1
00:00:00,000 --> 00:00:10,000
Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the

2
00:00:10,000 --> 00:00:15,740
latest trends, tips and tools in the world of marketing AI, helping you get the best

3
00:00:15,740 --> 00:00:18,560
results from your marketing efforts.

4
00:00:18,560 --> 00:00:23,120
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:23,120 --> 00:00:27,320
Welcome to episode 11 of Artificially Intelligent Marketing.

6
00:00:27,320 --> 00:00:31,240
I'm here with my very good friend Martin Broadhurst. How are you doing Martin?

7
00:00:31,240 --> 00:00:40,400
Oh Paul, it's great to see you. I'm back in the saddle, ready to talk AI on a Friday afternoon.

8
00:00:40,400 --> 00:00:43,480
Still doesn't get any better than that, does it? We've got a bunch of really good stories

9
00:00:43,480 --> 00:00:47,600
to get through today as always. We're going to go deep into a few topics. We're going

10
00:00:47,600 --> 00:00:53,520
to talk about a company called Causal who used AI to generate a ton of content for their

11
00:00:53,520 --> 00:00:58,480
website and draw in a lot of traffic from search. Fairly controversial in many ways,

12
00:00:58,480 --> 00:01:03,160
so we're going to get into that. We're going to talk about some amendments to the proposed

13
00:01:03,160 --> 00:01:07,680
amendments to the EU AI Act and what that would mean for both the companies producing

14
00:01:07,680 --> 00:01:14,000
tools but also marketers as users of those tools. And we're going to talk about ChatGPT's

15
00:01:14,000 --> 00:01:19,960
plugin rollout to all users. We're also going to hear from an industry colleague of ours,

16
00:01:19,960 --> 00:01:25,040
Liam Lally, who's a PPC specialist from Zaddle. He's going to input his expert opinion on

17
00:01:25,040 --> 00:01:31,040
the automated PPC debate that we had in last week's episode. We'll get into that a bit

18
00:01:31,040 --> 00:01:36,320
later on. And we're also going to look at a tool of the week, which this week is going

19
00:01:36,320 --> 00:01:42,040
to be the voice generation tool, Jenny. And there's more. We're also going to look at

20
00:01:42,040 --> 00:01:48,320
some short snippets from across the worlds of AI and marketing. And we'll do that before

21
00:01:48,320 --> 00:01:52,120
we get into everything else. So let's take a look at some of those short snippets. So

22
00:01:52,120 --> 00:02:01,000
a couple of quick things then. ChatGPT is launching an iPhone app. Not massive news,

23
00:02:01,000 --> 00:02:05,520
but it's probably going to take the place of a few apps that sprung up in that space

24
00:02:05,520 --> 00:02:11,840
when ChatGPT and OpenAI weren't able to get into it fast enough. Perhaps the most interesting

25
00:02:11,840 --> 00:02:17,400
thing about it is that you can speak to it. So it will do transcription using OpenAI's

26
00:02:17,400 --> 00:02:21,000
whisper and in essence, you'll be able to talk to your phone and then get it to do some

27
00:02:21,000 --> 00:02:27,680
stuff for you. So that will start potentially to make ChatGPT a bit more of a competitor

28
00:02:27,680 --> 00:02:34,640
to Siri and Alexa and Google Assistant and those types of things.

29
00:02:34,640 --> 00:02:39,600
The biggest feature that they introduced in this that I was impressed with is the simple

30
00:02:39,600 --> 00:02:45,120
search functionality to be able to search your conversations rather than having to just

31
00:02:45,120 --> 00:02:48,280
scroll endlessly through your chat history.

32
00:02:48,280 --> 00:02:50,840
You can search your conversations now.

33
00:02:50,840 --> 00:02:54,880
Apparently so. I haven't tried it yet because I'm not on iPhone and I'm not in America,

34
00:02:54,880 --> 00:02:56,680
but apparently you can search your conversations.

35
00:02:56,680 --> 00:03:00,760
They'll have to bring that to the desktop app, surely?

36
00:03:00,760 --> 00:03:05,440
I mean, you would think that's a given, but I would have thought it was a given having

37
00:03:05,440 --> 00:03:10,980
it from the off. I guess this ChatGPT interface was developed by about four people in an afternoon

38
00:03:10,980 --> 00:03:12,120
from what I understand.

39
00:03:12,120 --> 00:03:15,680
So we should be amazed with what we have.

40
00:03:15,680 --> 00:03:16,680
Yeah, exactly.

41
00:03:16,680 --> 00:03:21,080
All right. Well, that's good. I don't have an iPhone either, so I won't be playing with

42
00:03:21,080 --> 00:03:25,520
that anytime soon. But if anybody who listens to the podcast does get to have a play, we'd

43
00:03:25,520 --> 00:03:31,160
love to hear what your experiences are. Hit us up on the Linkedins and the Twitters.

44
00:03:31,160 --> 00:03:38,040
Meta is planning to create it own AI chips. So they're joining Google, Microsoft and Amazon

45
00:03:38,040 --> 00:03:43,040
in the custom chip game, one assumes because they don't have to completely rely on other

46
00:03:43,040 --> 00:03:51,040
suppliers like Nvidia, but also because GPUs, which power a lot of machine learning research

47
00:03:51,040 --> 00:03:57,840
and also the final tools that are launched also run on GPUs. But it's likely that custom

48
00:03:57,840 --> 00:04:05,360
chips is going to be better. That is also connected to a story this week in Forbes about

49
00:04:05,360 --> 00:04:11,800
just how expensive it is to run these generative AI models in data centers. I think they were

50
00:04:11,800 --> 00:04:17,160
projecting that by 2028, the cost of running the models on data centers would be more than

51
00:04:17,160 --> 00:04:23,800
76 billion dollars. So I guess anything you can do to make that better, including custom

52
00:04:23,800 --> 00:04:25,400
chips makes sense.

53
00:04:25,400 --> 00:04:32,040
When I saw that Forbes story mine, I was prompted to think to myself, prompt being the iterative

54
00:04:32,040 --> 00:04:40,000
word that if the product is free or near free and these data centers are costing billions

55
00:04:40,000 --> 00:04:45,840
of dollars to run, then you are the product. So that might be something for us all to remember

56
00:04:45,840 --> 00:04:52,760
as we rush to leverage Bard and Bing, especially because they are free at the moment and even

57
00:04:52,760 --> 00:04:55,800
ChatGPT, it's relatively low price.

58
00:04:55,800 --> 00:05:01,440
Yeah. And if you think about what they're getting from all of the feedback from users

59
00:05:01,440 --> 00:05:07,880
giving the thumbs up and thumbs down to responses in ChatGPT, that is reinforcement learning

60
00:05:07,880 --> 00:05:11,840
from human feedback. That's something that they've trained the models on historically

61
00:05:11,840 --> 00:05:17,140
with people being paid 10 to 15 dollars an hour. Now they've got the whole world doing

62
00:05:17,140 --> 00:05:25,240
it for free. So yeah, I was going to say it's a cost saving, but I'm sure that the GPU expense

63
00:05:25,240 --> 00:05:30,600
is slightly more than the wage bill for those humans previously.

64
00:05:30,600 --> 00:05:34,760
Yes. Well, that's certainly something they're getting out of it. Who else knows what other

65
00:05:34,760 --> 00:05:38,840
purposes they might have with the data that we share and the information and the experiences

66
00:05:38,840 --> 00:05:42,360
we're having with the tool. But I suspect that's a big chunk of it, Martin. I think

67
00:05:42,360 --> 00:05:49,000
you're right. Another new story this week was that Google's new medical AI MedPalm 2,

68
00:05:49,000 --> 00:05:54,920
which Martin talked a little bit about last week, scored over 80% on a medical exam that

69
00:05:54,920 --> 00:06:00,000
a typical human would only get about 60% on. So that's pretty interesting stuff.

70
00:06:00,000 --> 00:06:05,400
And when we say a typical human, we are talking people trained on training to be doctors.

71
00:06:05,400 --> 00:06:10,680
Not like me, if I did the exam, I don't think I'd get 60%.

72
00:06:10,680 --> 00:06:18,120
Yes. Trained people who understand and should do well on that test versus say a four year

73
00:06:18,120 --> 00:06:22,400
old or Martin. I'm not suggesting they're the same thing.

74
00:06:22,400 --> 00:06:25,920
No comparison at all.

