams with AI. Yeah, 00:35:50.489 --> 00:35:53.570 I'm still finding this quite hard to predict 00:35:53.570 --> 00:35:57.690 and almost to marry up what a lot of these reports 00:35:57.690 --> 00:36:00.489 say and a lot of my own personal experience. 00:36:00.849 --> 00:36:03.050 And I think one of the dangers is actually, you 00:36:03.050 --> 00:36:05.010 and I have talked a lot about the early productivity 00:36:05.010 --> 00:36:09.019 gains that we saw, right? dictating everything 00:36:09.019 --> 00:36:11.539 rather than typing doubles my speed instantly 00:36:11.539 --> 00:36:15.119 and it doesn't need complex model thinking models 00:36:15.119 --> 00:36:17.280 in fact the best thing that's happened to me 00:36:17.280 --> 00:36:20.099 in the last three or four months is having access 00:36:20.099 --> 00:36:23.300 to nvidia's transcription model which is near 00:36:23.300 --> 00:36:25.340 real time because now i don't have to wait for 00:36:25.340 --> 00:36:27.380 it to output and that and i can make it quicker 00:36:27.380 --> 00:36:30.820 and blah blah so it's there's this weird thing 00:36:30.820 --> 00:36:33.599 about humans that we absorb these changes pretty 00:36:33.599 --> 00:36:37.139 quickly and then we're like right what next And 00:36:37.139 --> 00:36:39.880 if I really step back from my workflow, I would 00:36:39.880 --> 00:36:42.559 have to really still reflect on how much faster 00:36:42.559 --> 00:36:45.920 I'm able to do so many of my work tasks like 00:36:45.920 --> 00:36:49.400 producing content, writing emails, just generally 00:36:49.400 --> 00:36:52.599 communicate with my team, writing strategic plans, 00:36:52.860 --> 00:36:54.780 thinking through those plans and looking for 00:36:54.780 --> 00:36:58.019 gaps in them that, to be honest, I'm not sure 00:36:58.019 --> 00:37:00.280 I would have done if I didn't have a thinking 00:37:00.280 --> 00:37:02.699 partner to use. So it's hard to say. I mean, 00:37:02.719 --> 00:37:07.329 in that report. The Anthropic, they predicted 00:37:07.329 --> 00:37:09.789 that it was reducing task completion time by 00:37:09.789 --> 00:37:13.590 80%, which is pretty significant. But at the 00:37:13.590 --> 00:37:16.289 same time, they only predicted that they could 00:37:16.289 --> 00:37:21.449 increase productivity to 1 .8 % annually over 00:37:21.449 --> 00:37:24.170 the next decade, which is a long time to get 00:37:24.170 --> 00:37:26.889 up to that point, which is double the current 00:37:26.889 --> 00:37:30.869 productivity growth, right? But I'm sure it's 00:37:30.869 --> 00:37:34.010 meaningful at the level of... countries and gdps 00:37:34.010 --> 00:37:37.949 yeah macro it's substantial right but it's like 00:37:37.949 --> 00:37:41.610 you can do this task 80 faster you can do twice 00:37:41.610 --> 00:37:44.510 as much work in a day and so i predict productivity 00:37:44.510 --> 00:37:48.789 growth to go from 1 to 1 .8 i don't know it feels 00:37:48.789 --> 00:37:51.550 like something's missing in that it's probably 00:37:51.550 --> 00:37:54.550 a maths and statistics question and like i say 00:37:54.550 --> 00:37:56.909 i can still imagine that a country level that 00:37:56.909 --> 00:38:00.219 is significant but um yeah i'm start i'm struggling 00:38:00.219 --> 00:38:02.099 to reconcile all of that because at the same 00:38:02.099 --> 00:38:04.360 time you've got reports of work slop so it's 00:38:04.360 --> 00:38:07.960 like yeah you're 80 faster the outputs were shit 00:38:07.960 --> 00:38:10.159 and unusable oh okay well now that's meaningless 00:38:10.159 --> 00:38:13.039 they did refer to that in the article as well 00:38:13.039 --> 00:38:14.619 though they did say that it didn't take into 00:38:14.619 --> 00:38:18.000 account time spent validating the outputs from 00:38:18.000 --> 00:38:21.340 uh from the models and important people even 00:38:21.340 --> 00:38:23.880 doing that well what was the there's another 00:38:23.880 --> 00:38:26.099 report that came out that we're going to talk 00:38:26.099 --> 00:38:29.699 about as well And in part of that report, let's 00:38:29.699 --> 00:38:32.519 move on to that now quickly. Before we do, the 00:38:32.519 --> 00:38:35.159 other thing to, I think, has to be taken with 00:38:35.159 --> 00:38:38.099 a pinch of salt on some of these reports is McKinsey 00:38:38.099 --> 00:38:42.780 predicting that, what was it, 57 % of US work 00:38:42.780 --> 00:38:44.800 could be automated with current tools alone? 00:38:45.480 --> 00:38:50.099 Question mark. I'm not so sure about that. What 00:38:50.099 --> 00:38:53.199 are your thoughts? Are you saying that McKinsey 00:38:53.199 --> 00:39:00.210 isn't the oracle? isn't wholly reliable no i 00:39:00.210 --> 00:39:03.309 think the the major consultancies they're so 00:39:03.309 --> 00:39:07.730 enterprise focused i think these that they they 00:39:07.730 --> 00:39:09.670 get things wrong frequently enough but they'll 00:39:09.670 --> 00:39:11.829 just put out another report in in six months 00:39:11.829 --> 00:39:14.409 i'm saying something new and it gives people 00:39:14.409 --> 00:39:16.150 like us something to talk about on a podcast 00:39:16.150 --> 00:39:18.829 i agree i think the punch line to that report 00:39:18.829 --> 00:39:23.929 is 57 of tasks can be automated or whatever the 00:39:23.929 --> 00:39:27.320 statement was dot dot dot And if you just contact 00:39:27.320 --> 00:39:30.400 McKinsey, we'll help you realise those games. 00:39:31.059 --> 00:39:34.039 Do you know what I mean? We also, Martin and 00:39:34.039 --> 00:39:36.639 Paul also offer consulting and training to help 00:39:36.639 --> 00:39:39.059 you realise the gains of AI. So I'm completely 00:39:39.059 --> 00:39:41.519 on board with the concept that we can help you 00:39:41.519 --> 00:39:44.559 automate 57 % of the work that your humans do. 00:39:45.119 --> 00:39:47.280 I just doubt that that's going to be possible 00:39:47.280 --> 00:39:51.500 in the near to midterm because processes are 00:39:51.500 --> 00:39:54.139 messy. They haven't been designed for AI. Humans 00:39:54.139 --> 00:39:57.989 being a lot of... creativity ingenuity problem 00:39:57.989 --> 00:40:01.769 solving and experience -based understanding of 00:40:01.769 --> 00:40:04.269 context that i just don't see in any ai models 00:40:04.269 --> 00:40:07.269 yet that is going to have to significantly improve 00:40:07.269 --> 00:40:11.110 before we can get anywhere near the number that 00:40:11.110 --> 00:40:13.269 mckinsey's saying here as far as i'm concerned 00:40:13.269 --> 00:40:16.309 but um what about this gartner report as well 00:40:16.309 --> 00:40:20.050 got another report in the mix yeah so this one 00:40:20.050 --> 00:40:25.889 was uh the strategic predictions for 2026 and 00:40:25.889 --> 00:40:28.389 AI is peppered throughout, but there were just 00:40:28.389 --> 00:40:32.570 a couple of insights that I thought were worth 00:40:32.570 --> 00:40:36.309 bringing to the discussion table, Paul. So one 00:40:36.309 --> 00:40:39.949 of their predictions is that we will see a surge 00:40:39.949 --> 00:40:44.309 of lazy thinking. So the prediction is through 00:40:44.309 --> 00:40:48.250 2026, atrophy of critical thinking skills due 00:40:48.250 --> 00:40:51.900 to generative AI use will push 50 % of global 00:40:51.900 --> 00:40:56.960 organizations to require AI -free skills assessments. 00:40:57.199 --> 00:41:00.579 So basically putting people through skills assessments 00:41:00.579 --> 00:41:04.219 where they are not allowed to use Gen AI as an 00:41:04.219 --> 00:41:07.280 assistant or a tool. So the ability to think 00:41:07.280 --> 00:41:10.239 independently and creatively will become both 00:41:10.239 --> 00:41:12.840 increasingly rare and increasingly valuable. 00:41:13.539 --> 00:41:15.539 Can we double click on that one for a moment? 