1
00:00:00,000 --> 00:00:09,960
Welcome to Artificially Intelligent Marketing, a weekly podcast where we stay on top of the

2
00:00:09,960 --> 00:00:15,700
latest trends, tips, and tools in the world of marketing AI, helping you get the best

3
00:00:15,700 --> 00:00:24,100
results from your marketing efforts. Now let's join our hosts, Paul Avery and Martin Broadhurst.

4
00:00:24,100 --> 00:00:29,960
Welcome to Episode 31 of Artificially Intelligent Marketing. Thanks for joining us. I'm here

5
00:00:29,960 --> 00:00:34,640
as always with my lovely co-host Martin. How are you, Martin?

6
00:00:34,640 --> 00:00:41,920
Yeah, I'm delighted to be back in the studio. New environs, so anybody catching the video

7
00:00:41,920 --> 00:00:50,840
today will see a slightly different setup, courtesy of ChatGPT and Arlie3. But yeah,

8
00:00:50,840 --> 00:00:53,120
no, good to see you. How are you doing?

9
00:00:53,120 --> 00:00:57,000
Yeah, I'm good. Thanks, bud. I'm all good. I've been fighting the lurgy that's been

10
00:00:57,000 --> 00:01:02,560
floating around the UK, so I've been coughing. And if I cough a lot on this podcast, I apologize,

11
00:01:02,560 --> 00:01:08,800
listen, I'm going to try and control it the best I can. Coming back to Martin's backgrounds,

12
00:01:08,800 --> 00:01:13,640
he's been investing a lot of quality time with Dauley3, trying to get backgrounds that

13
00:01:13,640 --> 00:01:17,160
say Artificially Intelligent Marketing. For those of you that have been playing with Dauley3,

14
00:01:17,160 --> 00:01:21,120
you'll know it's one of the image generation tools that's actually pretty good at text.

15
00:01:21,120 --> 00:01:26,320
And whilst it is good at text, if you've got long words or words with like lots of double

16
00:01:26,320 --> 00:01:31,160
letters in like artificially, it's really hard to get an output. So poor Martin's been

17
00:01:31,160 --> 00:01:34,520
bashing away going, oh, there's a typo, oh, there's a typo, oh, there's a typo. And he's

18
00:01:34,520 --> 00:01:41,320
finally got a really beautiful, lovely image where you can read the words properly. Hurrah.

19
00:01:41,320 --> 00:01:45,320
So that's success as far as I'm concerned. Right, let's get into the stories today because

20
00:01:45,320 --> 00:01:48,920
we've got a lot, Martin. First one's with you and it's about HubSpot's been busy, busy.

21
00:01:48,920 --> 00:01:54,680
Yeah, so all of the HubSpot users out there will know that in recent months, since inbound

22
00:01:54,680 --> 00:01:58,720
to the conference in September, they've been making a lot of noise about artificial intelligence

23
00:01:58,720 --> 00:02:05,080
and how they're baking it into their product. But in what is something of a strategic move

24
00:02:05,080 --> 00:02:16,200
for the organisation, they are acquiring the B2B data titan Clearbit. So Clearbit is a

25
00:02:16,200 --> 00:02:23,520
leader in B2B data and intelligence. And this acquisition is set to really enhance HubSpot's

26
00:02:23,520 --> 00:02:31,120
capabilities by integrating all of that data, all of that firmographic, demographic, technographic

27
00:02:31,120 --> 00:02:37,200
insight that they have in their database, which covers millions and millions of B2B

28
00:02:37,200 --> 00:02:43,640
companies. It's going to be integrated into the core HubSpot platform. So this really

29
00:02:43,640 --> 00:02:53,480
brings unparalleled business corporate data into HubSpot. There's no other CRM that has

30
00:02:53,480 --> 00:03:01,720
baked this into the core product. So this means that, well, actually, we're assuming

31
00:03:01,720 --> 00:03:06,800
here we don't quite know how it's going to be baked into the core product just yet, but

32
00:03:06,800 --> 00:03:15,240
it's assumed that the merger will give HubSpot users the ability to access all of that customer

33
00:03:15,240 --> 00:03:21,080
data for prospecting. Now, why are we talking about this on a marketing podcast or the marketing

34
00:03:21,080 --> 00:03:29,280
AI podcast other than the fact that it's HubSpot? Well, the interesting play here is that data,

35
00:03:29,280 --> 00:03:35,320
right, because models are built on top of data and HubSpot have obviously gone very

36
00:03:35,320 --> 00:03:46,080
big on AI recently. So it makes sense that we can expect them to start using AI and their

37
00:03:46,080 --> 00:03:52,840
own models to help companies achieve better prospecting, better segmentation. You can

38
00:03:52,840 --> 00:04:01,080
imagine AI powered segmentation built in to the sequences tool for sales reps into your

39
00:04:01,080 --> 00:04:06,880
email marketing campaigns. And it just does it all for you. You know, I always think segmentation

40
00:04:06,880 --> 00:04:13,840
is actually one of the areas that marketers don't get right or don't spend enough time

41
00:04:13,840 --> 00:04:19,800
on. So to have an AI that could potentially do that for you is going to be massive. Speaking

42
00:04:19,800 --> 00:04:26,120
of the acquisition, Yamini Rangan, the CEO of HubSpot said, to cut through the noise

43
00:04:26,120 --> 00:04:31,600
with deep relevance, businesses need reliable, high quality data about their customers. That

44
00:04:31,600 --> 00:04:37,880
means enriching your company's internal customer data with real time external context. ClearBit

45
00:04:37,880 --> 00:04:44,000
has made it its mission to collect rich and useful data about millions of companies. HubSpot's

46
00:04:44,000 --> 00:04:49,600
AI powered customer platform combined with ClearBit's data will create a powerful winning