75
00:06:25,920 --> 00:06:31,680
No comments on the domain. No comment. Right. We'll move on. There is a general feeling

76
00:06:31,680 --> 00:06:37,800
that we've been sensing around the web that Palm 2, Google's new model is meaning that

77
00:06:37,800 --> 00:06:42,880
BARD is faster and provides better output. So there seems to be a fair bit of chat on

78
00:06:42,880 --> 00:06:47,800
the Twittersphere and it's certainly been my experience and I think yours as well, Martin,

79
00:06:47,800 --> 00:06:53,600
that it's better than it was. It also has a really cool new export to Gmail or G-Docs

80
00:06:53,600 --> 00:06:56,640
button, which I really love because if you then ask it to draft an email for you, you

81
00:06:56,640 --> 00:07:01,800
can click a button and it puts that straight into a new composed message for you in Gmail,

82
00:07:01,800 --> 00:07:12,680
which is pretty cool. At the same time, it does seem that very creative things like really

83
00:07:12,680 --> 00:07:18,320
interesting poetry and some of the other things that ChatGPT 4 has been able to do. BARD is

84
00:07:18,320 --> 00:07:23,520
not up to that level when it comes to being creative. So we're drifting into already starting

85
00:07:23,520 --> 00:07:28,800
to drift into which of the models is best for what you're or which of the tools is best

86
00:07:28,800 --> 00:07:34,160
for what you're trying to do because in general, BARD is a heck of a lot faster than ChatGPT

87
00:07:34,160 --> 00:07:40,640
4 as well. So it's sort of like horses for courses on that one. Many of you might have

88
00:07:40,640 --> 00:07:47,560
seen that Sam Altman, the CEO of OpenAI and other industry experts spent three hours testifying

89
00:07:47,560 --> 00:07:53,480
before the US Senate and that there was a general consensus that AI needs to be regulated

90
00:07:53,480 --> 00:07:59,160
but it's such a complicated issue, nobody seems to be able to suggest a good mechanism

91
00:07:59,160 --> 00:08:04,000
of regulation that everyone thinks would actually work. So we'll all have to keep an eye on

92
00:08:04,000 --> 00:08:12,040
that. And then there was a couple of other cool stories. There is a company called Hippocratic

93
00:08:12,040 --> 00:08:21,320
AI, which has just received a ton of funding, I think $50 million recently to develop an

94
00:08:21,320 --> 00:08:28,400
AI driven platform to provide medical advice. So like WebMD but driven by AI, co-founded

95
00:08:28,400 --> 00:08:35,520
by a team of physicians and hospital administrators, Medicare pros, AI researchers, etc. So that's

96
00:08:35,520 --> 00:08:43,000
pretty cool. It's got some fairly significant and well-known investors including Andreessen

97
00:08:43,000 --> 00:08:49,480
Horowitz. So that might be something to keep an eye on, especially as the CEO of Hippocratic

98
00:08:49,480 --> 00:09:00,680
AI, Munjal Shah has said that their tool outperforms GPT-4 and Claude in over 100 healthcare certifications.

99
00:09:00,680 --> 00:09:06,600
So we've just talked about how good MedPalm 2 is, that would imply Hippocratic AI is even

100
00:09:06,600 --> 00:09:12,120
better. So that'll be interesting to see how that plays out. There is another company that

101
00:09:12,120 --> 00:09:18,400
raised a ton of cash recently, which is Everseen, who raised $71 million for its computer vision

102
00:09:18,400 --> 00:09:25,400
that prevents self-checkout theft. I'm sure we're all used now to being a number of places,

103
00:09:25,400 --> 00:09:28,640
for example, supermarkets where we go through the self-checkout, but they have this cool

104
00:09:28,640 --> 00:09:34,200
new computer vision tool to make sure that all the things that you're walking around

105
00:09:34,200 --> 00:09:39,740
with and that you actually pass them through the checkout, which is pretty interesting.

106
00:09:39,740 --> 00:09:40,740
So yeah.

107
00:09:40,740 --> 00:09:42,240
You can't be stealing those carrier bags anymore.

108
00:09:42,240 --> 00:09:43,960
Yeah, carrier bags, chewing gum.

109
00:09:43,960 --> 00:09:46,000
How many carrier bags did you take? None.

110
00:09:46,000 --> 00:09:47,000
Didn't use any.

111
00:09:47,000 --> 00:09:52,440
How many did you use today? Well, it depends, like how you're going to work it out. Oh,

112
00:09:52,440 --> 00:10:00,200
crumbs. No, you can't. So that's some short snippets from around the web this week. Let's

113
00:10:00,200 --> 00:10:07,400
jump into this first story then, Martin, because I think it's an interesting one. So a financial

114
00:10:07,400 --> 00:10:14,640
modeling SaaS startup called Causal successfully increased their SEO, their traffic from search

115
00:10:14,640 --> 00:10:26,840
to 750,000 visits a month using 100% AI generated content. So many, what feels like years ago,

116
00:10:26,840 --> 00:10:32,640
but it's probably mere weeks or maybe a month at best, we talked on the podcast, Martin,

117
00:10:32,640 --> 00:10:41,100
about generating content at scale with, with ChatGPT and other tools and with the quality

118
00:10:41,100 --> 00:10:46,480
of the content and would Google and other search engines basically penalize you for

119
00:10:46,480 --> 00:10:54,400
having AI generated content? And what this would imply is at least in this case, the

120
00:10:54,400 --> 00:11:01,200
answer is no, presumably because it's too hard for them to detect that it's AI generated

121
00:11:01,200 --> 00:11:06,620
or that not that much time has gone by for Google to actually identify another search

122
00:11:06,620 --> 00:11:13,800
engines and do something about it. The way that they did this is that they identified

123
00:11:13,800 --> 00:11:18,560
a bunch of topics with high potential keywords, chose six of them, and then created thousands

124
00:11:18,560 --> 00:11:28,560
of pages targeting those keywords at scale using a tool called ByWord. So that was their

125
00:11:28,560 --> 00:11:34,320
strategy. Then they put it all up on a website and did all the usual things to ensure that

126
00:11:34,320 --> 00:11:38,520
it was, the website was technically optimized for all the other things that Google and other

127
00:11:38,520 --> 00:11:45,960
search engines would expect in order to rank a website on search engine results pages.

128
00:11:45,960 --> 00:11:50,620
It's proven to be highly effective so far. And then the key question I think for us to

129
00:11:50,620 --> 00:11:59,720
debate here, Martin, is with the success of this approach by Causal, the fact that ByWord

130
00:11:59,720 --> 00:12:07,080
is a tool that anybody can now go and use to create AI content at scale, presumably,

131
00:12:07,080 --> 00:12:13,280
I'm going to put a massive asterisk on that even before we start our debate, is will there

132
00:12:13,280 --> 00:12:19,440
now be a race on for brands to try and generate as much content as possible and throw it up

133
00:12:19,440 --> 00:12:25,360
on their websites to see if they can reproduce what Causal did? So what's your thought on

134
00:12:25,360 --> 00:12:27,800
this story, Martin?

135
00:12:27,800 --> 00:12:34,100
I think the race started a while ago and this is just one of the first examples that we're

136
00:12:34,100 --> 00:12:40,160
likely to see in the coming months. From reading the case study, I think they chose a few interesting

137
00:12:40,160 --> 00:12:46,000
approaches. It looks like they did kind of glossary index pages that then linked out

138
00:12:46,000 --> 00:12:51,800
to subtopics and so on. So there was a bit of thought into how to structure the content

139
00:12:51,800 --> 00:12:59,320
at least. And I think if we go back to the question that you've brought, the idea that

140
00:12:59,320 --> 00:13:05,240
the search engines haven't yet discovered or found a way of identifying this as AI content,

141
00:13:05,240 --> 00:13:08,840
I think Google have been quite clear so far. So they don't really care as long as the content

142
00:13:08,840 --> 00:13:16,400
is helpful and relevant. So if it's passing some authority check and if we look at the

143
00:13:16,400 --> 00:13:24,200
kind of tools out there, say authority, kind of relevance or topic depth, they have their

144
00:13:24,200 --> 00:13:31,720
sentiment analysis and semantic frameworks that's, sorry, sentiment analysis is completely

145
00:13:31,720 --> 00:13:37,740
the wrong phrase there. That kind of semantic graph that says, okay, these are the topics

146
00:13:37,740 --> 00:13:41,680
if you want to rank for this thing, these are the topics you've got to be talking about

147
00:13:41,680 --> 00:13:48,640
to be an authority on this. Clearly, Causal have done their research there, hence identifying

148
00:13:48,640 --> 00:13:54,240
those topics in the first place. Is this going to work for other people? Absolutely. I think

149
00:13:54,240 --> 00:13:58,720
if you've got the resources to invest in this, and it sounds like you don't need particularly

150
00:13:58,720 --> 00:14:04,880
deep pockets and you just need a bit of time to run the AIs and do a bit of the research

151
00:14:04,880 --> 00:14:10,920
upfront and then post it into WordPress or wherever, you will see big results. Will it

152
00:14:10,920 --> 00:14:22,640
last in the long term? That's the question. It strikes me that this is a, there is a limited

153
00:14:22,640 --> 00:14:32,040
window to make hay here. And that I think comes from two observations. One is that everybody's

154
00:14:32,040 --> 00:14:42,080
going to be doing this, therefore Google will find other ways of ranking or finding signals.

155
00:14:42,080 --> 00:14:47,480
We know that Google has the EAT, expertise, authority, trustworthiness, but they've also

156
00:14:47,480 --> 00:14:53,960
included experience in that. So it's now experience, expertise, authority, and trustworthiness.

157
00:14:53,960 --> 00:15:00,120
And the experience is looking for real people that have real credibility in this space.

158
00:15:00,120 --> 00:15:04,760
So things like podcasts and videos and speaking events and published papers, these are all

159
00:15:04,760 --> 00:15:10,840
going to make a difference and you can't do that at scale. So that's going to become increasingly

160
00:15:10,840 --> 00:15:16,760
important. And the second thing is there's a general question about how important SEO

161
00:15:16,760 --> 00:15:26,640
or how impactful SEO is going to be with that demo that we saw at Google AI where quite

162
00:15:26,640 --> 00:15:33,340
literally the generative AI, a big green box on the Google search page just pushed away

163
00:15:33,340 --> 00:15:41,680
the organic search results as if to say, you're no longer relevant. That is going to cannibalise

164
00:15:41,680 --> 00:15:47,540
organic traffic for publishers. So sure, make hay while the sun is shining, but do not expect

165
00:15:47,540 --> 00:15:50,120
this to be a long-term sustainable strategy.

166
00:15:50,120 --> 00:15:56,520
Yeah, I'd agree. And I think we're back as well to just because you can produce a lot

167
00:15:56,520 --> 00:16:03,640
of content and get a load of traffic, even if it's in the short term, you need that content

168
00:16:03,640 --> 00:16:08,260
to be speaking about the genuine pains and challenges and interests of your customers.