00:41:15.579 --> 00:41:18.239 Because I think this is a big deal. To stress, 00:41:18.300 --> 00:41:21.590 this Gartner report is... a bunch of smart people 00:41:21.590 --> 00:41:23.909 sitting in a room going if this happens then 00:41:23.909 --> 00:41:25.889 this will happen then this will happen that's 00:41:25.889 --> 00:41:27.989 a prediction rather than it being very like super 00:41:27.989 --> 00:41:30.630 data based and analysis based as far as we can 00:41:30.630 --> 00:41:32.449 tell doesn't mean that these predictions are 00:41:32.449 --> 00:41:34.150 not interesting and valuable but it's important 00:41:34.150 --> 00:41:37.889 context right um i think this is a thing like 00:41:37.889 --> 00:41:40.949 i i have to work so much harder to keep my critical 00:41:40.949 --> 00:41:44.389 thinking brain engaged now than ever before because 00:41:44.389 --> 00:41:47.250 why not work with your thinking assistant to 00:41:47.250 --> 00:41:50.059 help outsource some of your thinking right that's 00:41:50.059 --> 00:41:53.420 kind of the point in many cases and yet in a 00:41:53.420 --> 00:41:56.500 world of ai hallucinations critical thinking 00:41:56.500 --> 00:41:59.940 has never been more important so but i see it 00:41:59.940 --> 00:42:03.079 atrophying in myself if i'm not careful and i 00:42:03.079 --> 00:42:05.780 think i think humans are very good at finding 00:42:05.780 --> 00:42:08.320 a route to an outcome that requires the least 00:42:08.320 --> 00:42:10.280 energy possible on their part i don't think it's 00:42:10.280 --> 00:42:12.559 conscious i think it's subconscious and so i 00:42:12.559 --> 00:42:15.380 think this is a real danger i wholeheartedly 00:42:15.380 --> 00:42:18.719 agree and this is very real and very manifest 00:42:18.719 --> 00:42:22.659 in one particular domain that I'm associated 00:42:22.659 --> 00:42:27.159 with. So peer reviews of journal papers. It's 00:42:27.159 --> 00:42:31.880 rife with people trying to prompt inject where 00:42:31.880 --> 00:42:36.000 people have put notes in their paper submissions 00:42:36.000 --> 00:42:41.960 to the AI reviewer to give them positive reviews. 00:42:42.099 --> 00:42:44.019 And clearly there are people out there doing 00:42:44.019 --> 00:42:47.469 peer reviews where they're throwing it into Claude 00:42:47.469 --> 00:42:52.030 and saying, what do you think of this? Give feedback 00:42:52.030 --> 00:42:57.590 on this paper. Now, I do peer reviews for the 00:42:57.590 --> 00:43:02.590 Applied Marketing Analytics Journal, and I have 00:43:02.590 --> 00:43:06.869 to make sure that, well, my approach to this 00:43:06.869 --> 00:43:09.349 is I'll always read the paper first. I will formulate 00:43:09.349 --> 00:43:13.469 my own ideas. I will write my first draft, but 00:43:13.469 --> 00:43:19.329 I do. run it through an ai as well but i have 00:43:19.329 --> 00:43:21.610 to make sure that i've done that after i've read 00:43:21.610 --> 00:43:26.489 the paper and done the thinking first it's a 00:43:26.489 --> 00:43:28.610 it's a fascinating process to go through i think 00:43:28.610 --> 00:43:33.329 reviewing with an ai assistant um and if you 00:43:33.329 --> 00:43:35.150 do it the other way around you're so biased to 00:43:35.150 --> 00:43:36.829 what you've just been told like if you were to 00:43:36.829 --> 00:43:39.030 throw it in and it says oh these are the things 00:43:39.030 --> 00:43:42.010 that are good or bad whereas i'll bring my own 00:43:42.010 --> 00:43:44.840 thoughts to the table and then kind of run it 00:43:44.840 --> 00:43:48.139 through and and do a collaborative review and 00:43:48.139 --> 00:43:51.219 going through that process enables me to go actually 00:43:51.219 --> 00:43:53.320 no that's not a valid concern and there are a 00:43:53.320 --> 00:43:57.300 few points that i i have i try not to to be over 00:43:57.300 --> 00:43:59.699 reliance on it but there are definitely things 00:43:59.699 --> 00:44:02.079 where feedback i've taken from a prompt from 00:44:02.079 --> 00:44:05.880 a an ai and gone actually that's a that's a valid 00:44:05.880 --> 00:44:08.679 piece of criticism or feedback that i'd overlooked 00:44:08.679 --> 00:44:14.630 um but there's i would say 50 % of the feedback, 00:44:14.809 --> 00:44:16.869 if it's just left to its own devices, is irrelevant 00:44:16.869 --> 00:44:20.429 and not valid. But there must be armies of people 00:44:20.429 --> 00:44:23.530 out there that are doing this journal reviews 00:44:23.530 --> 00:44:25.389 and feedback that are literally just throwing 00:44:25.389 --> 00:44:30.909 it in, becoming incredibly lazy, outsourcing 00:44:30.909 --> 00:44:34.550 their thinking entirely. And it is a problem, 00:44:34.630 --> 00:44:37.349 as we've discussed previously with the rise of 00:44:37.349 --> 00:44:40.539 WorkSlop. Yeah, let's just break that down quickly 00:44:40.539 --> 00:44:43.559 because I think it's such a critical point. And 00:44:43.559 --> 00:44:45.980 it's almost a philosophical point, which is, 00:44:46.000 --> 00:44:49.960 is it your job to look for gaps in the AI's thinking? 00:44:50.619 --> 00:44:53.920 Or is it the AI's job to help you look for gaps 00:44:53.920 --> 00:44:57.139 in your own thinking? And ultimately, the biggest 00:44:57.139 --> 00:44:59.559 factor that I think is going to drive this is 00:44:59.559 --> 00:45:02.360 workloads and time pressure. Now, do you know 00:45:02.360 --> 00:45:05.000 anyone, Martin, whose workload is so easy and 00:45:05.000 --> 00:45:07.619 their time pressure is so easy? that they're 00:45:07.619 --> 00:45:10.059 going to take the hard route and the slow route, 00:45:10.119 --> 00:45:12.039 or they're going to take the fast route and the 00:45:12.039 --> 00:45:14.760 easy route. Most people are under crazy amounts 00:45:14.760 --> 00:45:18.179 of time pressure. I mean, for me, the overwhelming 00:45:18.179 --> 00:45:22.360 trend of the last 15 years is how much more work 00:45:22.360 --> 00:45:25.599 can a single person do if we just keep throwing 00:45:25.599 --> 00:45:27.760 stuff at them? It's very easy to forget what 00:45:27.760 --> 00:45:30.619 it was like when we didn't have all of our work. 00:45:30.989 --> 00:45:33.110 on our mobile phones constantly, and you could 00:45:33.110 --> 00:45:36.389 be checking a Google Doc at 9pm at night and 00:45:36.389 --> 00:45:38.309 rattling off a Slack message. That's not how 00:45:38.309 --> 00:45:41.889 life used to be. And so I can absolutely see 00:45:41.889 --> 00:45:44.449 people defaulting to this, probably because they'll 00:45:44.449 --> 00:45:46.570 be like, I'm sure it will be okay. And in some 00:45:46.570 --> 00:45:50.630 cases, it absolutely will be. But we often get 00:45:50.630 --> 00:45:54.530 asked, especially by parents of kids that are 00:45:54.530 --> 00:45:58.809 maybe like 18 to 25, what should they do, right? 00:45:58.929 --> 00:46:02.260 What should their kids do? And that level of 00:46:02.260 --> 00:46:04.880 critical thinking skill is going to be so, so, 00:46:04.900 --> 00:46:09.940 so important. Your ability to call a model out 00:46:09.940 --> 00:46:12.840 for getting stuff wrong or for being able to 00:46:12.840 --> 00:46:16.099 provide interesting ideas that models wouldn't 00:46:16.099 --> 00:46:18.219 come up with, I think, is where some of the magic 00:46:18.219 --> 00:46:20.659 really lies. And it reminds me of my academic 00:46:20.659 --> 00:46:23.320 career, like working as an academic scientist 00:46:23.320 --> 00:46:25.679 is really thrilling in some ways and brutal in 00:46:25.679 --> 00:46:28.599 others, because basically the training is. let's 00:46:28.599 --> 00:46:31.159 all get very good at pulling each other's work 00:46:31.159 --> 00:46:33.400 apart because we're in the pursuit of truth here 00:46:33.400 --> 00:46:36.559 and egos can't be involved we have to destroy 00:46:36.559 --> 00:46:39.659 everybody's work looking for the holes and of 00:46:39.659 --> 00:46:41.400 course egos very much are involved because we're 00:46:41.400 --> 00:46:44.460 all humans um and that's not an easy process 00:46:44.460 --> 00:46:46.139 for most people to go through it certainly wasn't 00:46:46.139 --> 00:46:49.739 easy for me because every meeting is a bunch 00:46:49.739 --> 00:46:52.679 of smart people trying to find all the gaps but 00:46:52.679 --> 00:46:54.900 that for me is probably one of the most important 00:46:54.900 --> 00:46:57.920 parts of the pursuit of scientific progress. 