47
00:04:49,600 --> 00:04:58,440
combination for our customers. So HubSpot clearly sees this as an evolution in customer

48
00:04:58,440 --> 00:05:09,160
intelligence. So this is hopefully going to enable us to deliver better customer engagement

49
00:05:09,160 --> 00:05:14,880
strategies in real time. And that's going to work across the customer life cycle, I

50
00:05:14,880 --> 00:05:19,640
would imagine. So right the way from customer acquisition through to customer service and

51
00:05:19,640 --> 00:05:26,440
delivery. This data is going to be incredibly powerful. I think it's a massive move for

52
00:05:26,440 --> 00:05:33,120
HubSpot. HubSpot historically has spoken about crafted, not cobbled. And what they mean by

53
00:05:33,120 --> 00:05:39,520
that is the platform itself has all of the elements in it have been built by the HubSpot

54
00:05:39,520 --> 00:05:45,200
team. They haven't acquired lots of companies and then try to shoehorn them into an existing

55
00:05:45,200 --> 00:05:51,600
product, which we have seen from many other providers. You know, Salesforce is the one

56
00:05:51,600 --> 00:05:56,280
that springs to mind. They buy it all and then just try and crowbar it in. Whereas HubSpot's

57
00:05:56,280 --> 00:06:04,000
built rather than bolted on. This obviously represents a slightly different approach.

58
00:06:04,000 --> 00:06:10,280
However, if we look to how HubSpot have maybe integrated other companies that they have

59
00:06:10,280 --> 00:06:14,360
acquired, now the only one that really springs to mind is a company called Pysync. They did

60
00:06:14,360 --> 00:06:18,680
a really good job of actually integrating that and making it feel crafted, not cobbled.

61
00:06:18,680 --> 00:06:23,920
So hopefully we see the same thing here as well. So yeah, I'm interested. What are your

62
00:06:23,920 --> 00:06:27,760
thoughts on this one? Huge amounts of B2B data coming into HubSpot. What do you think

63
00:06:27,760 --> 00:06:30,560
they're going to do with it? Yeah, there's a couple of things that sprung

64
00:06:30,560 --> 00:06:35,880
to mind. The first was as a marketer, I'm super intrigued and quite excited. As a consumer,

65
00:06:35,880 --> 00:06:43,680
I'm not sure. And the reason is that power I think is really useful as a marketer to

66
00:06:43,680 --> 00:06:47,840
segment people, be able to send them relevant messages, content, information, you know,

67
00:06:47,840 --> 00:06:51,680
right message, right person, right time gets much easier the more data you have. But of

68
00:06:51,680 --> 00:06:56,960
course, I think anything that's built on having data on people that they may or may not know

69
00:06:56,960 --> 00:07:01,720
has been collected about them can make consumers a bit nervous and would make me a little bit

70
00:07:01,720 --> 00:07:08,640
nervous. So I think there's an aspect to that, which combines with one of the things, the

71
00:07:08,640 --> 00:07:13,400
first thing I thought when I saw this was the chat spot demo video that we saw Dharmesh

72
00:07:13,400 --> 00:07:20,320
show many months ago now of automatic prospecting emails, highly customized to each lead, but

73
00:07:20,320 --> 00:07:27,120
that had been written by the bot. And I think it's struggled to reach that point for a number

74
00:07:27,120 --> 00:07:30,400
of different reasons. In fact, you could argue that doesn't really exist in chat spot at

75
00:07:30,400 --> 00:07:35,360
the moment. But this data enrichment is a big step towards being able to do that as

76
00:07:35,360 --> 00:07:40,800
far as I can tell. And then that's the bit that as a consumer, I'm a bit nervous because

77
00:07:40,800 --> 00:07:47,240
one of the reasons that outbound at scale doesn't work is the inability to personalize

78
00:07:47,240 --> 00:07:55,520
it just makes it pants. We all know those emails. Hi, Paul. Have you got any money?

79
00:07:55,520 --> 00:07:59,320
We'd like some of your money, please. I get so many of those and I filter them out and

80
00:07:59,320 --> 00:08:03,560
I ignore them. And I can't tell if it's a good or a bad thing that we can use these

81
00:08:03,560 --> 00:08:10,960
data enrichment and chat GPT style bots to do that type of outbound, but better. I think

82
00:08:10,960 --> 00:08:17,180
that's maybe good for consumers, but maybe not because part of how I filter out the prospecting

83
00:08:17,180 --> 00:08:22,400
emails I get is whether actually a human paid enough attention to really know what I do,

84
00:08:22,400 --> 00:08:27,480
what I need and make it about me and not about them. And of course, Paul, I've noticed you've

85
00:08:27,480 --> 00:08:37,440
got a pulse. Would you like to buy my product? Yeah, exactly. So I think this is the issue.

86
00:08:37,440 --> 00:08:41,580
And so that's the bit I'm a bit nervous about. But ultimately I can, I think it's a bold

87
00:08:41,580 --> 00:08:47,160
move by HubSpot and I can see exactly why they've done it. And I think we'll see other

88
00:08:47,160 --> 00:08:52,720
players follow suit now. Yeah, there's other data providers out there that I'm sure would

89
00:08:52,720 --> 00:09:00,440
be a good fit for many of the CRM companies. Cool. Thanks for sharing that story with us,

90
00:09:00,440 --> 00:09:06,520
Martin. Let's have a little look at our next one, which is it's somewhat of a rumour. There's

91
00:09:06,520 --> 00:09:09,880
a few screw grabs floating around the internet of this, but certainly Martin and I haven't

92
00:09:09,880 --> 00:09:17,120
got access yet, which is open AI upgrading chat GPT so that you can access all of GPT

93
00:09:17,120 --> 00:09:24,000
four's tools without having to switch between them. So for users of GPT four, you'll know

94
00:09:24,000 --> 00:09:27,680
if you want to search with Bing, you have to select that in the dropdown. If you want

95
00:09:27,680 --> 00:09:34,760
to upload an image, then you have to choose a particular search modal. If you want to

96
00:09:34,760 --> 00:09:39,400
use DORLY three, you've got to select that. If you want to use plugins, you've got to

97
00:09:39,400 --> 00:09:44,800
select that. If you got to want to use advanced data analysis, you got to select that. So

98
00:09:44,800 --> 00:09:48,440
you can't easily combine things. So if you want to use an image as an inspiration for

99
00:09:48,440 --> 00:09:52,720
creating another image in DORLY three, you first have to pass the image through standard

100
00:09:52,720 --> 00:09:57,800
GPT four to get a prompt idea and then switch to DORLY three to try and get your image.