169
00:16:08,260 --> 00:16:12,600
You need to be able to, at some point, position the value of your offering. You need to be

170
00:16:12,600 --> 00:16:18,200
able to do something with that traffic and you need that traffic to be relevant. So at

171
00:16:18,200 --> 00:16:22,440
the very least, it can probably do a job at the top of the funnel to drag people to your

172
00:16:22,440 --> 00:16:25,800
website. But by the way, if those people are not the right people for your business, then

173
00:16:25,800 --> 00:16:31,400
in essence, is there any point in doing it at all for most brands? Question mark. But

174
00:16:31,400 --> 00:16:37,240
I think we both know what we think about that, Mark, which is that I probably wouldn't bother.

175
00:16:37,240 --> 00:16:43,200
I think the other thing that's interesting about the tool is it's one of the more expensive

176
00:16:43,200 --> 00:16:50,240
content generators on the market. So if you wanted to really do this at scale with basically

177
00:16:50,240 --> 00:16:54,300
unlimited credits, you'd have to spend two and a half thousand dollars a month, which

178
00:16:54,300 --> 00:17:00,320
is a lot more than, you know, the Jaspers and the copy.ai's and the Go Charlies of the

179
00:17:00,320 --> 00:17:05,840
world. And one of the things that actually lists in its monthly plans as a feature is

180
00:17:05,840 --> 00:17:15,200
detector evasion. So it's clearly trying to bake this, you can get away with this Google

181
00:17:15,200 --> 00:17:17,880
won't know into its USP.

182
00:17:17,880 --> 00:17:26,440
Well, that annoys me because the detectors themselves are just trash at the moment anyway.

183
00:17:26,440 --> 00:17:33,880
So as a selling point, I mean, yeah, detector evasion. I think you're just speaking out

184
00:17:33,880 --> 00:17:38,400
to, well, to be fair, you know, every time I speak to people about AI content, one of

185
00:17:38,400 --> 00:17:42,620
the big questions is can it be detected by Google? And I think people are unaware of

186
00:17:42,620 --> 00:17:48,520
just how many false positives and false negatives come out of these detections. They're just

187
00:17:48,520 --> 00:17:52,920
totally inaccurate at the moment. Certainly not accurate enough for the search engines

188
00:17:52,920 --> 00:17:59,560
to be looking to actually deploy them into their systems at the moment. So, but yeah,

189
00:17:59,560 --> 00:18:04,120
good on them if that's how they want to position their products. I can't comment on this product

190
00:18:04,120 --> 00:18:08,880
at all. I haven't seen it. I haven't used it. But if this is working for that company,

191
00:18:08,880 --> 00:18:16,200
then I think to your point there, if this is driving traffic, great. You then just moving

192
00:18:16,200 --> 00:18:19,880
on to the next part of the process, how much of that are you actually converting? And that's

193
00:18:19,880 --> 00:18:25,440
the interesting part of the story. Great. If this is a SaaS company wanting to drive

194
00:18:25,440 --> 00:18:32,080
a lot of traffic, fantastic. But SaaS company is not a publisher that gets paid per impression.

195
00:18:32,080 --> 00:18:37,640
SaaS company gets paid on new subs. And that's the interesting part of the story for me.

196
00:18:37,640 --> 00:18:42,400
Yeah, which is the bit that we don't know, right? And it's certainly, dear listeners,

197
00:18:42,400 --> 00:18:47,680
we don't want this to come across as a byword. But actually that was never the intent of

198
00:18:47,680 --> 00:18:52,480
what we were intending from this discussion. It sounds like a powerful tool that at least

199
00:18:52,480 --> 00:18:59,360
one company has used to great effect. There's a gap in the market to do this type of thing.

200
00:18:59,360 --> 00:19:04,240
And I'm sure a number of other companies are doing something similar, just maybe not quite

201
00:19:04,240 --> 00:19:09,320
as sophisticatedly as perhaps byword enables. And again, I haven't played with it either,

202
00:19:09,320 --> 00:19:17,560
so I can't say for sure. I still can't help but feel, and my recommendation would be just

203
00:19:17,560 --> 00:19:21,520
because you can doesn't mean you should. And I personally, I wouldn't recommend to any

204
00:19:21,520 --> 00:19:30,340
of our clients to go down this route. I still think that the opportunity here is how can

205
00:19:30,340 --> 00:19:39,880
you create insightful content that can't be produced easily by AI, by a bot basically,

206
00:19:39,880 --> 00:19:43,240
because that's what's going to differentiate your brand and deliver genuinely interesting

207
00:19:43,240 --> 00:19:48,340
and valuable information to your audience, whether that's content that you create by

208
00:19:48,340 --> 00:19:53,740
interviewing your technical SMEs or working with your customers and talking about their

209
00:19:53,740 --> 00:19:57,380
use cases and applications of your products or whatever it may be, because that's kind

210
00:19:57,380 --> 00:20:05,400
of B2B. And then how do you leverage that content at scale would be a better question,

211
00:20:05,400 --> 00:20:12,200
I think, than how can I produce in a content farm style a load of content about a given

212
00:20:12,200 --> 00:20:17,240
topic but the bot wrote it all and doesn't necessarily know what's going to be interesting

213
00:20:17,240 --> 00:20:21,880
over the next one, two or three years. Because at least as it stands, I don't think any of

214
00:20:21,880 --> 00:20:28,280
those bots are going to the key industry conferences and hearing from the key opinion leaders speaking

215
00:20:28,280 --> 00:20:33,800
in the keynotes. So I still think that's where the magic is going to be in terms of brands

216
00:20:33,800 --> 00:20:37,880
being able to differentiate the content they produce. The other thing is doing cool things

217
00:20:37,880 --> 00:20:43,640
like industry reports, again, surveys and gathering insights about what's going on in

218
00:20:43,640 --> 00:20:47,360
your market that no one else can possibly know, certainly not bots, but probably none

219
00:20:47,360 --> 00:20:52,760
of your competitors either. And then again, figuring out after you've done the hard work

220
00:20:52,760 --> 00:20:57,760
of doing the survey and analysis or whatever it may be, how do you then promote that at

221
00:20:57,760 --> 00:21:02,400
scale and then I'm sure there are ways that you can use AI to write short summaries and

222
00:21:02,400 --> 00:21:07,680
create social media snippets and all these other things. What do you think, Martin?

223
00:21:07,680 --> 00:21:14,320
Yeah, exactly. The insight is where the value is going to be again, going back to the Google

224
00:21:14,320 --> 00:21:21,840
eat principle, real insight, real people, real expertise, real authority. That's ultimately

225
00:21:21,840 --> 00:21:26,320
where in a world of AI generated content, the winners will actually be created.

226
00:21:26,320 --> 00:21:34,120
Yes, I agree. Right. Let's move on to our next story. Let's talk about this AI Act amendment

227
00:21:34,120 --> 00:21:39,280
within the EU that's proposed, which should be clear, not passed. Tell us about this,

228
00:21:39,280 --> 00:21:40,280
Martin.

229
00:21:40,280 --> 00:21:44,800
Yeah. So at the end of last week, two European Parliament committees published the latest

230
00:21:44,800 --> 00:21:50,920
version of the EU AI Act, which we have discussed a few times in recent episodes. And there

231
00:21:50,920 --> 00:21:56,840
were some big amendments really, because they'd received a lot of feedback from the industry

232
00:21:56,840 --> 00:22:04,760
and from the community at large. And there were, I would say four main takeaways. There's

233
00:22:04,760 --> 00:22:08,880
a really good article that will pop in the show notes from Lexology, which has done top

234
00:22:08,880 --> 00:22:15,740
10 takeaways, which we'll make sure we share. But in terms of the, I think the key implications

235
00:22:15,740 --> 00:22:24,160
that business owners and marketers in the EU need to be aware of. First of all, that

236
00:22:24,160 --> 00:22:30,200
there are higher costs and obligations for companies developing or using AI in the EU.

237
00:22:30,200 --> 00:22:37,440
These obligations in terms of reporting and disclosures are going to be increased. And

238
00:22:37,440 --> 00:22:46,280
in fact, actually the penalties for non-compliance are actually greater than GDPR. So the GDPR

239
00:22:46,280 --> 00:22:53,240
data breach and non-compliance fines were up to 4% of global annual turnover. This was

240
00:22:53,240 --> 00:23:02,160
4%, sorry, this was 6% and I think it's being increased to 7%, I understand. So yeah, the

241
00:23:02,160 --> 00:23:06,880
costs are quite big. There's a new definition of AI in there. So the EU has adopted the

242
00:23:06,880 --> 00:23:14,480
OECD definition of AI, which is very much more focused on machine learning and deep

243
00:23:14,480 --> 00:23:21,600
learning networks. It was previously a somewhat overly broad range of software applications

244
00:23:21,600 --> 00:23:26,160
that were falling under the umbrella of AI, which software developers were not too pleased

245
00:23:26,160 --> 00:23:32,800
with as you can well imagine. One thing that is new is that they've really increased the

246
00:23:32,800 --> 00:23:39,400
rights for EU consumers. The rules as it were, or the proposed legislation didn't really

247
00:23:39,400 --> 00:23:45,680
talk about individuals and how individuals were affected. Whereas now in this amendment,

248
00:23:45,680 --> 00:23:50,240
individuals have the right to lodge complaints with supervisory authorities and to request

249
00:23:50,240 --> 00:23:59,160
explanations of decisions made by high-risk AI systems. That's a key part of this legislation

250
00:23:59,160 --> 00:24:06,000
as well. The EU is taking a risk-based approach to regulation. So they're basically saying,

251
00:24:06,000 --> 00:24:15,120
we are going to regulate AI systems based on the level of harm that they could cause.

252
00:24:15,120 --> 00:24:23,260
Things that are related to healthcare or policing, these are the things that they're really going

253
00:24:23,260 --> 00:24:33,400
to clamp down on. Foundation models such as GPT-4 and the big generative AI foundation

254
00:24:33,400 --> 00:24:43,080
models, these are being regulated, but they are not going to be classed as high-risk.