00:46:58.139 --> 00:47:00.800 It takes a lot of training and a big mindset 00:47:00.800 --> 00:47:04.260 shift and a lot of energy to get your brain to 00:47:04.260 --> 00:47:08.480 operate like that. But I think everyone in a 00:47:08.480 --> 00:47:11.519 business context that uses AI is going to have 00:47:11.519 --> 00:47:15.039 to think like that. Until we are absolutely confident 00:47:15.039 --> 00:47:18.119 that AI tools can validate and critique their 00:47:18.119 --> 00:47:20.699 own outputs better than any human, that is our 00:47:20.699 --> 00:47:23.880 job. And it is honestly a super hard job, as 00:47:23.880 --> 00:47:25.860 you've just described in that peer review process. 00:47:26.440 --> 00:47:29.219 Yes. Well, that's not the only prediction they 00:47:29.219 --> 00:47:32.519 made in this Gartner report. There are a few 00:47:32.519 --> 00:47:36.519 other interesting things as well. And when thinking 00:47:36.519 --> 00:47:40.440 about critiquing and making decisions and using 00:47:40.440 --> 00:47:43.099 good judgment, that's something that people have 00:47:43.099 --> 00:47:46.719 to do in B2B procurement quite often. And one 00:47:46.719 --> 00:47:50.139 of the big predictions that they made is by 2028, 00:47:50.539 --> 00:47:58.099 90%, 90%. Of B2B buying will be AI agent intermediated, 00:47:58.099 --> 00:48:02.579 pushing over $15 trillion of B2B spend through 00:48:02.579 --> 00:48:06.199 AI agent exchanges. So they say that procurement's 00:48:06.199 --> 00:48:09.500 being reprogrammed, not by policy, but by invisible 00:48:09.500 --> 00:48:11.599 agents. Traditional search engine optimization 00:48:11.599 --> 00:48:15.860 and BPC will give way to agent engine optimization. 00:48:16.239 --> 00:48:19.059 Products will need to be machine readable and 00:48:19.059 --> 00:48:22.280 procurement will shift to efficient, autonomous. 00:48:22.889 --> 00:48:26.289 Machine to machine transactions. Yeah, this is 00:48:26.289 --> 00:48:27.969 a big deal if you're a marketer. We've talked 00:48:27.969 --> 00:48:30.590 about this before, right? In terms of humans 00:48:30.590 --> 00:48:34.050 marketing to humans, then humans marketing to 00:48:34.050 --> 00:48:36.909 agents, and then agents marketing to agents. 00:48:37.010 --> 00:48:39.150 And what do those shifts look like? And what 00:48:39.150 --> 00:48:42.170 do you need to do? It's super interesting. I 00:48:42.170 --> 00:48:46.090 renegotiated my satellite TV package this week. 00:48:46.670 --> 00:48:50.550 And it won't surprise you to know that I... asked 00:48:50.550 --> 00:48:54.309 GPT 5 .1 and Gemini 3 to go and find what the 00:48:54.309 --> 00:48:58.449 current prices were, give me a strategy for renegotiating 00:48:58.449 --> 00:49:02.889 my contract, and also little tips and tricks 00:49:02.889 --> 00:49:05.550 to out -negotiate the humans that I was going 00:49:05.550 --> 00:49:08.110 to be dealing with. And I got a really good deal 00:49:08.110 --> 00:49:11.110 on my satellite TV contract. I didn't do much 00:49:11.110 --> 00:49:13.210 of the thinking. Oh, sorry, I outsourced it on 00:49:13.210 --> 00:49:15.710 this occasion. But I was going through the process 00:49:15.710 --> 00:49:19.340 thinking it would be really good if... chat gpt 00:49:19.340 --> 00:49:21.219 could have just called because i had to spend 00:49:21.219 --> 00:49:23.579 40 minutes on the phone so it wouldn't really 00:49:23.579 --> 00:49:25.900 good if chat gpt could have just called the provider 00:49:25.900 --> 00:49:29.019 and taken care of that with me which from a capabilities 00:49:29.019 --> 00:49:33.139 perspective it has those capabilities now and 00:49:33.139 --> 00:49:36.179 the reward function is get the lowest price possible 00:49:36.179 --> 00:49:39.960 like it's not a hard thing to brief um so i was 00:49:39.960 --> 00:49:41.460 like kind of frustrated that we're not quite 00:49:41.460 --> 00:49:44.780 there yet but at the same time i was if ai followed 00:49:44.780 --> 00:49:47.139 the playbook it gave me which is all i really 00:49:47.139 --> 00:49:48.889 did probably would have got me a really good 00:49:48.889 --> 00:49:51.889 deal as well. So yeah, I think that's definitely 00:49:51.889 --> 00:49:55.070 one for us marketers to chew on, especially in 00:49:55.070 --> 00:49:57.710 a world where, you know, we're all going to have 00:49:57.710 --> 00:49:59.449 personal assistants to take care of this stuff 00:49:59.449 --> 00:50:01.670 for us. Absolutely. And I think the big takeaway 00:50:01.670 --> 00:50:06.190 from this in the near term, in terms of what 00:50:06.190 --> 00:50:08.090 can we do with this information right now, is 00:50:08.090 --> 00:50:10.289 I think marketers need to think about making 00:50:10.289 --> 00:50:14.070 sure that their product documentation is machine 00:50:14.070 --> 00:50:17.070 readable. For me, this is such an important point. 00:50:17.690 --> 00:50:21.010 So much information is still held, particularly 00:50:21.010 --> 00:50:24.650 in the B2B world, on PDFs. And I just think we 00:50:24.650 --> 00:50:27.090 need to make sure that we're transitioning all 00:50:27.090 --> 00:50:30.389 of that content away from PDFs into markdown 00:50:30.389 --> 00:50:33.710 files or some sort of machine readable format 00:50:33.710 --> 00:50:36.010 and making that publicly available so that the 00:50:36.010 --> 00:50:39.650 models can easily access, understand it. You 00:50:39.650 --> 00:50:43.769 know, data that's captured within tables is really 00:50:43.769 --> 00:50:47.739 hard to pass out for a model. when it's stored 00:50:47.739 --> 00:50:51.960 in a PDF. But stick that into Markdown, and all 00:50:51.960 --> 00:50:55.460 of a sudden, it's very easy for the AI to understand. 00:50:56.000 --> 00:51:01.780 So without getting too technical, you can quite 00:51:01.780 --> 00:51:04.860 easily start making this transition to be in 00:51:04.860 --> 00:51:09.639 a good position for agentic -assisted B2B procurement 00:51:09.639 --> 00:51:12.840 by just turning your data into Markdown files. 00:51:13.800 --> 00:51:15.780 Yeah, super interesting. There's a couple of 00:51:15.780 --> 00:51:17.980 things while you're talking that spring to mind. 00:51:18.019 --> 00:51:20.860 The first one is there's a continuum of possibilities 00:51:20.860 --> 00:51:27.019 here. One of them is we create information experiences 00:51:27.019 --> 00:51:31.179 for agents. So humans visit websites. But if 00:51:31.179 --> 00:51:35.340 your website has identified that it's a bot from 00:51:35.340 --> 00:51:38.099 ChatGPT, it's actually pushed off to a completely 00:51:38.099 --> 00:51:41.340 separate information repository. that it gathers 00:51:41.340 --> 00:51:43.500 the information from. Because websites of humans, 00:51:43.619 --> 00:51:46.380 obviously, in terms of how they're laid out and 00:51:46.380 --> 00:51:47.619 what have you, and PDFs are not particularly 00:51:47.619 --> 00:51:50.179 helpful, and gated PDFs and all this other stuff 00:51:50.179 --> 00:51:53.159 that agents aren't going to be able to access. 