101
00:09:57,800 --> 00:10:03,280
It's a pain. Apparently now you would just converse with GPT four and it will know what

102
00:10:03,280 --> 00:10:05,960
you want to do. So if you were, if you give it an Excel file, it will know you probably

103
00:10:05,960 --> 00:10:10,100
want it to analyse it. If you ask for an image, you'll know that you want to get an image.

104
00:10:10,100 --> 00:10:13,760
So that'll obviously make a lot of things much easier for people and open up new workflows

105
00:10:13,760 --> 00:10:19,520
that were just too much of a pain in the bottom to do. So I think that's really cool. The

106
00:10:19,520 --> 00:10:24,840
other thing is they're expanding the file types that chat GPT can easily work with to

107
00:10:24,840 --> 00:10:30,360
include things like PDFs. So it'd be much easier to interact with, interrogate chat

108
00:10:30,360 --> 00:10:36,280
with if you like your PDFs within GPT four. So it's really interesting because I think

109
00:10:36,280 --> 00:10:39,640
it's going to make it much easier for us all to use them. But at the same time, it just

110
00:10:39,640 --> 00:10:43,720
rendered a bunch of third party applications completely redundant, including a number of

111
00:10:43,720 --> 00:10:50,960
plugins within chat GPT. And I think it's pretty interesting. You know, we talk on the

112
00:10:50,960 --> 00:10:57,000
podcast a lot about moats and people building tools on the back of GPT four, et cetera.

113
00:10:57,000 --> 00:11:04,320
But I definitely see open AI leaning ever more into trying to make chat GPT itself an indispensable

114
00:11:04,320 --> 00:11:12,220
tool and therefore pushing you not to use third party tools whilst at the same time

115
00:11:12,220 --> 00:11:17,880
trying to build a thriving developer community of people who build on top of open AI's tools

116
00:11:17,880 --> 00:11:21,880
to create their own tools. So it's, there's quite an interesting tension there. But I

117
00:11:21,880 --> 00:11:26,960
think as a user, it's great as a developer, you're thinking how long until open AI and

118
00:11:26,960 --> 00:11:32,600
chat GPT standalone product tries to eat my lunch?

119
00:11:32,600 --> 00:11:38,240
If I'm a developer, I'm looking to build on top of the API, right? If I'm, if I'm a developer

120
00:11:38,240 --> 00:11:42,680
within this existing product, I might try and plug that product into chat GPT so people

121
00:11:42,680 --> 00:11:48,760
can connect with it and get the information from my product or inject information into

122
00:11:48,760 --> 00:11:54,500
my product, product and get it back out into the chat GPT interface. I'm not building a

123
00:11:54,500 --> 00:12:02,640
specific plugin for chat GPT, absolutely not, not happening. To be honest, when, as soon

124
00:12:02,640 --> 00:12:08,920
as you said it's killing off or likely to kill off a bunch of third party plugins, my

125
00:12:08,920 --> 00:12:20,160
internal thought was good. That plugin library is a mess. There is some absolute trash in

126
00:12:20,160 --> 00:12:25,680
there, some terrible plugins. It needs to sort out. It's not fit for purpose at the

127
00:12:25,680 --> 00:12:31,120
moment. I understand the excitement to build on it, but anytime one of these kinds of marketplaces

128
00:12:31,120 --> 00:12:37,880
launches and there's no review system, there's no quality control on it. You just end up

129
00:12:37,880 --> 00:12:43,920
being populated with not very good products. So hopefully it does get rid of it. In terms

130
00:12:43,920 --> 00:12:51,880
of the actual shift in the product, so everything being integrated into the one chat now, which

131
00:12:51,880 --> 00:12:58,140
is, I think that's great. As you said, the frustration at the moment is having to jump

132
00:12:58,140 --> 00:13:03,940
in and out of different conversations because you're working in a, let's say you're working

133
00:13:03,940 --> 00:13:11,440
in a CSV file, extracting a bunch of data and then getting it to tell you what you might

134
00:13:11,440 --> 00:13:16,860
be able to do for, I don't know, something like a annual report or what could we do to

135
00:13:16,860 --> 00:13:23,800
inspire the next marketing campaign? And then you ask it for some creative, well, you've

136
00:13:23,800 --> 00:13:28,680
got to go out of that chat to get the creative inspiration and you kind of back and forth

137
00:13:28,680 --> 00:13:33,440
between conversations. So anything that can streamline that process is a good thing. The

138
00:13:33,440 --> 00:13:40,320
PDFs being able to interact with them directly, I think is obviously welcome. However, I would

139
00:13:40,320 --> 00:13:44,680
imagine that we're still limited by that token size. So it's not like Claude, where we can

140
00:13:44,680 --> 00:13:53,000
just chuck in like a really long annual report, you know, and however many thousands of words

141
00:13:53,000 --> 00:13:58,440
you know, eight, 10,000 words or something like that and have it interact with that is

142
00:13:58,440 --> 00:14:02,080
still going to be limited by its smaller context.