255
00:24:43,080 --> 00:24:47,720
So that's quite a positive thing for us to recognise is that they're not going to be

256
00:24:47,720 --> 00:24:54,920
classed as high-risk systems. However, this increased regulation around them does have

257
00:24:54,920 --> 00:25:03,480
some restrictions and requirements for the model trainers, the companies creating these

258
00:25:03,480 --> 00:25:09,760
models around data governments. They've got to have increased technical documentation

259
00:25:09,760 --> 00:25:14,680
and there are going to be quality management systems and requirements to disclose things

260
00:25:14,680 --> 00:25:20,600
like the amount of copyright information that was used in the training of the models. So

261
00:25:20,600 --> 00:25:27,440
the companies have to be much more aware of all of that in disclosing various parts of

262
00:25:27,440 --> 00:25:31,080
the model and the use cases and things like that.

263
00:25:31,080 --> 00:25:38,640
One other observation was that there are more obligations for deployers. Now this is different

264
00:25:38,640 --> 00:25:45,580
from the company that's creating an AI model. This is someone that is deploying a model.

265
00:25:45,580 --> 00:25:54,040
If you're building on top of a foundational model, this would, as I understand it, make

266
00:25:54,040 --> 00:25:55,600
you a deployer.

267
00:25:55,600 --> 00:26:01,560
Like a Jasper or a copy AI or a Go Charlie or anything that's using open AI as its back

268
00:26:01,560 --> 00:26:02,920
end basically.

269
00:26:02,920 --> 00:26:08,620
Yeah, exactly. And if you're deploying somebody else's model, so it's not just the foundation

270
00:26:08,620 --> 00:26:15,400
of models, any kind of AI model, you're building on AWS using one of their models, you are

271
00:26:15,400 --> 00:26:24,440
a deployer. And the deployers themselves now face increased compliance obligations. Now

272
00:26:24,440 --> 00:26:31,140
this includes conducting a fundamental rights impact assessment and providing certain information

273
00:26:31,140 --> 00:26:35,560
to individuals affected by high risk AI systems.

274
00:26:35,560 --> 00:26:42,360
Now you might be saying, well, what does that mean? What are fundamental rights impact assessments?

275
00:26:42,360 --> 00:26:50,440
Now in GDPR, there's a requirement to have a data impact assessment, which is all around,

276
00:26:50,440 --> 00:26:57,400
I probably haven't got the terminology right there, but if you're using people's data,

277
00:26:57,400 --> 00:27:01,400
you have to do an impact assessment about what would happen if there was a data breach

278
00:27:01,400 --> 00:27:06,240
and you have to give high risks and what are you doing to mitigate the risks in high risk

279
00:27:06,240 --> 00:27:12,000
situations and low risk situations? What have you got in place to cover those kinds of things?

280
00:27:12,000 --> 00:27:16,840
Now this is a fundamental rights. This isn't data rights. This is fundamental rights impact

281
00:27:16,840 --> 00:27:22,200
assessment. So what are fundamental rights? It's difficult to kind of really pin down.

282
00:27:22,200 --> 00:27:26,960
And there's a few EU papers that cover this topic. I've found one from December, 2022

283
00:27:26,960 --> 00:27:32,840
that goes into this and it identifies things like privacy rights. So rights to personal

284
00:27:32,840 --> 00:27:38,040
autonomy, the sanctity of the home, data protection does actually fall under that. Things like

285
00:27:38,040 --> 00:27:43,880
expressional rights. So press freedom, commercial expression, right to assembly, right to vote,

286
00:27:43,880 --> 00:27:51,280
procedural rights. So the right to access a court, the right to a fair trial, equality

287
00:27:51,280 --> 00:27:56,880
rights, which will be things like rights to equal application of the law to everyone to

288
00:27:56,880 --> 00:28:01,600
whom the law applies. And then socioeconomic rights. So equal access to healthcare, equal

289
00:28:01,600 --> 00:28:06,320
access to affordable housing, equal access to education, equal access to social benefits.

290
00:28:06,320 --> 00:28:14,280
These are fundamental rights as they are relevant in an AI context. So you as an organisation

291
00:28:14,280 --> 00:28:22,160
that is thinking of deploying any AI in your product has to be aware of that. And you have

292
00:28:22,160 --> 00:28:26,200
to think about, you're going to have to start thinking about a fundamental rights impact

293
00:28:26,200 --> 00:28:33,040
assessment. Because if your product might be used for insurance, let's say you're doing

294
00:28:33,040 --> 00:28:37,800
healthcare insurance and it's assessing risk, that's going to be a really, that's going

295
00:28:37,800 --> 00:28:47,400
to be quite a high risk deployment of AI. If you're using generative AI, if you're Jasper,

296
00:28:47,400 --> 00:28:54,240
for instance, and you've got your open AI and various other models powered, powering

297
00:28:54,240 --> 00:28:58,160
your product and your product helps people write better subject lines for email marketing

298
00:28:58,160 --> 00:29:02,920
and things like that, that's relatively low risk, isn't it? The fundamental rights impact

299
00:29:02,920 --> 00:29:06,440
is going to be very low. You're not probably going to have to report that, create an extensive

300
00:29:06,440 --> 00:29:13,360
report on that. So yeah, I think that's a big consideration for business owners and

301
00:29:13,360 --> 00:29:24,040
marketers building with AI. Generally speaking, I think this is a good thing. This is the

302
00:29:24,040 --> 00:29:29,760
world's first piece of AI regulation. And as you mentioned, Sam Altman has been at the

303
00:29:29,760 --> 00:29:38,800
Senate this week, speaking to the senators in the US about AI regulation. The EU is now

304
00:29:38,800 --> 00:29:42,280
actually kind of leading the way. What are your thoughts on this?

305
00:29:42,280 --> 00:29:46,640
Yeah, it's an interesting one. Firstly, I'm glad I have you here, Martin, because when

306
00:29:46,640 --> 00:29:52,560
these changes are made, my brain struggles with all of the ins and outs of this information

307
00:29:52,560 --> 00:29:57,080
needs. So hopefully this, I'm sure it's very useful for all our listeners, but it's even

308
00:29:57,080 --> 00:30:02,520
just useful for me to hear you describe it, because then I understand it better. I think

309
00:30:02,520 --> 00:30:12,600
as a marketer, I think, A, I can see a lot of products and services getting AI augmented.

310
00:30:12,600 --> 00:30:18,600
And so product managers and marketers need to be aware of what's going on inside their

311
00:30:18,600 --> 00:30:23,200
products. Even if it's for things like crisis comms, right? If you're in a comms team and

312
00:30:23,200 --> 00:30:27,520
you need to have a plan in place for how you're going to deal with any issues that come up

313
00:30:27,520 --> 00:30:37,760
with how your AI driven tool is now no longer meeting the requirements of this legislation,

314
00:30:37,760 --> 00:30:43,400
assuming it goes through. I think more broadly, if I'm a marketer, one of the key questions

315
00:30:43,400 --> 00:30:49,200
I'm asking myself and one of the key questions I field is what tools should we use? How should

316
00:30:49,200 --> 00:30:53,360
we roll them out? That type of thing. I think it's a really difficult question to answer.

317
00:30:53,360 --> 00:30:58,640
We've talked about it many times on the podcast and it's a temptation to wait for big companies

318
00:30:58,640 --> 00:31:04,840
to just roll out broad tools like Google in its workplace. And it's perhaps like Google

319
00:31:04,840 --> 00:31:10,240
docs and Gmail, et cetera, and Microsoft copilot, just because there's a sort of trust you have

320
00:31:10,240 --> 00:31:14,960
in those organizations that it's going to make it easy to roll out and train. But if

321
00:31:14,960 --> 00:31:21,040
not, you're investing in tools that you have to embed in your organization. You have to

322
00:31:21,040 --> 00:31:24,920
make them a part of your processes. You have to train your team on them. And then what

323
00:31:24,920 --> 00:31:29,560
if this legislation impacts one of your supplies in a way that means that they can no longer

324
00:31:29,560 --> 00:31:35,000
provide the tool or they can, but the costs of meeting the requirements of this and other

325
00:31:35,000 --> 00:31:42,320
legislation quadruples the cost of it, right? Or the functionality changes because there's

326
00:31:42,320 --> 00:31:45,920
certain aspects of the tool that they just can't even offer anymore until this stuff

327
00:31:45,920 --> 00:31:50,720
is figured out. So there's a bit of a wait and see mindset that I see a lot of people

328
00:31:50,720 --> 00:31:54,880
starting to take in terms of crumbs. I want to start rolling some of these tools out in

329
00:31:54,880 --> 00:31:59,920
my organization, but what if some of those things happen? And of course it's legislation

330
00:31:59,920 --> 00:32:05,960
like these that could kickstart shutdowns and pauses and functionality changes and et

331
00:32:05,960 --> 00:32:06,960
cetera.

332
00:32:06,960 --> 00:32:13,480
So we're already seeing the EU's moves in this domain have an impact, right? So Bard,

333
00:32:13,480 --> 00:32:17,960
for instance, has different functionality in the rest of the world than it does in the

334
00:32:17,960 --> 00:32:29,520
EU. And it has been speculated that this is a bit of a, you know, FU to the EU from Google

335
00:32:29,520 --> 00:32:33,680
to say that, you know, we're bringing out this cool new tech and you can't have it because

336
00:32:33,680 --> 00:32:37,720
we can't make it comply with your products and you know, you're going to be left wanting

337
00:32:37,720 --> 00:32:42,480
whether that is the case or not, who knows? There, you know, there is functionality, the

338
00:32:42,480 --> 00:32:46,600
plugins that you get in Bard in the US and the rest of the world are not available to

339
00:32:46,600 --> 00:32:48,080
people in the EU at the moment.