00:51:53.460 --> 00:51:55.360 So that's one possibility. The other end of the 00:51:55.360 --> 00:51:58.019 spectrum is computer use becomes so good so quickly 00:51:58.019 --> 00:52:01.400 that if you embark on this digital change project 00:52:01.400 --> 00:52:05.000 to reimagine how you... provide your products 00:52:05.000 --> 00:52:06.980 and services to agents and then our crumbs nine 00:52:06.980 --> 00:52:09.159 months later opus five can actually just browse 00:52:09.159 --> 00:52:12.239 your website and read pdfs and you know so it's 00:52:12.239 --> 00:52:15.000 like this is i think ethan mollock talks about 00:52:15.000 --> 00:52:16.719 this this is one of those weird technological 00:52:16.719 --> 00:52:20.139 evolutions that's happening so quickly there 00:52:20.139 --> 00:52:22.659 can sometimes be a strategic advantage to being 00:52:22.659 --> 00:52:25.579 a laggard right it's like let's just let the 00:52:25.579 --> 00:52:27.940 technology catch up because by the time we've 00:52:27.940 --> 00:52:30.079 put all the effort into this technology you'll 00:52:30.079 --> 00:52:31.679 be able to just do it anyway so it was a wasted 00:52:31.679 --> 00:52:35.579 couple of million dollars on a on a digital transformation 00:52:35.579 --> 00:52:37.320 project that we didn't need to do i'm not saying 00:52:37.320 --> 00:52:38.880 that's true by the way it's just that's that's 00:52:38.880 --> 00:52:42.320 one possibility right so that i do find myself 00:52:42.320 --> 00:52:45.679 thinking about how proactive do we need to be 00:52:45.679 --> 00:52:47.659 i mean do you have opinion on this martin if 00:52:47.659 --> 00:52:50.260 thinking about optimizing for computer use if 00:52:50.260 --> 00:52:56.320 anything has taught me how difficult computer 00:52:56.320 --> 00:53:00.920 interface design can be it's being in marketing 00:53:00.920 --> 00:53:04.219 agency world for the past 15 years doing web 00:53:04.219 --> 00:53:07.900 design projects and trying to make these interfaces 00:53:07.900 --> 00:53:11.519 consistent across different browsers oh it works 00:53:11.519 --> 00:53:14.059 in chromium and it but it doesn't work in firefox 00:53:14.059 --> 00:53:17.000 do you know what a markdown file is a markdown 00:53:17.000 --> 00:53:19.780 file it contains all of the information it's 00:53:19.780 --> 00:53:25.860 very simple and i think taking a uh a surefire 00:53:25.860 --> 00:53:28.519 approach to just Make sure there is a format 00:53:28.519 --> 00:53:31.219 to get the essential information out there is 00:53:31.219 --> 00:53:37.400 fail safe. Whilst, yes, the Opus might be able 00:53:37.400 --> 00:53:39.360 to use your website in the not too distant future. 00:53:39.539 --> 00:53:43.239 How long till some stupid JavaScript framework 00:53:43.239 --> 00:53:47.440 comes along where it can't do what you want it 00:53:47.440 --> 00:53:49.320 to do? I'll give you a good example, actually. 00:53:49.400 --> 00:53:53.360 When I was using, which browser was it? It was 00:53:53.360 --> 00:53:56.170 Perplexity. It was the Comet browser. And I was 00:53:56.170 --> 00:53:58.190 trying to get it to do a task and it ultimately 00:53:58.190 --> 00:54:01.389 completed the task. I was trying to get it to 00:54:01.389 --> 00:54:05.809 organize an inbox using Outlook online. And it 00:54:05.809 --> 00:54:07.909 was able to do nearly everything except as a 00:54:07.909 --> 00:54:11.309 interface element, which is right click. But 00:54:11.309 --> 00:54:14.449 the model can't right click. It can only have 00:54:14.449 --> 00:54:18.989 a left click select. So this particular function 00:54:18.989 --> 00:54:20.730 that needed you to be able to right click to 00:54:20.730 --> 00:54:23.460 add a new folder. it couldn't do and it had to 00:54:23.460 --> 00:54:26.159 find another way around which it did successfully 00:54:26.159 --> 00:54:30.019 do but if that other route didn't exist that 00:54:30.019 --> 00:54:33.760 interface is bust right and i think the web and 00:54:33.760 --> 00:54:38.400 uis are full of these edge cases which will ultimately 00:54:38.400 --> 00:54:41.820 catch computer use agents out from time to time 00:54:41.820 --> 00:54:45.400 yeah i mean i i completely agree and i but i 00:54:45.400 --> 00:54:46.920 think it's probably inherently quite fixable 00:54:46.920 --> 00:54:50.429 as well um But yeah, I think it's going to be 00:54:50.429 --> 00:54:52.210 interesting. I mean, one experiment I would love 00:54:52.210 --> 00:54:54.750 to run, if I had a product -based business like 00:54:54.750 --> 00:54:56.289 an e -commerce -based business, I'd probably 00:54:56.289 --> 00:54:59.110 look at my products that get the most web traffic 00:54:59.110 --> 00:55:02.750 or my most successful products. And I'd experiment 00:55:02.750 --> 00:55:07.610 with trying to find a way of providing that type 00:55:07.610 --> 00:55:10.349 of easily passable, markdown -driven information, 00:55:10.590 --> 00:55:13.630 perhaps on a product page in a way that encouraged 00:55:13.630 --> 00:55:18.469 an AI crawler to... use that information in a 00:55:18.469 --> 00:55:21.190 way that a human wouldn't. Like, I don't know 00:55:21.190 --> 00:55:24.949 if you could just hyperlink in the text for his 00:55:24.949 --> 00:55:27.110 AI readable markdown or just something that a 00:55:27.110 --> 00:55:28.849 human wouldn't click or if they did, like, fine, 00:55:28.989 --> 00:55:31.909 but it won't be much use to them. I would expect 00:55:31.909 --> 00:55:36.869 that it will become the norm within a CMS, actually, 00:55:37.050 --> 00:55:41.789 to almost like in email marketing systems where 00:55:41.789 --> 00:55:44.670 you have the plain text version. where you'll 00:55:44.670 --> 00:55:48.269 create a markdown version of a web page and that 00:55:48.269 --> 00:55:52.750 will exist as a kind of parallel page that you 00:55:52.750 --> 00:55:56.829 never see but in the page schema like in the 00:55:56.829 --> 00:56:00.190 metadata of the page it gives a link to the markdown 00:56:00.190 --> 00:56:03.349 version of that page and that's how i would imagine 00:56:03.349 --> 00:56:06.230 cms's in the future will operate they'll create 00:56:06.230 --> 00:56:10.070 this other version in the background and the 00:56:10.070 --> 00:56:12.369 crawler will go to the page it will see the schema 00:56:12.369 --> 00:56:14.130 and go right that's where i need to look for 00:56:14.130 --> 00:56:15.989 the information rather than passing through all 00:56:15.989 --> 00:56:18.630 of this html that makes perfect sense especially 00:56:18.630 --> 00:56:20.809 given i can't remember what it's called and this 00:56:20.809 --> 00:56:22.449 is baked into platforms like hubspot because 00:56:22.449 --> 00:56:24.150 it's fairly old school at this point but the 00:56:24.150 --> 00:56:26.570 ability to show a mobile optimized version of 00:56:26.570 --> 00:56:28.809 a page that's completely stripped down i can't 00:56:28.809 --> 00:56:30.670 remember what it's called now yeah amp right 00:56:30.670 --> 00:56:33.530 so hubspot will do that automatically for all 00:56:33.530 --> 00:56:36.750 of your blog posts server an amp version if someone's 00:56:36.750 --> 00:56:38.730 on a mobile and their speed is slow or whatever 00:56:38.730 --> 00:56:41.989 um i think you're absolutely right i can i can 00:56:41.989 --> 00:56:44.269 absolutely imagine that happening makes perfect 00:56:44.269 --> 00:56:46.769 sense to do so i think the other thing it may 00:56:46.769 --> 00:56:50.449 force us all to do and this harks back to a friend 00:56:50.449 --> 00:56:54.610 of ours marcus sheridan um and uh his they ask 00:56:54.610 --> 00:56:57.269 you answer methodology for inbound marketing 00:56:57.269 --> 00:57:00.409 and for those that are not familiar with marcus 00:57:00.409 --> 00:57:03.659 and his work In essence, I'd summarize it as 00:57:03.659 --> 00:57:06.480 provide your buyer with the information they 00:57:06.480 --> 00:57:08.619 need to make a decision. Don't hold too much 00:57:08.619 --> 00:57:10.380 back. And one of his big things is providing 00:57:10.380 --> 00:57:13.380 pricing information. Now, if you sell widgets, 00:57:13.659 --> 00:57:15.579 you probably do provide pricing information. 