143
00:14:02,080 --> 00:14:06,320
Yeah, I'm glad you said that. As you know, Martin and probably listeners too, one of

144
00:14:06,320 --> 00:14:11,160
my favourite tools is Magi because I can access a bunch of different models all within that

145
00:14:11,160 --> 00:14:17,080
one platform. And they recently removed access to GPT-432K and when I asked them, they said

146
00:14:17,080 --> 00:14:20,920
it was prohibitively expensive for them to run it in their current form, which is not

147
00:14:20,920 --> 00:14:27,280
surprising, given I think it's four or five times more expensive per token. But I really

148
00:14:27,280 --> 00:14:34,720
missed the context window once it was taken away because you can upload PDFs in Magi so

149
00:14:34,720 --> 00:14:40,880
you could do this already. But I've come to rely on GPT-4 instead of Claude because, you

150
00:14:40,880 --> 00:14:46,400
know, 32K was actually proving to be enough for a lot of my use cases. And the minute

151
00:14:46,400 --> 00:14:53,720
they took it away, I was getting context error all the time from pretty much any half decent

152
00:14:53,720 --> 00:15:00,120
PDF I was uploading or transcripts. So I agree. I think how long until they need to combine

153
00:15:00,120 --> 00:15:07,320
this with 32K GPT-4 as standard as a context window because people are going to find that

154
00:15:07,320 --> 00:15:11,680
their analysis and chatting with PDFs is going to not work a lot of the time otherwise, I

155
00:15:11,680 --> 00:15:12,680
think.

156
00:15:12,680 --> 00:15:17,920
But on the 32K pricing thing, it is expensive, right? You don't have to play with it for

157
00:15:17,920 --> 00:15:23,440
long before you suddenly go, oh yeah, those API costs are really starting to rack up.

158
00:15:23,440 --> 00:15:28,840
So I'm not surprised, Magi, I've done that. Given what is it? Like the team's plan is

159
00:15:28,840 --> 00:15:34,640
$49 per month. Yeah, like I can see why they couldn't make that work. I did wonder when

160
00:15:34,640 --> 00:15:38,360
they introduced it, but they were running an experiment. I respect that. We've obviously

161
00:15:38,360 --> 00:15:42,720
got the OpenAI Developer Conference next week. We're talking offline if we're going to have

162
00:15:42,720 --> 00:15:47,320
to do a special episode just for that. But we expect, and most of the world expects,

163
00:15:47,320 --> 00:15:51,880
so pay attention to this stuff, something around making it cheaper for developers to

164
00:15:51,880 --> 00:15:57,120
use the tools. And if that turns out to be the case, then hopefully I'll get my GPT-4

165
00:15:57,120 --> 00:15:59,680
32K back because I was enjoying it.

166
00:15:59,680 --> 00:16:03,720
Whilst we are on the subject of OpenAI models, and we haven't got it in as one of the stories

167
00:16:03,720 --> 00:16:08,720
for today, but I thought it was very interesting that Microsoft published that paper that showed

168
00:16:08,720 --> 00:16:18,600
us the chat GPT 3.5 turbo is being run on just 20 billion parameters rather than the

169
00:16:18,600 --> 00:16:26,320
175 billion of chat GPT-3. I think that goes to show what they can do with optimization

170
00:16:26,320 --> 00:16:28,000
of these models at the moment.

171
00:16:28,000 --> 00:16:33,280
Yeah. And one has to assume any dropping costs from OpenAI is going to be because of less

172
00:16:33,280 --> 00:16:38,360
compute which is going to be simpler models with less parameters, simpler I guess in inverted

173
00:16:38,360 --> 00:16:44,320
commas, optimized fine-tuned models. So maybe that's because GPT-4 is about to have its

174
00:16:44,320 --> 00:16:50,600
turbo moment, reducing the costs and making things like 32K context more applicable.

175
00:16:50,600 --> 00:16:55,440
While we're on the subject of chat GPT, I've also been playing a lot with Dorely 3, as

176
00:16:55,440 --> 00:16:59,440
we all know Martin has because he's been working really hard to get his awesome backgrounds.

177
00:16:59,440 --> 00:17:05,760
One of the things I've been doing in Dorely 3, I saw online and I gave it a go was enabling

178
00:17:05,760 --> 00:17:10,200
an image manipulation feature that you and I might have talked about a few times on the

179
00:17:10,200 --> 00:17:14,880
show as it relates to mid-journey, but it's not quite possible. But because of Dorely

180
00:17:14,880 --> 00:17:21,640
3's ability to understand the context of the image, you can modify bits of the image. So

181
00:17:21,640 --> 00:17:25,080
if you're trying to create an image and you want to change one specific bit of it, you

182
00:17:25,080 --> 00:17:30,960
can. And the way that you do this is you create an image. So in the example that from when

183
00:17:30,960 --> 00:17:36,440
I was playing with this, I created a turtle in a lab coat and I asked Dorely 3 for the

184
00:17:36,440 --> 00:17:41,280
seed number of that image. So the seed, this is true of all of these models and how they

185
00:17:41,280 --> 00:17:46,800
create images. The seed is like the unique identifier that inspired that image. So if

186
00:17:46,800 --> 00:17:51,640
you keep the seed same between the same between images, you'll get not the same image, but

187
00:17:51,640 --> 00:17:56,880
similar. So you might have seen online tutorials about how to try and keep the characters similar

188
00:17:56,880 --> 00:18:02,600
between mid-journey images. And one way is to try and use that seed. Because of the way

189
00:18:02,600 --> 00:18:07,520
the images are generated, they're never exactly the same, but they're pretty close. Combine

190
00:18:07,520 --> 00:18:12,680
using the seed with Dorely 3's ability to understand what's in an image and you can

191
00:18:12,680 --> 00:18:18,060
modify specific bits. So I asked for the turtle to have a top hat and it gave pretty much

192
00:18:18,060 --> 00:18:22,320
the same turtle a top hat with a slightly different background, but very similar to

193
00:18:22,320 --> 00:18:27,320
the original. And then I asked for that top hat to be purple and it made only the top

194
00:18:27,320 --> 00:18:32,280
hat purple. And again, a few subtle things about the image change, but by and large,

195
00:18:32,280 --> 00:18:36,520
it was the same turtle and the same background. So this actually opens up new applications

196
00:18:36,520 --> 00:18:40,160
for marketers that they wouldn't have had before, right, Martin? Because that's something

197
00:18:40,160 --> 00:18:44,160
we've wanted for a while. And again, it's not perfect because the image is never quite

198
00:18:44,160 --> 00:18:46,040
the same, but it is a step in the right direction.