340
00:32:48,080 --> 00:32:53,440
Just for clarity, the plugins in Bard or the plugins in ChatGPT?

341
00:32:53,440 --> 00:32:54,440
Plugins in Bard.

342
00:32:54,440 --> 00:32:55,440
Plugins in Bard.

343
00:32:55,440 --> 00:32:56,440
Yeah, the extensions in Bard. Yeah.

344
00:32:56,440 --> 00:32:58,560
Cool. I know that there's been...

345
00:32:58,560 --> 00:33:04,720
I say plugins, the extensions, the additional features, the extended capabilities.

346
00:33:04,720 --> 00:33:11,160
And you just can't get them. Yeah, yeah. I'd noticed that as well and saw some stories.

347
00:33:11,160 --> 00:33:16,600
Attached to this as well is just thinking about as a marketer or as a business owner,

348
00:33:16,600 --> 00:33:21,360
as anybody working in this space, one thing to be aware of is they have reduced the time

349
00:33:21,360 --> 00:33:26,840
for compliance. There was originally a three year time scale to implement this, which companies

350
00:33:26,840 --> 00:33:32,640
were saying was already too short and they've reduced it from three years to two years.

351
00:33:32,640 --> 00:33:37,200
So if it was too short and you were whinging, better spend less time whinging and more time

352
00:33:37,200 --> 00:33:40,000
acting because that timeline is ticking.

353
00:33:40,000 --> 00:33:46,520
Oh. Although what the world would look like in three or two years, given the pace of change,

354
00:33:46,520 --> 00:33:48,920
is another interesting question, I would say.

355
00:33:48,920 --> 00:33:52,440
It is. There's one additional point to this, though, that I just want to throw in. There

356
00:33:52,440 --> 00:33:58,360
was a story in The Guardian today. So Prime Minister Rishi Sunak is out in Japan doing

357
00:33:58,360 --> 00:34:07,360
his tour, doing a media press tour while he prepares for the G7 meeting. And when asked

358
00:34:07,360 --> 00:34:16,480
about AI, he said that the UK will lead on guardrails to limit the dangers of AI, which

359
00:34:16,480 --> 00:34:21,000
is a stark contrast to the UK government's position in March, where they said they were

360
00:34:21,000 --> 00:34:29,240
going to take a laissez faire approach to AI regulation, enable to encourage and foster

361
00:34:29,240 --> 00:34:36,320
innovation and whatnot. But it turns out when a letter is written by senior people in the

362
00:34:36,320 --> 00:34:42,080
world of AI saying, well, we need to consider this, the UK government suddenly sits up and

363
00:34:42,080 --> 00:34:46,400
listens. I would suggest they probably don't know their arse from their elbow in this area,

364
00:34:46,400 --> 00:34:48,200
but what do I know?

365
00:34:48,200 --> 00:34:56,160
They certainly seem a little confused, depending on who you ask, I guess. But I'm sure we'll

366
00:34:56,160 --> 00:35:03,120
see how that plays out over time. It has a bit of an in the thick of it, yes, Minister,

367
00:35:03,120 --> 00:35:10,760
feel that quote. For those of you that are not familiar with those very British sitcoms,

368
00:35:10,760 --> 00:35:16,360
they lampoon the inner workings of the UK government and in effect make it look like

369
00:35:16,360 --> 00:35:19,560
no one in the government knows what they're doing, which again is not what we're saying

370
00:35:19,560 --> 00:35:25,360
here, but certainly an interesting parallel. Right. Thank you for that, Martin. That's

371
00:35:25,360 --> 00:35:29,760
super helpful. Right. Last detailed story of the week is that we're going to have a

372
00:35:29,760 --> 00:35:35,400
brief discussion about chat GPT plugins. So we've talked quite a lot about these over

373
00:35:35,400 --> 00:35:41,680
the last couple of weeks, maybe month or two since they were announced, but it was pretty

374
00:35:41,680 --> 00:35:47,400
hard to get access to them. But this week, the lovely folks at OpenAI have been rolling

375
00:35:47,400 --> 00:35:55,480
them out to all of their users. Martin got access before me, which just annoyed me immensely

376
00:35:55,480 --> 00:36:01,320
and I didn't get out of bed for two days, but I also have access now, so I am joyous

377
00:36:01,320 --> 00:36:07,080
again. If you are interested in having a play with these, you can access them if you have

378
00:36:07,080 --> 00:36:11,640
chat GPT. I'm guessing you have to have plus. We've both got plus, Martin. Yeah. It does

379
00:36:11,640 --> 00:36:15,520
require plus. Yeah. Yeah. I figured it might. So if you've got plus, you can go into your

380
00:36:15,520 --> 00:36:20,400
settings and you'll see some new options in there that allow you to turn on the plugins

381
00:36:20,400 --> 00:36:26,880
beta and then you'll be able to start leveraging the plugins when you're searching. It changes

382
00:36:26,880 --> 00:36:31,880
the way that you select the model you're using at the top and that's where you can add the

383
00:36:31,880 --> 00:36:38,560
plugins. There are, I think, nearly a hundred from my quick scan through of plugins with

384
00:36:38,560 --> 00:36:45,680
lots of different use cases at this point that you can have a play with. We did see

385
00:36:45,680 --> 00:36:53,080
a really interesting article from the folks over at AI Marketing School where they actually

386
00:36:53,080 --> 00:37:01,240
took us through three really cool use cases using chat GPT plugins for marketing. So just

387
00:37:01,240 --> 00:37:05,960
to briefly describe what they talked about in their newsletter, which is well worth signing

388
00:37:05,960 --> 00:37:13,400
up to, they used the plugin called Video Insights to be able to turn videos into articles. So

389
00:37:13,400 --> 00:37:19,320
you have to install the Video Insights plugin and then you can go to YouTube. You can search

390
00:37:19,320 --> 00:37:25,440
for videos on a particular topic and then you can paste the URLs of those videos and

391
00:37:25,440 --> 00:37:30,960
it doesn't have to be one. It could be multiple ones and then ask chat GPT to write a blog

392
00:37:30,960 --> 00:37:37,960
post or an article or some sort of summary based on what the tool was able to pull out

393
00:37:37,960 --> 00:37:45,560
of those videos. And as the team say, it's not perfect, but it can give you a pretty

394
00:37:45,560 --> 00:37:51,600
decent first draft and some ideas they suggest things like turning a webinar into articles,

395
00:37:51,600 --> 00:37:57,200
social media posts and those types of things. They make a note and we would massively second

396
00:37:57,200 --> 00:38:02,280
this. I think this is true of anything you do with Genitive AI. Double check the outputs,

397
00:38:02,280 --> 00:38:05,800
make sure it hasn't hallucinated in the middle of what I was doing, make sure it hasn't pulled

398
00:38:05,800 --> 00:38:10,000
in some materials from elsewhere that you didn't want it to pull in and all that good

399
00:38:10,000 --> 00:38:15,360
stuff. The second use case they talk about, which is pretty cool, is connecting it with

400
00:38:15,360 --> 00:38:22,360
Zapier, Zapier, which basically allows you to connect chat GPT to any other app. So it's

401
00:38:22,360 --> 00:38:28,760
a bit like for Zapier users, Zapier is the great connector, right? It allows you to have

402
00:38:28,760 --> 00:38:35,040
a piece of data that's in one app, trigger an action in another app. So an example might

403
00:38:35,040 --> 00:38:46,320
be when a new contact appears in my CRM, add their details to a row in a Google Sheet as

404
00:38:46,320 --> 00:38:52,680
an example. But now with chat GPT, you can access Zapier's capabilities within the interface.

405
00:38:52,680 --> 00:38:57,320
So as they say in the newsletter, you can create emails or Slack messages and send them

406
00:38:57,320 --> 00:39:03,740
straight without leaving the interface. You can send data directly from chat GPT to other

407
00:39:03,740 --> 00:39:09,200
tools like Google Sheets and Notion and those types of things. And you can push things like

408
00:39:09,200 --> 00:39:14,720
if you're doing a brainstorm and asking chat GPT for headline ideas or article outlines,

409
00:39:14,720 --> 00:39:19,520
those types of things, you can push them straight into a Google Doc. So that's pretty interesting.

410
00:39:19,520 --> 00:39:24,800
I think people will have to figure out where that really saves them time versus just like

411
00:39:24,800 --> 00:39:29,800
copy pasting stuff over themselves. But I think it's pretty interesting one. And then

412
00:39:29,800 --> 00:39:33,440
the last one that I think I probably like the most and definitely want to go and have

413
00:39:33,440 --> 00:39:39,440
a play with is the WebPilot plugin. So in this, they give an example of how you could

414
00:39:39,440 --> 00:39:47,360
use it to SEO, optimize your content by in essence, getting it to create an SEO strategy

415
00:39:47,360 --> 00:39:54,200
to outrank your competitors. So in the example that they gave, they asked it to scrape the

416
00:39:54,200 --> 00:40:01,520
top five search engine results for the keyword using AI for marketing and give me a detailed

417
00:40:01,520 --> 00:40:08,720
SEO strategy to help me outcompete them. For example, include exact keyword density, comprehensive

418
00:40:08,720 --> 00:40:14,520
h2 outline for an article that contains everything these articles touch upon, plus anything that

419
00:40:14,520 --> 00:40:18,640
you feel is missing that they should have touched upon that will help me rank highly

420
00:40:18,640 --> 00:40:23,280
for this keyword. And it's too long and laborious for me to read the output to you. But it's

421
00:40:23,280 --> 00:40:29,160
pretty darn impressive in terms of touching on the suggested keyword density, suggesting

422
00:40:29,160 --> 00:40:35,320
a really effective outline for the article, highlighting a few gaps that the top ranking

423
00:40:35,320 --> 00:40:40,440
pages don't focus on. And when I was seeing this example, it reminded me of when you took

424
00:40:40,440 --> 00:40:47,640
us through tool of the week, a month or two ago, Market Muse, Martin, and it's for obvious

425
00:40:47,640 --> 00:40:53,000
reasons, nowhere near as usable and sophisticated as Market Muse. But as a tool, you can, in

426
00:40:53,000 --> 00:41:01,880
essence, as I understand it, use for free within the ChatGPT interface, although we

427
00:41:01,880 --> 00:41:07,560
should check because you might have to have a login for WebPilot as well, I can imagine.