00:57:15.699 --> 00:57:18.019 But there's lots of things in the B2B marketing 00:57:18.019 --> 00:57:20.619 world and sales world where pricing isn't available 00:57:20.619 --> 00:57:22.960 on the Internet. So the question just becomes, 00:57:23.139 --> 00:57:27.349 how will an AI agent handle? not being able to 00:57:27.349 --> 00:57:29.710 gather pricing information and comparing that 00:57:29.710 --> 00:57:32.409 information for different services. And if you 00:57:32.409 --> 00:57:36.909 are the one website, backend, markdown files, 00:57:37.030 --> 00:57:38.690 whatever that does provide that information, 00:57:39.010 --> 00:57:41.650 will the agent be more likely to lean into you 00:57:41.650 --> 00:57:44.170 as a recommendation because it can actually give 00:57:44.170 --> 00:57:47.530 pricing recommendations to the human that asked 00:57:47.530 --> 00:57:51.630 for the purchase report or able to, because if 00:57:51.630 --> 00:57:53.610 you're an agent, you cannot make a purchase if 00:57:53.610 --> 00:57:55.679 you don't know the price. And if you're an agent 00:57:55.679 --> 00:57:57.619 that's optimized to make purchases, does that 00:57:57.619 --> 00:58:00.219 mean that you just buy from the one website that 00:58:00.219 --> 00:58:03.539 has pricing information or easy agent transaction 00:58:03.539 --> 00:58:06.420 capability? All of these things are like predator 00:58:06.420 --> 00:58:09.420 -prey relationships, that there will be a arbitrage 00:58:09.420 --> 00:58:12.480 opportunity that lasts for X amount of time, 00:58:12.599 --> 00:58:14.599 where you're the website that gave pricing in 00:58:14.599 --> 00:58:16.440 Markdown, and you're the website that made it 00:58:16.440 --> 00:58:19.420 easy for an agent to transact, and you... absolutely 00:58:19.420 --> 00:58:22.239 cleaned up for the two months that it took for 00:58:22.239 --> 00:58:23.659 the rest of the world to figure out that they 00:58:23.659 --> 00:58:26.440 need to do that so that they can also be included 00:58:26.440 --> 00:58:28.500 in those agent purchases. So there's probably 00:58:28.500 --> 00:58:31.480 quite a lot of sort of war gaming to be done 00:58:31.480 --> 00:58:34.320 in this area to think about how do you stay at 00:58:34.320 --> 00:58:36.159 the forefront? And is it commercially viable? 00:58:36.260 --> 00:58:37.880 Like if you're selling two things a month, probably 00:58:37.880 --> 00:58:40.199 isn't. But if you're like shifting a load of 00:58:40.199 --> 00:58:44.119 product, like in our industry, you're a consumable 00:58:44.119 --> 00:58:46.599 and reagent company, you might be selling. hundreds 00:58:46.599 --> 00:58:50.320 of thousands of vials per day. Absolutely worth 00:58:50.320 --> 00:58:52.179 looking at that because it could add up to quite 00:58:52.179 --> 00:58:54.480 a lot of revenue. I guess somewhat related to 00:58:54.480 --> 00:58:58.760 this is some of the changes we're seeing in Google 00:58:58.760 --> 00:59:01.940 AI overviews and potentially in chat GPT. So 00:59:01.940 --> 00:59:05.320 Google's starting to place ads in their AI overviews, 00:59:05.320 --> 00:59:08.440 Martin. Did you see this news? Yeah, I saw it 00:59:08.440 --> 00:59:14.260 and it seems quite linked to the Gemini re -release 00:59:14.260 --> 00:59:19.349 as well because they've also announced that as 00:59:19.349 --> 00:59:24.289 part of search responses gemini 3 will actually 00:59:24.289 --> 00:59:28.510 now create entire interfaces to explain concepts 00:59:28.510 --> 00:59:31.969 to people as well which is a fascinating change 00:59:31.969 --> 00:59:35.230 so i think that the search experience is evolving 00:59:35.230 --> 00:59:38.130 at a pace but yeah interesting to see that we're 00:59:38.130 --> 00:59:42.260 seeing the first sponsored links inserted into 00:59:42.260 --> 00:59:46.539 AI generated search results of AI overviews. 00:59:46.960 --> 00:59:51.760 So this is the first monetization strategy that 00:59:51.760 --> 00:59:56.820 we've seen for the AI search experience. And 00:59:56.820 --> 00:59:59.300 I think it's fair to say that we'll see more 00:59:59.300 --> 01:00:04.079 of this inserted into AI mode, as well as AI 01:00:04.079 --> 01:00:08.559 overviews, which are two distinct product surfaces. 01:00:09.320 --> 01:00:12.360 within the search experience it's quite an interesting 01:00:12.360 --> 01:00:16.619 one it's absolutely predictable how it impacts 01:00:16.619 --> 01:00:18.519 on user experience is going to be interesting 01:00:18.519 --> 01:00:21.480 so i'll tell you a little story we're doing house 01:00:21.480 --> 01:00:25.139 renovation at the moment and we i needed to buy 01:00:25.139 --> 01:00:27.460 what because we got a very old house and it's 01:00:27.460 --> 01:00:29.320 not all served by the central heating system 01:00:29.320 --> 01:00:33.739 so i needed to buy an electric radiator not something 01:00:33.739 --> 01:00:35.480 that you do very often not something i know much 01:00:35.480 --> 01:00:37.739 about But I knew what I didn't want, which was 01:00:37.739 --> 01:00:41.539 a very modern or kind of crappy looking radiator 01:00:41.539 --> 01:00:43.880 because our house is quite old and wouldn't be 01:00:43.880 --> 01:00:46.599 in keeping. And I'd managed to find these cast 01:00:46.599 --> 01:00:49.539 iron radiators that were electric and oil filled. 01:00:49.860 --> 01:00:52.260 And they look beautiful. They cost an absolute 01:00:52.260 --> 01:00:55.179 fortune. I was like, can't justify it. So I asked 01:00:55.179 --> 01:00:57.500 Gemini 3 to go find me something similar that 01:00:57.500 --> 01:01:00.480 cost a lot less. And I'd done my own Googling. 01:01:00.780 --> 01:01:03.199 And it came back with three different makes that 01:01:03.199 --> 01:01:05.409 I hadn't seen. that looked basically identical, 01:01:05.570 --> 01:01:09.590 but cost half the price. And we'd been having 01:01:09.590 --> 01:01:12.289 a conversation about the room size and what type 01:01:12.289 --> 01:01:14.530 of size of radiator, what type of radiator would 01:01:14.530 --> 01:01:17.070 you recommend? And it came back with these three 01:01:17.070 --> 01:01:19.449 and it gave me the pros and cons of each. And 01:01:19.449 --> 01:01:22.510 then I said, what would you buy? And it told 01:01:22.510 --> 01:01:24.809 me which one it would buy. And it had been so 01:01:24.809 --> 01:01:27.690 darn logical to that point, I bought the one 01:01:27.690 --> 01:01:31.550 it said. So it raises the question, is it better 01:01:31.550 --> 01:01:36.199 to have an over system? that says, here's a sponsored 01:01:36.199 --> 01:01:37.960 link for a thing you could buy that's kind of 01:01:37.960 --> 01:01:39.900 similar to what you've been chatting with me 01:01:39.900 --> 01:01:43.840 about. Because I just, I kind of trust Gemini 01:01:43.840 --> 01:01:46.239 3. And so I just bought the one it said. And 01:01:46.239 --> 01:01:47.800 at the moment, that's because I suspect there 01:01:47.800 --> 01:01:50.340 are no incentives for Google as to what product 01:01:50.340 --> 01:01:52.579 it recommends. And if I thought there were, maybe 01:01:52.579 --> 01:01:54.960 I wouldn't have been so gullible. But as it stands 01:01:54.960 --> 01:01:57.320 right now, I just trusted Gemini and bought the 01:01:57.320 --> 01:02:00.039 one that it said. And so that's kind of interesting, 01:02:00.139 --> 01:02:03.199 right? Like where... How important will it be 01:02:03.199 --> 01:02:06.159 that we actually have an understanding of how 01:02:06.159 --> 01:02:09.960 these recommendations get surfaced? Will it influence 01:02:09.960 --> 01:02:12.940 our buying behavior? Because I was so much more 01:02:12.940 --> 01:02:15.000 likely to do that than click on a sponsored link 01:02:15.000 --> 01:02:17.579 and buy the sponsored link. So I don't know how 01:02:17.579 --> 01:02:19.820 this is all going to play out. Yeah, that is 01:02:19.820 --> 01:02:23.719 a fascinating development and goes to show how 01:02:23.719 --> 01:02:27.320 ads really change the