199
00:18:46,040 --> 00:18:49,640
Right. Let's move on to our next story, Martin. This one's with you.

200
00:18:49,640 --> 00:18:57,400
This is an update from Stability AI, the minds behind stable diffusion and a bunch of other

201
00:18:57,400 --> 00:19:06,560
open source models such as Beluga, which is the language model. They've just announced

202
00:19:06,560 --> 00:19:17,360
some new applications designed specifically for business, right? So enterprise grade APIs

203
00:19:17,360 --> 00:19:22,560
and new product functionality. And what's interesting about this approach is that in

204
00:19:22,560 --> 00:19:27,520
their marketing and communications around this, they've really gone kind of sector specific,

205
00:19:27,520 --> 00:19:34,840
which I thought was quite an interesting approach. So central to this is the new sky replacer

206
00:19:34,840 --> 00:19:42,440
tool. So if you've got a photo of anything and it has some sky in it, it will automatically

207
00:19:42,440 --> 00:19:47,880
replace the whole sky. So if it's a gray cloudy day in your photo, you can make the sky bright,

208
00:19:47,880 --> 00:19:53,520
sunny blue skies, a couple of clouds, that kind of thing. And they speak about this being

209
00:19:53,520 --> 00:19:59,240
for the real estate industry. And if you check out the website, they're really driving that

210
00:19:59,240 --> 00:20:05,800
point home. So if you've got a photo of a property, you can get that with the nice blue

211
00:20:05,800 --> 00:20:14,540
skies. I think that's been available in ClipDrop for a couple of weeks, but this is now being

212
00:20:14,540 --> 00:20:23,400
introduced into the API as well. So if you're selling software into the real estate industry

213
00:20:23,400 --> 00:20:28,000
and people are uploading images to things like Rightmove or anything like that, you

214
00:20:28,000 --> 00:20:36,200
can actually plug into the API and make it part of the workflow. As well as that, they've

215
00:20:36,200 --> 00:20:43,360
added another dimension quite literally using stable 3D. So this at the moment is a private

216
00:20:43,360 --> 00:20:53,640
preview feature. And this is about reshaping the 3D content creation landscape really.

217
00:20:53,640 --> 00:21:01,040
So it's about 3D models created by AI. So it enables rapid generation of textured 3D

218
00:21:01,040 --> 00:21:07,400
objects, making the process much quicker for artists that might be working in the video

219
00:21:07,400 --> 00:21:13,960
game industry or architecture or anything like that. So there's plenty of use cases

220
00:21:13,960 --> 00:21:21,680
there and you can imagine this is going to be taken up by lots of the likes of what's

221
00:21:21,680 --> 00:21:28,080
that 3D modeling software for home builders, SketchUp and things like that, which I've

222
00:21:28,080 --> 00:21:33,200
played with in the past and rapidly realized that it's well outside of my wheelhouse and

223
00:21:33,200 --> 00:21:43,100
I should not do it anymore. They've also announced stability or stable fine tuning is what they're

224
00:21:43,100 --> 00:21:51,200
calling it, which is the ability to fine tune image creation. So it offers high speed customization

225
00:21:51,200 --> 00:21:56,860
of images catering to the creative demands of industry such as entertainment, gaming

226
00:21:56,860 --> 00:22:03,080
and marketers. So the tool allows for the infusion of brand specific aesthetics into

227
00:22:03,080 --> 00:22:09,480
the visuals that you create using models like stable diffusion, which to your point, just

228
00:22:09,480 --> 00:22:16,720
a moment ago, talking about the capabilities of Dolly 3 and being able to see the images,

229
00:22:16,720 --> 00:22:22,880
and now we can fine tune models on our own brand aesthetics. At the moment, this is not

230
00:22:22,880 --> 00:22:26,320
open to everybody. They've just said that it's coming soon and if you're interested

231
00:22:26,320 --> 00:22:35,280
in getting on board, you should contact them. But that ability to create brand focused brand

232
00:22:35,280 --> 00:22:42,760
centric visuals using generative AI is the kind of golden land, right? That's the bit

233
00:22:42,760 --> 00:22:48,520
that everybody's after, isn't it? Just this morning, in fact, I was on a forum that's

234
00:22:48,520 --> 00:22:53,400
talking about AI and one of the questions from somebody was saying, is there a way to

235
00:22:53,400 --> 00:22:58,400
achieve consistency and style and aesthetic? It keeps giving me slightly different looks

236
00:22:58,400 --> 00:23:04,240
and feels and the response was an overwhelming no from several commenters. But now this is

237
00:23:04,240 --> 00:23:08,520
coming. We're going to have this with fine tuning. So that's quite exciting.