428
00:41:07,560 --> 00:41:08,560
But we can look into that.

429
00:41:08,560 --> 00:41:14,000
I've tried that plugin and there's no restriction on it as far as I can see.

430
00:41:14,000 --> 00:41:21,080
Lovely stuff. So three pretty cool applications there of plugins and many more to come, I'm

431
00:41:21,080 --> 00:41:25,920
sure. So go and have a play with those. And if you come up with anything else, cool. Of

432
00:41:25,920 --> 00:41:30,520
course, we would love to hear about it. We've also been playing with it a bit ourselves.

433
00:41:30,520 --> 00:41:33,400
You've been playing with it, how many, Martin, what have you been up to?

434
00:41:33,400 --> 00:41:39,880
So there was, as well as the actual plugins, they also announced this week the web browsing

435
00:41:39,880 --> 00:41:47,680
capability in beta itself. So whereas there is WebPilot, which is a plugin, ChatGPT now

436
00:41:47,680 --> 00:41:51,960
has its own web browsing capability that isn't a plugin, but does need turning on in the

437
00:41:51,960 --> 00:42:00,200
same settings section. So I thought I would, I don't know if I've told you, but Derby County

438
00:42:00,200 --> 00:42:05,680
got knocked out of, well, didn't get into the playoffs this year. So our season ended

439
00:42:05,680 --> 00:42:12,280
early. And as such, I'm left looking for Derby County news that is, you know, non-match day

440
00:42:12,280 --> 00:42:18,400
related, things like player transfers and such like that. So typically trying to find

441
00:42:18,400 --> 00:42:23,480
any news around that requires crawling the forums, browsing Twitter. And I thought, I

442
00:42:23,480 --> 00:42:27,840
can't be bothered with any of that. So instead I will ask the web browsing functionality

443
00:42:27,840 --> 00:42:37,080
on ChatGPT to tell me, gather any latest transfer rumours for my beloved Derby County. And it

444
00:42:37,080 --> 00:42:43,360
did exactly that. It went out and crawled the web and brought back a really good summary

445
00:42:43,360 --> 00:42:47,240
actually. So it said Derby County has been linked with several players recently, and

446
00:42:47,240 --> 00:42:52,120
then it gives me one, two, three, four, five players that we've been linked with. And with

447
00:42:52,120 --> 00:42:59,920
each one, it tells me who they are and a kind of summary of the position and the news around

448
00:42:59,920 --> 00:43:06,440
them. Now, the interesting thing was with this, each one of those stories, and you get

449
00:43:06,440 --> 00:43:13,040
this with the web browsing model of ChatGPT, but not with WebPilot. So it actually links

450
00:43:13,040 --> 00:43:20,760
you to the story. With WebPilot, I don't remember actually having links to the stories.

451
00:43:20,760 --> 00:43:26,040
So I read these stories, I read this and then clicked on one of the links and it took me

452
00:43:26,040 --> 00:43:31,200
through to this football transfer rumour website. And actually I've posted a screenshot of it

453
00:43:31,200 --> 00:43:36,480
on Twitter, if you want to check it out. And the interface on this website was atrocious.

454
00:43:36,480 --> 00:43:42,080
It was, there were so many adverts. I was being pelted with adverts left, right and

455
00:43:42,080 --> 00:43:47,160
centre. There is literally on my screen about four lines of copy and the rest of the screen

456
00:43:47,160 --> 00:43:52,520
is adverts overlaid with adverts with more adverts overlaid on those. It's an absolute

457
00:43:52,520 --> 00:43:59,600
mess of a user experience. Yet I avoided all of that using this web browsing feature that

458
00:43:59,600 --> 00:44:05,280
just went on, did the crawling for me and brought the results back. And it made me think,

459
00:44:05,280 --> 00:44:08,400
oh, that's going to impact a lot of publishers.

460
00:44:08,400 --> 00:44:14,760
I mean, there's so many great things in that. Firstly, I get so many messages, people ask

461
00:44:14,760 --> 00:44:19,280
me what's going on with Derby County. Finally, they have the information that they need.

462
00:44:19,280 --> 00:44:23,480
So thank you, Martin. If you looked at our subscriber base, about half of them are Derby

463
00:44:23,480 --> 00:44:29,800
County fans and then the other half are marketers interested in AI. The other thing is I suspect

464
00:44:29,800 --> 00:44:36,400
these tools may still be somewhat prone to hallucination. I think it's unlikely that

465
00:44:36,400 --> 00:44:43,120
you're going to be signing Messi and Ronaldo. I know that's what it said in the output mind,

466
00:44:43,120 --> 00:44:47,120
but you should definitely double check the sources on those. And yes, as we've talked

467
00:44:47,120 --> 00:44:54,520
about before, how will people access information on the web as these tools become the mainstay

468
00:44:54,520 --> 00:45:00,240
of how we do things on our computers and on our phones and I guess eventually in VR and

469
00:45:00,240 --> 00:45:06,200
AR. And yeah, if you can just get the information easier and you don't have to fight pop up

470
00:45:06,200 --> 00:45:11,920
central you do as well. Right. So yeah, absolutely fascinating. I think one of the things I've

471
00:45:11,920 --> 00:45:16,880
noticed from playing with the tool and I found myself sort of replicating some of the things

472
00:45:16,880 --> 00:45:21,600
you talked about a couple of weeks ago, Martin. So I've made the move to edge now and I'm

473
00:45:21,600 --> 00:45:27,280
in my sidebar. I've got Bing, I've got Track GPT and I've got Bard and I can basically

474
00:45:27,280 --> 00:45:34,040
very quickly run the same query through all three tools. And whilst I love the power that

475
00:45:34,040 --> 00:45:40,860
you get with the plugins, my goodness, it is slow. It is significantly slower than the

476
00:45:40,860 --> 00:45:46,720
other tools. And so when I need an answer quickly, I'm going to Bing or to Bard. And

477
00:45:46,720 --> 00:45:51,040
if I want to do something where I feel like the request is kind of a bit complicated or

478
00:45:51,040 --> 00:45:58,360
taps into one of those more power, not more powerful, but more non browser based plugins,

479
00:45:58,360 --> 00:46:05,000
then I'll push it through Chat GPT. So this may be the first indicator at least for me

480
00:46:05,000 --> 00:46:12,840
of how as these tools mature, a lot of the things that underpin good UX as users, we're

481
00:46:12,840 --> 00:46:16,920
going to probably have to be giving them a little bit less license, right? Like if it's

482
00:46:16,920 --> 00:46:21,480
slow, like how when web pages that you load are slow, you're like, right, I haven't got

483
00:46:21,480 --> 00:46:26,880
eight seconds to wait for this to load. I'm going to go just read a different source instead.

484
00:46:26,880 --> 00:46:33,440
And we may just see that start to differentiate which tool you go to maybe in general or for

485
00:46:33,440 --> 00:46:38,000
different use cases like I've described. The other thing is I think it will highlight some

486
00:46:38,000 --> 00:46:41,440
of the interesting and powerful work that's happening in the open source community at

487
00:46:41,440 --> 00:46:47,200
the moment, trying to figure out how do you create a large language model that can deliver

488
00:46:47,200 --> 00:46:53,640
quality outputs at GPT four level, but with a much smaller model trained with much fewer

489
00:46:53,640 --> 00:46:58,000
parameters and much simpler weightings and all those other things, the likes of which

490
00:46:58,000 --> 00:47:03,400
would be fast. And as we've talked about previously, fast enough to run natively on a phone or

491
00:47:03,400 --> 00:47:09,920
even on as some people have shown on something like Raspberry Pi. And that need for speed

492
00:47:09,920 --> 00:47:13,960
is actually probably not just going to be something the open source community thinks

493
00:47:13,960 --> 00:47:17,680
is a cool experiment. And I'm sure they're not the only people who think that obviously,

494
00:47:17,680 --> 00:47:23,280
but actually could end up being quite a commercial driver in terms of reducing the costs and

495
00:47:23,280 --> 00:47:27,840
computation required to offer these tools, but just making a great fast user experience.

496
00:47:27,840 --> 00:47:34,320
Yeah. What's this space for when Apple launches? Indeed. I can't wait because the conference

497
00:47:34,320 --> 00:47:40,800
is in June and I want to see this VR headset they've been toiling away on to see if it's

498
00:47:40,800 --> 00:47:45,600
the car crash that some stories will tell you that everybody inside Apple thinks this

499
00:47:45,600 --> 00:47:50,240
is a terrible idea and we should never launch it or a few of the stories that are popping

500
00:47:50,240 --> 00:48:00,760
up that are suggesting that it might actually be quite cool, including the founder of Oculus

501
00:48:00,760 --> 00:48:05,600
who went on Twitter and basically implied it's absolutely awesome. So, oh, can't wait

502
00:48:05,600 --> 00:48:12,000
to see. Anyway, I digress and we need to respect our lovely list this time. So we'll move on.