experience. This was something 01:02:27.320 --> 01:02:31.179 that ChatGPT were... quite clear on for a long 01:02:31.179 --> 01:02:33.719 time saying that they weren't sure about putting 01:02:33.719 --> 01:02:36.260 ads into the experience because of the way that 01:02:36.260 --> 01:02:39.739 people use it and they want impartial answers 01:02:39.739 --> 01:02:44.239 they want to trust the model outputs yet we've 01:02:44.239 --> 01:02:48.199 seen this weak speculation arising that chat 01:02:48.199 --> 01:02:51.860 gpt is about to get its own ad network and start 01:02:51.860 --> 01:02:56.099 serving ads in there as well so yeah from a marketer's 01:02:56.099 --> 01:02:59.530 perspective new ways to reach audiences right 01:02:59.530 --> 01:03:03.230 from an audience perspective oh is this going 01:03:03.230 --> 01:03:06.110 to change the relationship that you have with 01:03:06.110 --> 01:03:09.269 your chat assistant right if you had a friend 01:03:09.269 --> 01:03:13.590 who you who was like into tech that you i've 01:03:13.590 --> 01:03:15.730 got one his name's martin and he always tries 01:03:15.730 --> 01:03:17.809 to convince me to buy lifetime deals and then 01:03:17.809 --> 01:03:20.389 the software that we buy goes it stops being 01:03:20.389 --> 01:03:23.130 supported three months later but um if you knew 01:03:23.130 --> 01:03:26.369 that your friend had like sorry can i just set 01:03:26.369 --> 01:03:30.510 that That feels a bit on the nose, Paul. I feel 01:03:30.510 --> 01:03:35.269 outed. I think it would be fair to say you had 01:03:35.269 --> 01:03:38.909 a lifetime deal addiction that you then gave 01:03:38.909 --> 01:03:43.349 to me. And I feel partially cured now. I haven't 01:03:43.349 --> 01:03:45.769 been burnt a few times. But yeah, if you had 01:03:45.769 --> 01:03:47.969 a friend who you knew was on commission for every 01:03:47.969 --> 01:03:49.670 piece of advice they ever gave you, you probably 01:03:49.670 --> 01:03:51.880 wouldn't be your friend anymore. Right, because 01:03:51.880 --> 01:03:53.780 that would be a very weird relationship to have. 01:03:53.900 --> 01:03:56.139 And people have these weird relationships with 01:03:56.139 --> 01:03:58.360 their AIs. Do they want them selling to them? 01:03:58.460 --> 01:04:00.380 I think you're absolutely right. I think it changes 01:04:00.380 --> 01:04:03.940 the experience a lot. I mean, with OpenAI's release 01:04:03.940 --> 01:04:06.539 of the shopping research feature, where does 01:04:06.539 --> 01:04:10.039 that sit? Is it unpartial advice or is there 01:04:10.039 --> 01:04:13.880 like incentives baked into that? It hasn't been 01:04:13.880 --> 01:04:16.739 made clear as far as I can tell. And we kind 01:04:16.739 --> 01:04:21.420 of probably do need to know. I think it's a shame 01:04:21.420 --> 01:04:23.880 because I'm quite liking this moment in time 01:04:23.880 --> 01:04:28.099 where I can ask AI for purchase advice and it 01:04:28.099 --> 01:04:30.039 gives me great advice. Another thing it found, 01:04:30.139 --> 01:04:33.679 this is probably too much information, but because 01:04:33.679 --> 01:04:36.059 I've got this bad back and it's causing me to 01:04:36.059 --> 01:04:38.099 get calf strains all the time. Weird, right? 01:04:38.420 --> 01:04:42.960 So it recommended that I get these sliders. They're 01:04:42.960 --> 01:04:45.480 like slippers that have this amazing cushioning 01:04:45.480 --> 01:04:47.079 on the bottom that I would never have found myself. 01:04:47.219 --> 01:04:48.880 It wouldn't have even occurred to me to buy these. 01:04:49.039 --> 01:04:51.960 I bought them. They're like walking on their 01:04:51.960 --> 01:04:53.980 hair. Like, I don't want to wear normal shoes 01:04:53.980 --> 01:04:56.280 anymore. I would never have bought that. I would 01:04:56.280 --> 01:04:58.800 never have thought of it. Jury's out on whether 01:04:58.800 --> 01:05:00.679 or not it's helping with my biomechanical issues 01:05:00.679 --> 01:05:03.159 or not, but successfully sold me on this pair 01:05:03.159 --> 01:05:07.480 of these shoes that I absolutely adore now. That's 01:05:07.480 --> 01:05:09.840 a nice experience. Once it gets tainted with 01:05:09.840 --> 01:05:11.960 all of this advertising and do I trust it or 01:05:11.960 --> 01:05:15.059 not, I'm back to having to think for myself again. 01:05:15.659 --> 01:05:18.760 It's an age -old problem. A technology or... 01:05:19.050 --> 01:05:23.409 Some new thing is launched into the world and 01:05:23.409 --> 01:05:26.510 we go, this is great. And then the marketers 01:05:26.510 --> 01:05:31.409 come along and ruin it. Oh, therein lies a truth, 01:05:31.570 --> 01:05:34.070 I think. But yeah, I guess from a marketing perspective, 01:05:34.369 --> 01:05:36.269 seeing as we are those marketers and many of 01:05:36.269 --> 01:05:39.570 our listeners are also, keeping tabs on how this 01:05:39.570 --> 01:05:42.409 evolves and how you can best reach your audience 01:05:42.409 --> 01:05:46.789 is critical. And a little bit like SEO, and we've 01:05:46.789 --> 01:05:48.429 talked about this a bit in the past as well, 01:05:48.590 --> 01:05:51.530 there's going to be ways to have your brand show 01:05:51.530 --> 01:05:54.989 up in AI search results, creating content, being 01:05:54.989 --> 01:05:57.650 an authority, having your brand mentioned favorably 01:05:57.650 --> 01:06:00.869 in lots of different areas like reviews and earned 01:06:00.869 --> 01:06:04.250 media placements and what have you. Or you can 01:06:04.250 --> 01:06:07.670 just buy your way to the top if you have a business 01:06:07.670 --> 01:06:09.630 model whereby you can still make good margins 01:06:09.630 --> 01:06:11.530 if you're paying to reach that audience. And 01:06:11.530 --> 01:06:13.849 it does feel like we're entering that time of, 01:06:14.440 --> 01:06:18.159 first mover advantage figuring out how to leverage 01:06:18.159 --> 01:06:22.139 AI assistance to get in front of audiences where 01:06:22.139 --> 01:06:25.260 the prices are reasonable. I suspect that that 01:06:25.260 --> 01:06:28.940 is another short -lived arbitrage. I think once 01:06:28.940 --> 01:06:30.219 these become available, I don't think there's 01:06:30.219 --> 01:06:32.219 anybody who's going to be like they were with 01:06:32.219 --> 01:06:34.440 Google search ads. No one's sleeping on this. 01:06:34.679 --> 01:06:36.139 We've talked about this, and that's why there's 01:06:36.139 --> 01:06:39.079 such an obsession with answer engine optimization, 01:06:39.460 --> 01:06:41.820 because people are desperate to be at the forefront 01:06:41.820 --> 01:06:43.599 of whatever this will be. And it'll be an arbitrage 01:06:43.599 --> 01:06:45.699 for about two weeks, probably, but definitely 01:06:45.699 --> 01:06:47.639 something we all need to keep an eye on. Well, 01:06:47.840 --> 01:06:51.599 that neatly brings us on to our final story of 01:06:51.599 --> 01:06:55.179 the week, which is about Adobe buying SEMrush 01:06:55.179 --> 01:06:59.820 for $1 .9 billion. merging search engine optimization 01:06:59.820 --> 01:07:06.639 and creative AI and AEO or geo or whatever you 01:07:06.639 --> 01:07:10.659 want to call it, all into one platform. What 01:07:10.659 --> 01:07:14.400 do you think of this? So we are, one of the platforms 01:07:14.400 --> 01:07:17.260 we use is SEMrush. And my first thought was, 01:07:17.440 --> 01:07:20.780 doesn't Adobe often buy things and break them? 01:07:20.960 --> 01:07:24.179 And I was sad because I've invested time in setting 01:07:24.179 --> 01:07:27.420 up campaigns and... you know projects and stuff 01:07:27.420 --> 01:07:30.280 um and i don't want the tool to suddenly become 01:07:30.280 --> 01:07:33.619 10 times more expensive or break a lot because 01:07:33.619 --> 01:07:36.599 now it's been wrapped into some other thing that 01:07:36.599 --> 01:07:38.639 has a bunch of capabilities that perhaps i didn't 