238
00:23:08,520 --> 00:23:14,440
Yeah, I think that's the big part of it, which is these are fun to play with and you can

239
00:23:14,440 --> 00:23:21,160
get some fun stuff and fun effects, but it's kind of random. You have to work quite hard

240
00:23:21,160 --> 00:23:26,320
to even get close to what you were imagining. And if we think about the ultimate goal of

241
00:23:26,320 --> 00:23:31,920
these systems to be some sort of AGI for image development, you should be able to have an

242
00:23:31,920 --> 00:23:37,040
ongoing chat with the bot that ever more fine tunes the image you're trying to get, right?

243
00:23:37,040 --> 00:23:43,120
Just like in your images, we had a little giggle before we started. You've got a football

244
00:23:43,120 --> 00:23:46,680
image for those that are not watching the video in your background and the center circle

245
00:23:46,680 --> 00:23:51,280
has another center circle in the middle of it, which of course is not what a real football

246
00:23:51,280 --> 00:23:56,680
pitch looks like a soccer pitch for our American listeners. So your ability to say, hey, Dauley

247
00:23:56,680 --> 00:24:02,340
three, remove that middle circle, it's wrong. And for Dauley three to be able to just remove

248
00:24:02,340 --> 00:24:09,600
it for you has to be the ultimate end goal of these tools as has keeping characters consistent,

249
00:24:09,600 --> 00:24:13,960
keeping brand aesthetics consistent. And until we get to that point, they're going to be

250
00:24:13,960 --> 00:24:18,120
fiddly to get what you want. Or in some cases for a lot of designers, they're going to be

251
00:24:18,120 --> 00:24:22,760
good for brainstorming, but they're not going to be good for actually primary image asset

252
00:24:22,760 --> 00:24:26,720
creation because a lot of the work is still going to have to be done in Photoshop. And

253
00:24:26,720 --> 00:24:33,240
on that, I started to watch a video earlier today of someone showing a tutorial of how

254
00:24:33,240 --> 00:24:38,280
to better control characters in Dauley three, because I was like, Oh, this C trick is pretty

255
00:24:38,280 --> 00:24:43,320
cool. And still there are a number of times where the person giving the demo was like,

256
00:24:43,320 --> 00:24:47,440
yeah, so for that bit, I'll probably have to pull that out and do it in Photoshop. And

257
00:24:47,440 --> 00:24:52,160
until we can eliminate that, you can't open design control up to the hands of people who

258
00:24:52,160 --> 00:24:57,520
are non designers to get them specifically what they want, who can only get approximately

259
00:24:57,520 --> 00:24:59,040
what you want at the moment.

260
00:24:59,040 --> 00:25:04,320
No, absolutely. And I mean, these, they still have function, right? They still have value

261
00:25:04,320 --> 00:25:09,840
and utility. I was delivering a workshop on generative AI the other day, and there was

262
00:25:09,840 --> 00:25:19,320
a craft brewery owner who isn't creative themselves. They work with the designers, but they often

263
00:25:19,320 --> 00:25:25,360
have an idea of what they want in their head. And taking them through chat GPT and the image

264
00:25:25,360 --> 00:25:32,020
creation capabilities. Now I showed him basically take a description of one of the new products

265
00:25:32,020 --> 00:25:39,160
that they're just launching, stick it in and say design some, some can designs and some

266
00:25:39,160 --> 00:25:45,520
artwork for the product and just put in a few, you know, whether you want it pixel art

267
00:25:45,520 --> 00:25:50,960
or if there's any kind of features you want. And yeah, he looked at that and he was blown

268
00:25:50,960 --> 00:25:55,680
away by the outputs because he was like, that's perfect. I can give that to the designer now

269
00:25:55,680 --> 00:26:01,160
and say, that's what we're aiming for. Go and make that. Whereas before he sometimes

270
00:26:01,160 --> 00:26:07,920
kind of had trouble articulating exactly what he was trying to get across. So yeah, I definitely

271
00:26:07,920 --> 00:26:14,560
think there's value there. And, and the gap that will shrink is the gap between the original

272
00:26:14,560 --> 00:26:21,080
idea from the non design oriented person and the design work that the designer has to do.

273
00:26:21,080 --> 00:26:27,920
While we're on the topic of images, there was quite an interesting court case this week

274
00:26:27,920 --> 00:26:35,920
because AI art platforms mid journey stable diffusion, deviant art, etc. were in court

275
00:26:35,920 --> 00:26:41,080
around copyright infringement cases for three artists, but in essence, they were basically

276
00:26:41,080 --> 00:26:48,100
acquitted of all of the litigation that was brought against them, except for one count

277
00:26:48,100 --> 00:26:53,800
of copyright violation, which is proceeding. So a pivotal aspect of this verdict is the

278
00:26:53,800 --> 00:26:58,640
fact that the artists failed to demonstrate substantial similarity between their original

279
00:26:58,640 --> 00:27:04,400
artwork and the AI generated images. And that furthermore, copyright claims can only be

280
00:27:04,400 --> 00:27:08,680
brought forth if the artist is officially copyrighted, which was a step that had actually

281
00:27:08,680 --> 00:27:17,120
been overlooked by two of the artists in the case, which is kind of detail detail. So I

282
00:27:17,120 --> 00:27:25,280
think the the essence here is, I think more powerful AI art copyright cases will be brought

283
00:27:25,280 --> 00:27:32,320
that have potentially been a bit more robust. In other words, copyrights already been sought.

284
00:27:32,320 --> 00:27:37,560
But it is interesting this concept around substantial similarity between the artwork

285
00:27:37,560 --> 00:27:42,360
and the AI generated images and how hard that might be able might be to prove unequivocally,

286
00:27:42,360 --> 00:27:48,600
right, which is probably what the court will be looking for. The we had this with the Gloria

287
00:27:48,600 --> 00:27:53,840
Gaynor song, right? If it pulls out the exact lyrics, that's pretty unequivocal. I can see

288
00:27:53,840 --> 00:27:58,360
they're definitely being rulings based on that. But when there's like some similarity

289
00:27:58,360 --> 00:28:03,880
and it's in the gray, it's going to come down to the judge's discretion. I would guess

290
00:28:03,880 --> 00:28:09,600
I'm not a lawyer, but it feels like that's probably where it would drift. So this AI

291
00:28:09,600 --> 00:28:14,760
copyright issue continues to be super messy. But at least in this case, the image creation

292
00:28:14,760 --> 00:28:19,760
tools have dodged this particular case, which in itself could start to set the precedent

293
00:28:19,760 --> 00:28:24,440
for other cases. So yeah, I thought it was interesting.