503
00:48:12,000 --> 00:48:18,360
Last couple of bits and pieces then everyone. First of all, we are going to hear, for all

504
00:48:18,360 --> 00:48:23,640
of us, as I talked about at the beginning, Liam Lally, who is a PPC expert. For those

505
00:48:23,640 --> 00:48:28,600
of you who haven't listened to last week's episode yet, we were talking, Martin and I,

506
00:48:28,600 --> 00:48:37,880
about Meta releasing some AI enabled tools to help advertisers get better results and

507
00:48:37,880 --> 00:48:41,640
create new ad formats automatically and all this type of stuff. And that got us into a

508
00:48:41,640 --> 00:48:48,520
discussion about the pros and cons of using automation in your pay-per-click advertising

509
00:48:48,520 --> 00:48:54,960
strategy. And the different perspectives that people have in the market, especially PPC

510
00:48:54,960 --> 00:49:02,720
specialists about the power of AI versus automation versus actually those tools just being something

511
00:49:02,720 --> 00:49:08,560
that Google and Meta can use to get you to spend more money, basically. So we had a really

512
00:49:08,560 --> 00:49:12,760
good debate about that. And then we invited other experts who probably know more about

513
00:49:12,760 --> 00:49:21,120
it than we do to share their thoughts. And that is what Liam did. So we're going to go

514
00:49:21,120 --> 00:49:25,760
to Liam in a minute. I think Martin and I were left with the idea that automation could

515
00:49:25,760 --> 00:49:32,100
be powerful, but only in the hands of an expert and the human in the loop and someone managing

516
00:49:32,100 --> 00:49:37,360
this process was critical. And then I think Liam's done a really fantastic job, a job

517
00:49:37,360 --> 00:49:43,460
of giving us some deeper insights there and really giving us a feel for where he feels

518
00:49:43,460 --> 00:49:47,360
automation sits and what to watch out for. So over to you, Liam.

519
00:49:47,360 --> 00:49:56,040
Hi, it's Liam from Saddle. I'm a PPC manager. I was listening to your discussion about the

520
00:49:56,040 --> 00:50:03,840
algorithmic ad managers. I think you were talking about Firms Max and Advantage Plus.

521
00:50:03,840 --> 00:50:07,680
And I thought I would just chip in because I've got quite a few thoughts on this one.

522
00:50:07,680 --> 00:50:13,320
So with regards to Performance Max, I think it was, I can remember it wasn't the thing

523
00:50:13,320 --> 00:50:19,920
that they mentioned. They thought it was the future of advertising for some businesses.

524
00:50:19,920 --> 00:50:25,120
One of the issues that some of us have got that have been doing this is that there is

525
00:50:25,120 --> 00:50:31,520
no denying that sometimes we turn a shopping campaign into a Performance Max campaign in

526
00:50:31,520 --> 00:50:36,560
the same way as previously we would have gone from a shopping campaign to what used to be

527
00:50:36,560 --> 00:50:43,080
called a smart shopping campaign. And they will perform, well, not all of them, but I

528
00:50:43,080 --> 00:50:50,360
think my caveat with some of this is the data, you have to have a huge amount of data for

529
00:50:50,360 --> 00:50:54,200
this to work out really well. And I think some of the people that kind of get on and

530
00:50:54,200 --> 00:50:57,880
say this is what you should be doing all the time and you should be, or to apply in all

531
00:50:57,880 --> 00:51:02,840
recommendations given to you by all of these platforms. I think if you're a huge brand

532
00:51:02,840 --> 00:51:07,720
with huge budgets, I think you can get away with it because I think that that data is

533
00:51:07,720 --> 00:51:14,520
collected quicker. I think for SMEs and smaller businesses, I think that there has to be more

534
00:51:14,520 --> 00:51:20,800
hesitation over that. And I'll come on to that in a second. So for example, if we put

535
00:51:20,800 --> 00:51:28,320
some Performance Max campaigns together for our clients, then we have, and this is coming

536
00:51:28,320 --> 00:51:35,280
in, so in theory we have no control over things like branded searches. So people physically

537
00:51:35,280 --> 00:51:39,600
looking for the client anyway, shopping app comes up, they click on that, they go through

538
00:51:39,600 --> 00:51:45,700
and they buy. Or if this is Performance Max, then clicking on a display ad or a video clip

539
00:51:45,700 --> 00:51:52,560
or whatever it is where they've decided to put it. Now you can, or you're beginning to

540
00:51:52,560 --> 00:51:56,440
have the ability to be able to exclude branded searches. So it'll be interesting to see how

541
00:51:56,440 --> 00:52:00,600
Performance Max changes, but just to put it into context, I just want to give you some

542
00:52:00,600 --> 00:52:06,400
figures because I work on figures. So if I said we have a client that's just spent 22

543
00:52:06,400 --> 00:52:14,840
grand in the last 30 days and they've had a conversion value of 1,054, 1,000, a million,

544
00:52:14,840 --> 00:52:21,680
1,000,000 pounds and 54,966 pounds worth of revenue. That sounds amazing. That sounds

545
00:52:21,680 --> 00:52:27,640
like it is amazing. It's fantastic. If I then turn around and say they've got some branded

546
00:52:27,640 --> 00:52:33,440
search ads that they can then monitor and make sure what's going on with them. And within

547
00:52:33,440 --> 00:52:41,320
all of those totals that I've just given you, they've spent 185 pound on their branded ads

548
00:52:41,320 --> 00:52:50,160
and they've had 243,000 pounds worth of sales. So my issue with Performance Max isn't necessarily

549
00:52:50,160 --> 00:52:55,520
the AI and all the other stuff around it. It's the fact that so much is hidden and we're

550
00:52:55,520 --> 00:53:00,560
not able to actually see what impact those, because if you think of some of those other

551
00:53:00,560 --> 00:53:08,960
campaigns, we've then got every single campaign could be looking like it's amazing because

552
00:53:08,960 --> 00:53:13,880
it's actually triggering for branded. And when you look at those kinds of figures, then

553
00:53:13,880 --> 00:53:21,460
it's very easy to see how the final results can be warped. My final point on machine learning

554
00:53:21,460 --> 00:53:28,640
and AI and things like that in Google, often people are, I don't want to say recommended,

555
00:53:28,640 --> 00:53:34,480
but they're suggested to them that they will turn on Google's auto apply recommendations.

556
00:53:34,480 --> 00:53:40,840
And when I have seen companies do this, generally everything goes south. Now, again, I think

557
00:53:40,840 --> 00:53:45,360
some of this is down to huge amounts of data. I think some of this could be down to set

558
00:53:45,360 --> 00:53:53,200
up a little bit, but actually if I look at some of the campaigns, we've run across multiple

559
00:53:53,200 --> 00:54:01,120
clients. So we're talking 100, 100 plus clients. When we've implemented some of the automated

560
00:54:01,120 --> 00:54:05,120
smart bidding as they would call it, it's done the polar opposite of what you wanted

561
00:54:05,120 --> 00:54:12,360
it to do. And the response that we've got back has been, oh, you should let it run for

562
00:54:12,360 --> 00:54:16,760
longer. And I think one of the issues for a lot of SMEs is we don't have budgets to

563
00:54:16,760 --> 00:54:22,080
be able to run something for six months to see if it works. It tends to be far shorter

564
00:54:22,080 --> 00:54:31,040
periods. So does performance max work? Yes, certainly looks like it works in some cases.

565
00:54:31,040 --> 00:54:37,080
I've also had clients where it hasn't worked at all. It seems to get, it seems to go downhill

566
00:54:37,080 --> 00:54:43,520
very quickly when it hasn't got enough conversions coming through. So I've seen campaigns literally

567
00:54:43,520 --> 00:54:50,040
fall over cliff and take three months to get back online. With it all, I'm worried about

568
00:54:50,040 --> 00:54:56,360
the transparency of it. And therefore we've always had the transparency. And I think when

569
00:54:56,360 --> 00:55:03,640
you look at P max for Google, if you look at advantage plus to an extent for Facebook,

570
00:55:03,640 --> 00:55:07,640
then we'll get into a point where we're getting less and less visibility and with less and

571
00:55:07,640 --> 00:55:13,040
less visibility, that worries me. So a really big thing. Thanks there to you,

572
00:55:13,040 --> 00:55:17,880
Liam, for sharing that with us. If you're a listener to the podcast and we touch on

573
00:55:17,880 --> 00:55:21,840
a topic that you're an expert in or something that you're passionate about and you want

574
00:55:21,840 --> 00:55:28,440
to come on, please let us know. We'd love to have you on. You can come in and join our

575
00:55:28,440 --> 00:55:32,680
conversation like me and Martin are having today. You can record an audio snippet and

576
00:55:32,680 --> 00:55:37,360
send it to us. However you want to do it. We would love to hear from you.

577
00:55:37,360 --> 00:55:41,180
Right. Last little bit. We're going to look at tool of the week and tool of the week this

578
00:55:41,180 --> 00:55:47,560
week is Jenny and Martin, you've been having a play with this voice generation tool. Tell

579
00:55:47,560 --> 00:55:57,680
us what you've learned. So Jenny is a tool from a company called Lovoo, L-O-V-O, Lovoo.ai

580
00:55:57,680 --> 00:56:04,440
and it is, as you've suggested, a voice generation tool, hence the name Jenny. Jenny is in generative,

581
00:56:04,440 --> 00:56:11,120
so it's G-E-N-N-Y and it's a really interesting interface. When you see it for the first time

582
00:56:11,120 --> 00:56:17,480
you go, wow, this is a proper audio editing system. Now the way that it works is that

583
00:56:17,480 --> 00:56:24,480
you can input your text from a script, it breaks it down into paragraphs for you and

584
00:56:24,480 --> 00:56:31,960
then you can go through your script, selecting each section and assigning one of the voices

585
00:56:31,960 --> 00:56:40,740
to it. And there are many, many voices to choose from, more than I care to count actually.