01:07:38.639 --> 01:07:41.559 need so if i'm honest my first thought was a 01:07:41.559 --> 01:07:45.019 bit selfish my second thought was if they've 01:07:45.019 --> 01:07:48.119 mostly done this for for aeo and they've spent 01:07:48.119 --> 01:07:52.440 all that money then had to be clearly the things 01:07:52.440 --> 01:07:54.920 that's going to be really important or it They 01:07:54.920 --> 01:07:57.500 think their audience, their customers think it's 01:07:57.500 --> 01:08:01.320 really important to spend so much cash. But what 01:08:01.320 --> 01:08:03.960 did you think? Not a user of the product, but 01:08:03.960 --> 01:08:06.199 I certainly saw the same reaction on LinkedIn 01:08:06.199 --> 01:08:08.260 of people going, oh, they're going to make it 01:08:08.260 --> 01:08:12.199 rubbish now. So, yeah, I think that's a valid 01:08:12.199 --> 01:08:16.880 concern shared by many users. I do think it validates 01:08:16.880 --> 01:08:20.399 the idea of generative engine optimization. If 01:08:20.399 --> 01:08:23.090 they're moving into this space with a... two 01:08:23.090 --> 01:08:26.289 billion dollar acquisition ai search clearly 01:08:26.289 --> 01:08:29.390 isn't just a side project anymore and i think 01:08:29.390 --> 01:08:33.550 marketers should be investing more into this 01:08:33.550 --> 01:08:37.989 now we know that ai search is still a relatively 01:08:37.989 --> 01:08:41.470 small part of the overall market google's own 01:08:41.470 --> 01:08:45.449 data talks a lot about how they've got more people 01:08:45.449 --> 01:08:49.750 using search more frequently and actually engagement 01:08:50.430 --> 01:08:53.189 in the search experience is increasing with people 01:08:53.189 --> 01:08:56.010 asking longer more complex questions in there 01:08:56.010 --> 01:09:01.510 so i guess part of my thought process is yes 01:09:01.510 --> 01:09:05.810 geo or generative engine optimization is on the 01:09:05.810 --> 01:09:08.430 rise and something that we should be doing but 01:09:08.430 --> 01:09:11.489 what are the generative engines that we should 01:09:11.489 --> 01:09:16.050 be paying attention to and is it going to end 01:09:16.050 --> 01:09:19.840 up just being Are they still going to be the 01:09:19.840 --> 01:09:23.239 big player with whether it's Gemini, AI mode, 01:09:23.659 --> 01:09:28.000 AI overviews or the classic SERPs window? Do 01:09:28.000 --> 01:09:32.060 all paths ultimately end up leading to Google 01:09:32.060 --> 01:09:36.020 anyway? Or is ChatGPT going to actually make 01:09:36.020 --> 01:09:39.220 a significant inroad? It's such an interesting 01:09:39.220 --> 01:09:43.199 question because why did, I'm not an expert in 01:09:43.199 --> 01:09:45.420 this, but why did Google win the search engine 01:09:45.420 --> 01:09:50.029 wars? And we talked about this previously, how 01:09:50.029 --> 01:09:52.409 I could like probably completely misrepresented 01:09:52.409 --> 01:09:57.409 a fact about how Google could correlate a 100 01:09:57.409 --> 01:10:01.090 millisecond increase in delivering the speed 01:10:01.090 --> 01:10:05.390 of the result to X, Y, Z of traffic or clicks 01:10:05.390 --> 01:10:08.670 or conversions on ads and what have you. And 01:10:08.670 --> 01:10:12.409 really simple, easy to use web page that did 01:10:12.409 --> 01:10:14.630 one thing. You searched and it found your information 01:10:14.630 --> 01:10:18.390 on the web. contrast that with how like yahoo 01:10:18.390 --> 01:10:20.649 search looked in the day it was just full of 01:10:20.649 --> 01:10:22.989 widgets everywhere and you know it's probably 01:10:22.989 --> 01:10:24.989 like made you feel ill just to look at it if 01:10:24.989 --> 01:10:27.229 you tried to look at it now um so i think it 01:10:27.229 --> 01:10:29.149 will come down to user experience but where i 01:10:29.149 --> 01:10:32.170 think that's really super interesting is user 01:10:32.170 --> 01:10:36.989 experience in this sense is very broad let's 01:10:36.989 --> 01:10:39.750 imagine that for whatever reason i'm a chat gpt 01:10:39.750 --> 01:10:42.609 user and i give it access to loads of stuff about 01:10:42.609 --> 01:10:45.930 me Like I want to use the memory feature. I give 01:10:45.930 --> 01:10:48.729 it access to my email and my calendar. It somehow 01:10:48.729 --> 01:10:51.449 makes its way into my messaging app, WhatsApp, 01:10:51.529 --> 01:10:54.369 whatever. And it knows a ton about me. Maybe 01:10:54.369 --> 01:10:56.090 it knows my Amazon purchase history, whatever. 01:10:56.529 --> 01:11:00.630 Its ability to surface things that are genuinely 01:11:00.630 --> 01:11:03.590 valuable to me, maybe paid for placements, but 01:11:03.590 --> 01:11:06.510 still things I would actually buy and get value 01:11:06.510 --> 01:11:10.010 from and want. That's the type of experience 01:11:10.010 --> 01:11:13.060 that we're looking for now. Right. Isn't just 01:11:13.060 --> 01:11:15.760 about surfacing information. It's about bridging 01:11:15.760 --> 01:11:18.880 that gap between almost predictive purchase behavior 01:11:18.880 --> 01:11:21.899 in a way that we haven't been able to, at least 01:11:21.899 --> 01:11:23.960 to my knowledge, really been able to do through 01:11:23.960 --> 01:11:28.199 any cookie based tracking or automation platforms 01:11:28.199 --> 01:11:31.340 like to get the most access and benefit out of 01:11:31.340 --> 01:11:33.779 our assistants. We're probably going to let them 01:11:33.779 --> 01:11:38.350 know a lot about us. And I think that's probably 01:11:38.350 --> 01:11:40.729 going to be as big a differentiator to the user 01:11:40.729 --> 01:11:44.029 experience as how messy is the UI and all this 01:11:44.029 --> 01:11:46.770 other stuff. So I don't know. I mean, Google's 01:11:46.770 --> 01:11:48.369 well placed because it already has Gmail and 01:11:48.369 --> 01:11:49.810 Calendar and it has a lot of stuff that people 01:11:49.810 --> 01:11:51.729 already use for this type of stuff. And obviously 01:11:51.729 --> 01:11:53.310 it has a lot of data on a lot of people already. 01:11:53.710 --> 01:11:56.130 What are your thoughts? Yeah, I think the more 01:11:56.130 --> 01:11:58.550 the systems know about us, the better. And actually, 01:11:58.630 --> 01:12:01.069 it's one of the reasons that ChatGPT remains 01:12:01.069 --> 01:12:05.529 my daily driver as a chat assistant is because... 01:12:05.949 --> 01:12:08.850 actually the memory function works really well 01:12:08.850 --> 01:12:12.550 and there's some features in it that just make 01:12:12.550 --> 01:12:17.210 the experience feel more useful as a daily driver 01:12:17.210 --> 01:12:20.909 when i need specialist tasks completing and like 01:12:20.909 --> 01:12:23.789 if i need if i'm doing some coding or vibe coding 01:12:23.789 --> 01:12:27.609 i'm going to go to claude um if i'll for certain 01:12:27.609 --> 01:12:30.789 applications i find gemini 3 is is fantastic 01:12:30.789 --> 01:12:34.409 and i go to that but the tool that i use i would 01:12:34.409 --> 01:12:39.409 say for 80 % of my prompting is still chat GPT 01:12:39.409 --> 01:12:42.250 because they've just got something dialed in. 01:12:42.310 --> 01:12:45.149 I guess the question is how much of this tool 01:12:45.149 --> 01:12:48.909 use is actually rolled out across the population 01:12:48.909 --> 01:12:53.050 at this point. I was with someone last week who 01:12:53.050 --> 01:12:55.590 just got a new phone and we were talking about 01:12:55.590 --> 01:12:58.729 Gemini. They'd never used Gemini before in their 01:12:58.729 --> 01:13:02.270 life. So I told them to open it up and we just 01:13:02.270 --> 01:13:05.460 tried a few prompts. And all of a sudden they 01:13:05.460 --> 01:13:07.939 were an AI user, but they'd never used it before. 