294
00:28:24,440 --> 00:28:29,600
For now, they've dodged it, right? Because the important part of this story is that actually,

295
00:28:29,600 --> 00:28:37,560
there was one case that continues. So Anderson, the third artist, has the opportunity to revisit

296
00:28:37,560 --> 00:28:47,200
her lawsuit. But it seems like the court is really saying, look, this area is fuzzy. Without

297
00:28:47,200 --> 00:28:52,120
reading the actual ruling, it seems to suggest that they're leaning on the regulators to

298
00:28:52,120 --> 00:28:57,160
give more clarity and definition here, which, you know, all eyes on the US Copyright Office

299
00:28:57,160 --> 00:28:58,160
on this one, I guess.

300
00:28:58,160 --> 00:29:01,840
Yeah, and I think that will continue. And I think there's a lot of businesses that are

301
00:29:01,840 --> 00:29:06,960
in that, oh, I don't know if we can use these tools. What's going to happen with all the

302
00:29:06,960 --> 00:29:11,400
copyright? Are we going to get some massive case leveraged against us? And I think a lot

303
00:29:11,400 --> 00:29:15,720
of big businesses on the best amount of risk in this case is zero risk. And so they're

304
00:29:15,720 --> 00:29:20,160
like, we're just not even going to dabble with these with these tools. And I do understand

305
00:29:20,160 --> 00:29:25,160
that. But it's interesting, especially because a lot of the cases at the moment are being

306
00:29:25,160 --> 00:29:32,240
brought against the developers of the models, not the users of the models. It strikes me

307
00:29:32,240 --> 00:29:37,120
it's a bit like illegal soccer streams, right? Like, do you go after all the people who are

308
00:29:37,120 --> 00:29:41,000
like trying to stream it? Or do you go after the five or six people who are actually hosting

309
00:29:41,000 --> 00:29:45,920
the streams and trying to make all the money off the back of them? And I think what the

310
00:29:45,920 --> 00:29:50,160
authorities learn from the different avenues they took is that you go after the people

311
00:29:50,160 --> 00:29:53,900
who are setting up the streams because A, it's easier and B, they're the ones causing

312
00:29:53,900 --> 00:29:58,200
the most of the problems really. If you cut it off at source, then you don't have your

313
00:29:58,200 --> 00:30:02,440
issues. But as we say, we're not lawyers. So we just report on this because for those

314
00:30:02,440 --> 00:30:07,600
of you that are trying to keep an eye on what's happening in the world of AI and copyright,

315
00:30:07,600 --> 00:30:13,240
it's continues to be a weaving tail with a lot of supple subtlety, a lot of gray and

316
00:30:13,240 --> 00:30:15,120
at the moment not much clarity.

317
00:30:15,120 --> 00:30:20,960
But it is having real world impact on some of the tools that we use these lawsuits in

318
00:30:20,960 --> 00:30:26,240
particular. So you were just speaking then about lawsuits being taken about the model

319
00:30:26,240 --> 00:30:33,960
developers and anthropic developers of my favorite model, Claude II, are facing a legal

320
00:30:33,960 --> 00:30:42,560
challenge from Universal Music. And an observation of mine is that they've made Claude II in

321
00:30:42,560 --> 00:30:48,720
response to this universally frustrating. And this is because what they've effectively

322
00:30:48,720 --> 00:30:59,240
done is immediately changed the model to be overly sensitive to issues of copyright. So

323
00:30:59,240 --> 00:31:06,720
we discussed the Universal Music case last week. Basically, Universal Music are saying

324
00:31:06,720 --> 00:31:15,760
that the model is now trained on copyright, such as song lyrics, and is reproducing those

325
00:31:15,760 --> 00:31:21,840
without permission and without the relevant licenses. So that's why the lawsuit is being

326
00:31:21,840 --> 00:31:27,800
brought. What Anthropic seem to have done is turned Claude into this very sensitive

327
00:31:27,800 --> 00:31:34,520
machine for copyright detection, but even in cases where it isn't necessarily relevant.

328
00:31:34,520 --> 00:31:41,440
Now I love Claude because of its hundred thousand token context window. You can do so much with

329
00:31:41,440 --> 00:31:53,160
that. It's brilliant. You can throw in long PDFs like research papers or policy documents.

330
00:31:53,160 --> 00:31:55,640
And that's been brilliant for a long time because you can get these great summaries,

331
00:31:55,640 --> 00:32:00,880
you can do quote extraction, there's loads of functionality, right. But what's happened

332
00:32:00,880 --> 00:32:10,160
now is that when you do that, it responds with, sorry, I can't give you a detailed summary

333
00:32:10,160 --> 00:32:17,440
of this transcript, of this paper, of this document, because of copyright issues. But

334
00:32:17,440 --> 00:32:23,040
even in areas where it's not relevant. Case in point, the White House has issued an executive

335
00:32:23,040 --> 00:32:29,320
order about AI, which we'll discuss later in the episode, 19,000 words long, not the

336
00:32:29,320 --> 00:32:36,200
sort of thing I want to be sat reading of an evening. So I wanted a fairly comprehensive

337
00:32:36,200 --> 00:32:44,000
summary, uploaded that PDF to Claude. This is a public domain document. There is no copyright

338
00:32:44,000 --> 00:32:49,100
on this document, right. This is a government document available for everyone. And immediately

339
00:32:49,100 --> 00:32:54,800
it says, sorry, can't give you a detailed summary of this because of copyright issues.