586
00:56:40,740 --> 00:56:45,160
Then you've got voices from all over the world and they're tagged by things like English

587
00:56:45,160 --> 00:56:53,120
American, English GB, then they've got things like young adult, education, entertainment,

588
00:56:53,120 --> 00:57:00,080
marketing, all these different use cases are in the mix. If you want something from a global

589
00:57:00,080 --> 00:57:04,480
or if you want a global voice, international voice, you've got that as well. You can actually

590
00:57:04,480 --> 00:57:12,160
look at different languages, everything from Afrikaans right down to Welsh and Zulu and

591
00:57:12,160 --> 00:57:16,640
Uzbek and everything in between. You can choose the age range of the speaker. So it's got

592
00:57:16,640 --> 00:57:23,400
a lot of customization in there. In fact, you can even customize things like pronunciation.

593
00:57:23,400 --> 00:57:28,360
So if there's a word that is written in your script, it just sounds weird when you generate

594
00:57:28,360 --> 00:57:33,120
it, you can actually type in a new pronunciation for that word and then every time the word

595
00:57:33,120 --> 00:57:39,800
appears in the script, the AI will know to say it in the correct way. There's more controls

596
00:57:39,800 --> 00:57:44,120
in this as well. So we were playing around and you can touch on things like emphasis

597
00:57:44,120 --> 00:57:51,200
on a particular word. You can increase the speed at which one of the sentences or paragraphs

598
00:57:51,200 --> 00:57:57,760
is spoken and you can even change the pitch as well. And the good thing is that you can

599
00:57:57,760 --> 00:58:04,360
actually do it with multiple speakers. So you don't just have to have one speaker throughout

600
00:58:04,360 --> 00:58:10,800
just creating one voice. So you can run it as a dialogue as well, which makes it quite

601
00:58:10,800 --> 00:58:17,560
a powerful tool to play with actually. It's not just a very static one tone, one voice

602
00:58:17,560 --> 00:58:21,980
over, which I think that's quite a few tools out there that will do one single voice if

603
00:58:21,980 --> 00:58:26,360
you give it some text. I think this is a real differentiator. In fact, the audio editing

604
00:58:26,360 --> 00:58:30,800
on here will allow you to add sound effects and background tracks as well. So it does

605
00:58:30,800 --> 00:58:38,280
have some more powerful audio editing capabilities beyond that as well. In terms of use cases,

606
00:58:38,280 --> 00:58:42,740
you can use it for anything like podcasts. For instance, I imagine podcasters might want

607
00:58:42,740 --> 00:58:49,800
to do audio intros and outros and things like that. I actually was, I've had the subscription

608
00:58:49,800 --> 00:58:53,460
to this for a while. I don't really use it that often, but I was reminded of it when

609
00:58:53,460 --> 00:58:59,160
I went to a client's office the other day and someone in the office mentioned that they

610
00:58:59,160 --> 00:59:07,120
needed to re-record the telephone message on their, you know, press one for such and

611
00:59:07,120 --> 00:59:12,200
such press two for such, such. Apparently there were some new options or something they

612
00:59:12,200 --> 00:59:15,920
needed to re-record it. And it was one of those where no one in the office wanted to

613
00:59:15,920 --> 00:59:21,080
do it. Everyone was like, I'm not really got the voice for it. And I thought, oh, you could

614
00:59:21,080 --> 00:59:25,160
use, you could use this tool. And I showed off Jenny to them for that particular use

615
00:59:25,160 --> 00:59:32,080
case. So we've actually got an example of it. Hopefully this comes through okay. I'm

616
00:59:32,080 --> 00:59:37,440
sure it will, because if, if it doesn't work in this live presentation version, Paul will

617
00:59:37,440 --> 00:59:45,120
edit it in, I'm sure. But this is a version, it's about a minute long, where I got Google's

618
00:59:45,120 --> 00:59:54,440
Bard to write a script talking about generative AI and audio generative AI specifically. And

619
00:59:54,440 --> 00:59:57,000
we're going to hear that in two different voices now.

620
00:59:57,000 --> 01:00:03,960
Hello everyone. My name is Bard and I'm a large language model from Google AI. I am

621
01:00:03,960 --> 01:00:11,160
here today to talk to you about generative AI and specifically about audio generation.

622
01:00:11,160 --> 01:00:17,280
Generative AI is a type of artificial intelligence that can create new content such as text,

623
01:00:17,280 --> 01:00:24,360
images and music. Audio generation is a specific type of generative AI that can create new

624
01:00:24,360 --> 01:00:31,080
audio content. Audio generation can be used for a variety of purposes, such as creating

625
01:00:31,080 --> 01:00:36,480
music, audio books and podcasts. Really interesting stuff there, Ma, compared to some of the other

626
01:00:36,480 --> 01:00:44,160
tools we've played with, some of the cadence of speech is so much more natural. When the

627
01:00:44,160 --> 01:00:50,120
second speaker says text, images and music, right, which is much more like a human would

628
01:00:50,120 --> 01:00:55,640
say like a human presenter on a radio show versus like text, images and music without

629
01:00:55,640 --> 01:01:02,840
any sort of variance. And yet in the next sentence, she says generative AI in a slightly

630
01:01:02,840 --> 01:01:09,160
weird way that a human would never do. So we're getting so much closer to something

631
01:01:09,160 --> 01:01:15,000
that sounds straight off the bat like a human could have done it with just a few more still

632
01:01:15,000 --> 01:01:19,360
a couple of telltales in there that it's AI generated.

633
01:01:19,360 --> 01:01:25,200
Yeah. And I would say actually in Jenny's defense there, we knock this up in seconds

634
01:01:25,200 --> 01:01:31,440
and you can edit the script to get those little details just right. This was basically zero

635
01:01:31,440 --> 01:01:36,520
editing so we could change the pronunciation of AI, we could add a pause. There are quite

636
01:01:36,520 --> 01:01:41,800
like I say some quite impressive bits of fine tuning control on there.

637
01:01:41,800 --> 01:01:48,360
Pretty cool. Yeah, I think as regular listeners will know when we put together the podcast,

638
01:01:48,360 --> 01:01:54,000
we had the intros and the outros generated by AI. So the lovely person who speaks at

639
01:01:54,000 --> 01:01:58,240
the beginning and at the end is not a real person, I'm afraid. But we had very little

640
01:01:58,240 --> 01:02:03,240
control in the tool we used at the time. I can't remember what it was to control cadence

641
01:02:03,240 --> 01:02:09,640
and how it was pronounced. So this is a really interesting step in terms of greater sophistication.

642
01:02:09,640 --> 01:02:14,000
I can see that being super valuable when you really want to invest quite a lot of time

643
01:02:14,000 --> 01:02:19,140
and effort into getting it just right. And hopefully the tools will get better at predicting

644
01:02:19,140 --> 01:02:23,480
what sort of changes someone would make in order to get that effect. If you were going

645
01:02:23,480 --> 01:02:27,440
to do this at scale and you're just creating quite a lot of content, like if you wanted

646
01:02:27,440 --> 01:02:31,880
someone to narrate every one of your blog posts on your website so that those who were

647
01:02:31,880 --> 01:02:36,800
visually impaired could listen to them or if people who were on the go wanted to listen

648
01:02:36,800 --> 01:02:42,800
to your content instead of read it, I'm sure that you probably wouldn't get the ROI of

649
01:02:42,800 --> 01:02:49,520
going through and absolutely making sure every single one of those was on the nose. But actually

650
01:02:49,520 --> 01:02:55,200
can still do a pretty good job regardless. And certainly very legible is the wrong word,

651
01:02:55,200 --> 01:03:00,520
I guess. Very easy to understand. There's no parts there where you didn't really realize

652
01:03:00,520 --> 01:03:07,480
what the person was saying. And I've got an ebook reader on my phone that has a text to

653
01:03:07,480 --> 01:03:13,880
speech synthesis built into it that just used the standard Google or Samsung one. And you

654
01:03:13,880 --> 01:03:18,760
can't really listen to a book like it's an audio book, right? It's a bit jarring. And

655
01:03:18,760 --> 01:03:23,240
every now and again, you can't quite understand because two sentences get smashed together

656
01:03:23,240 --> 01:03:29,920
or a word gets completely mispronounced. And it would sound like this is going to overcome

657
01:03:29,920 --> 01:03:34,000
quite a lot of those issues for people looking to make it so that their content can be listened

658
01:03:34,000 --> 01:03:45,400
to instead of just read. Cool. So that was Tool of the Week this week. Thank you, Martin,

659
01:03:45,400 --> 01:03:52,480
for telling us about Jenny. If you're interested in synthesizing audio, please do let us know.

660
01:03:52,480 --> 01:03:56,480
We hope you've enjoyed today's podcast. If you have, please subscribe. Tell your marketing

661
01:03:56,480 --> 01:04:01,200
friends to subscribe as well. If you've got any thoughts on anything we've discussed or

662
01:04:01,200 --> 01:04:05,280
if you want to come on the podcast or send us a snippet, hit us up on the LinkedIn or

663
01:04:05,280 --> 01:04:14,240
the Twitters. Our Twitter handle is AIMarketingPod or one word. Hurrah! We got the Twitter handle

664
01:04:14,240 --> 01:04:19,600
right first time and I promise you, dear listener, that's not because Martin is screen sharing

665
01:04:19,600 --> 01:04:25,360
and showing me Twitter as I read that out. Yeah, he actually is, but it's nice to get

666
01:04:25,360 --> 01:04:29,320
it right for once. We would love to hear from you. Thanks very much for joining us again

667
01:04:29,320 --> 01:04:34,280
and we'll speak to you all soon, I'm sure. Thank you, Martin. Cheers, Paul. See you next

668
01:04:34,280 --> 01:04:41,140
week and thanks for listening, everyone. Bye bye. Thank you for listening to Artificially

669
01:04:41,140 --> 01:04:47,000
Intelligent Marketing. To stay on top of the latest trends, tips and tools in the world

670
01:04:47,000 --> 01:04:52,960
of marketing AI, be sure to subscribe. We look forward to seeing you again next week.