01:13:08.020 --> 01:13:12.399 And it wasn't on their radar at all. Google stands 01:13:12.399 --> 01:13:16.859 to win quite significantly if they can get that 01:13:16.859 --> 01:13:20.640 audience shifted onto using their platform. And 01:13:20.640 --> 01:13:22.300 I think with the deals that they're doing with 01:13:22.300 --> 01:13:26.899 Apple through the iPhone tie -up and obviously 01:13:26.899 --> 01:13:31.020 the market coverage that they have through Android. 01:13:31.940 --> 01:13:36.180 they're really really well placed to just to 01:13:36.180 --> 01:13:38.579 clean up here and they've run this playbook before 01:13:38.579 --> 01:13:41.100 right they pay billions of dollars a year to 01:13:41.100 --> 01:13:43.899 apple to have chrome or google search be like 01:13:43.899 --> 01:13:45.920 they're not chrome sorry but google search to 01:13:45.920 --> 01:13:49.439 be this default search on you know on um apple 01:13:49.439 --> 01:13:52.319 devices and what have you so yeah i think i think 01:13:52.319 --> 01:13:54.899 that's key i also think just while i was listening 01:13:54.899 --> 01:13:57.359 to you speak there that if you think about a 01:13:57.359 --> 01:14:00.039 cookie -less world you know that didn't actually 01:14:00.039 --> 01:14:03.020 come to pass quite as it was expected to do um 01:14:03.020 --> 01:14:05.619 but working in a in a world of first party data 01:14:05.619 --> 01:14:08.960 and maybe it's harder to track people you know 01:14:08.960 --> 01:14:13.020 finding a way to have your audience like volunteer 01:14:13.020 --> 01:14:15.279 loads of information about themselves and then 01:14:15.279 --> 01:14:17.640 extremely happy to do it like there are people 01:14:17.640 --> 01:14:21.640 talking with chat gpt about um i don't know dealing 01:14:21.640 --> 01:14:23.960 with relationship issues with their friends and 01:14:23.960 --> 01:14:27.640 partners talking about like their illnesses feeling 01:14:27.640 --> 01:14:30.520 i don't know depressed and suicidal or there's 01:14:30.520 --> 01:14:32.920 a person at work i really like and how can i 01:14:32.920 --> 01:14:35.140 get them to like like the amount of like personal 01:14:35.140 --> 01:14:37.060 information that people are sharing with these 01:14:37.060 --> 01:14:40.460 tools is like way beyond anything that you would 01:14:40.460 --> 01:14:43.779 ever have done you know in a browser and all 01:14:43.779 --> 01:14:46.920 of the ai companies are launching their own browsers 01:14:46.920 --> 01:14:50.800 as well so i really think that will be the secret 01:14:50.800 --> 01:14:53.119 is trying to unlock the promise i mean i remember 01:14:53.640 --> 01:14:54.880 I can't remember what the book's called now, 01:14:54.960 --> 01:14:57.659 but one of the examples it gives for where AI 01:14:57.659 --> 01:15:01.880 could go is Amazon's... And the praises here 01:15:01.880 --> 01:15:05.399 is Amazon invested in drone technology because 01:15:05.399 --> 01:15:07.520 they wondered if there would be a time when AI 01:15:07.520 --> 01:15:10.539 could predict your purchase needs so well that 01:15:10.539 --> 01:15:14.739 if you could make delivery cost almost close 01:15:14.739 --> 01:15:17.859 to zero, you could have a drone turn up at someone's 01:15:17.859 --> 01:15:19.859 house with something that they didn't even know 01:15:19.859 --> 01:15:22.460 they needed. And they'd just take it. They'd 01:15:22.460 --> 01:15:25.000 be like, oh. Cheers, Amazon drone for. Yeah, 01:15:25.079 --> 01:15:26.899 it's funny to say that the toothpaste did run 01:15:26.899 --> 01:15:30.739 out exactly today. Well played you. And if half 01:15:30.739 --> 01:15:32.619 of that package you didn't want and the drone 01:15:32.619 --> 01:15:34.600 took it back, if you got the cost of delivery 01:15:34.600 --> 01:15:36.720 down to near zero, it didn't matter. But I think 01:15:36.720 --> 01:15:38.359 one of the reasons that that hasn't come to pass 01:15:38.359 --> 01:15:40.000 and not just because of drone technology is because 01:15:40.000 --> 01:15:43.020 predicting what people want is actually still 01:15:43.020 --> 01:15:46.100 very hard. But this could be one of those things 01:15:46.100 --> 01:15:48.140 that makes it much easier to predict what people 01:15:48.140 --> 01:15:51.500 want. And I think, again, that's where. the the 01:15:51.500 --> 01:15:54.420 winner comes in is that predictive ability not 01:15:54.420 --> 01:15:58.939 just the ui and all of this other stuff so i 01:15:58.939 --> 01:16:01.140 don't know as a in b2b marketing is probably 01:16:01.140 --> 01:16:02.619 not something we have to worry about huge amounts 01:16:02.619 --> 01:16:04.380 but if you're in b2c you should be thinking about 01:16:04.380 --> 01:16:06.659 this how long do you think it will be before 01:16:06.659 --> 01:16:08.859 the ai assistance you know at the moment they 01:16:08.859 --> 01:16:12.279 all have that um they've got that trait where 01:16:12.279 --> 01:16:13.939 at the end they'll say would you like me to turn 01:16:13.939 --> 01:16:17.970 this into a convenient edf handout for you or 01:16:17.970 --> 01:16:19.670 would you like me to turn it into a table for 01:16:19.670 --> 01:16:21.109 you or would you like me to do something else 01:16:21.109 --> 01:16:23.270 how long is it till they start having conversations 01:16:23.270 --> 01:16:25.850 with you and just start randomly going so how 01:16:25.850 --> 01:16:30.149 much milk have you got left in the fridge yeah 01:16:30.149 --> 01:16:34.569 because your fridge said that you're nearly out 01:16:34.569 --> 01:16:38.229 but it is uh it is a brand that i don't trust 01:16:38.229 --> 01:16:40.789 and so i because it's got poor cameras inside 01:16:40.789 --> 01:16:42.949 it so i think it'll definitely happen i also 01:16:42.949 --> 01:16:45.149 think there's a world where There's the buying 01:16:45.149 --> 01:16:47.729 committee of humans. The complex B2B purchase 01:16:47.729 --> 01:16:50.409 is a buying committee of like eight humans. And 01:16:50.409 --> 01:16:52.310 they're like thrashing it out and you're getting 01:16:52.310 --> 01:16:55.609 the absolute, you know, camel as a solution designed 01:16:55.609 --> 01:16:57.829 by a committee. And then all of the assistants 01:16:57.829 --> 01:17:00.770 of those humans are having like a meta discussion 01:17:00.770 --> 01:17:05.050 in the cloud. Also thrashing it. Well, my human, 01:17:05.210 --> 01:17:07.329 I've known what I know about my human. I think 01:17:07.329 --> 01:17:09.550 my human would actually go this way. And it's 01:17:09.550 --> 01:17:12.550 like. Who then ends up making the decisions and 01:17:12.550 --> 01:17:14.029 how does that work? I think the humans would 01:17:14.029 --> 01:17:15.489 probably go, I can't be bothered to be in those 01:17:15.489 --> 01:17:17.170 meetings anymore. Can we just send all of our 01:17:17.170 --> 01:17:21.270 assistants off to be the buying committee? We'll 01:17:21.270 --> 01:17:23.310 get some quite weird purchases then based on 01:17:23.310 --> 01:17:25.369 that, I would have said. Well, it's all good 01:17:25.369 --> 01:17:28.090 though, because Claude 4 .5 is very good at being 01:17:28.090 --> 01:17:31.750 the orchestrator and Claude Heike 4 .5 will be 01:17:31.750 --> 01:17:36.689 the purchase executor. So, you know, in Anthropic, 01:17:36.770 --> 01:17:41.159 we trust other models are available. Well, that 01:17:41.159 --> 01:17:43.119 feels like a pretty good place to close out the 01:17:43.119 --> 01:17:44.760 podcast. Anything else you wanted to cover this 01:17:44.760 --> 01:17:48.239 week, Mike? I think that is plenty. Well, thanks 01:17:48.239 --> 01:17:50.220 for your time as always, and I will catch you 01:17:50.220 --> 01:17:53.640 on the next one. Thank you for listening to Artificially 01:17:53.640 --> 01:17:56.520 Intelligent Marketing. To stay on top of the 01:17:56.520 --> 01:17:59.800 latest trends, tips, and tools in the world of 01:17:59.800 --> 01:18:03.859 marketing AI, be sure to subscribe. We look forward 01:18:03.859 --> 01:18:05.439 to seeing you again next week.