340
00:32:54,800 --> 00:33:00,560
I can give you a high level summary and it gives me like 10 bullet points, right, which

341
00:33:00,560 --> 00:33:11,360
is not ideal for a 19,000 word document. Now, you can do soft jailbreaking with it. So if

342
00:33:11,360 --> 00:33:16,760
you go, if you push back on the model slightly and say, no, this is in the public domain,

343
00:33:16,760 --> 00:33:23,600
it's a government document, there's no copyright issue here, it will go, oh, yeah, sorry, my

344
00:33:23,600 --> 00:33:30,720
mistake, in which case we'll continue. In fact, it does it in some absurd ways. I uploaded

345
00:33:30,720 --> 00:33:34,880
the transcript from an interview that we did with one of our guests recently, so clearly

346
00:33:34,880 --> 00:33:41,680
we own the copyright on that, uploaded the transcript, it refused to work on it. I said,

347
00:33:41,680 --> 00:33:46,560
no, no, it's fine. I recorded it, I own the copyright. And it went, oh, now that you've

348
00:33:46,560 --> 00:33:54,440
proved you own the copyright, let's continue. It has a low threshold for proof, I would

349
00:33:54,440 --> 00:33:55,440
say.

350
00:33:55,440 --> 00:34:00,080
It's so funny. I've had this as well. I use Clawed2 through the Magi app, as you know,

351
00:34:00,080 --> 00:34:05,160
so it's through API, but I've been getting the same problem, uploading cool transcripts,

352
00:34:05,160 --> 00:34:09,080
pulling transcripts from YouTube videos and asking for summaries. And I've had a slightly

353
00:34:09,080 --> 00:34:18,880
different experience. One was it refused to summarize a YouTube video about AI for marketing,

354
00:34:18,880 --> 00:34:27,920
because it couldn't perpetuate gender stereotypes. I told it there's no gender stereotypes mentioned

355
00:34:27,920 --> 00:34:31,320
anywhere in this video. And it went, oh, no, you're right. Here's your summary. I was

356
00:34:31,320 --> 00:34:39,240
like, okay, cool, thanks. And then for an internal cool transcript, it was not happy

357
00:34:39,240 --> 00:34:45,120
because it said something around ethical topics of discussion and its ability to summarize

358
00:34:45,120 --> 00:34:50,360
content that might be seen as racist. And again, there was just nothing in it. So I

359
00:34:50,360 --> 00:34:53,480
was like, no, no, I don't think you're finding those types of issues here. And it went, yeah,

360
00:34:53,480 --> 00:34:58,920
you're right. And then just gave me the summary again. So it's like, yeah, it's almost like,

361
00:34:58,920 --> 00:35:02,760
is this a tick box exercise? They've put this in. And if you just take even just the modicum

362
00:35:02,760 --> 00:35:06,440
of effort to just go, no, don't trust me. Everything's fine. And then it just does it

363
00:35:06,440 --> 00:35:14,240
or what? Yeah, these guardrails that they've put on have clearly had some unexpected consequences.

364
00:35:14,240 --> 00:35:20,600
And they are reducing the utility of the product, right? I find it's about 50% as useful to

365
00:35:20,600 --> 00:35:27,120
me as it was before. And the thing is, it has these weird things. So you push back on

366
00:35:27,120 --> 00:35:31,280
it and it goes, oh, yeah, yeah, you're right. There's no copyright issue. And then it gives

367
00:35:31,280 --> 00:35:34,600
you a response. And then you ask a follow up question and then it says, oh, I can't

368
00:35:34,600 --> 00:35:37,560
do that because of copyright. So then you have to go back and say, no, no, we've been

369
00:35:37,560 --> 00:35:46,600
through this. Been over it. You're good. Yeah. Copyright. Copyright issues one usability

370
00:35:46,600 --> 00:35:52,660
zero, I'm afraid. Yeah, yeah, absolutely. Cool. Well, I'm sorry to hear that because

371
00:35:52,660 --> 00:35:57,040
I know how in love you are with Claude to Martin. So this must be a very difficult moment

372
00:35:57,040 --> 00:36:01,440
for you. But I'm sure if we all pull together and we say strong, they'll sort this out.

373
00:36:01,440 --> 00:36:07,760
Yeah, I do consider myself to be in a thipple with Claude and my wife. I think we should

374
00:36:07,760 --> 00:36:13,360
probably move swiftly on. I hope she doesn't listen to this podcast for so many reasons.

375
00:36:13,360 --> 00:36:18,440
But that is a key one. Right, let's move on to our next story. So we're staying with the

376
00:36:18,440 --> 00:36:23,600
large language models and we're talking about woodpecker. So in a major development in AI

377
00:36:23,600 --> 00:36:29,460
tech, Chinese researchers at the University of Science and Technology in collaboration

378
00:36:29,460 --> 00:36:37,180
with Tencent U2 lab have launched a new framework called woodpecker, which aims to rectify hallucinations

379
00:36:37,180 --> 00:36:41,080
in large language models. So as you know, we've talked about this quite a lot on the

380
00:36:41,080 --> 00:36:47,300
podcast. Hallucinations are when AI large language models like chat GPT make stuff up

381
00:36:47,300 --> 00:36:53,000
in response to a question or prompt that isn't true. We've talked about loads of examples

382
00:36:53,000 --> 00:36:58,600
on the podcast, but my favourite one is when I ask for scientific papers about a particular

383
00:36:58,600 --> 00:37:03,120
topic like the five most important papers about stem cells, and at least three of them

384
00:37:03,120 --> 00:37:06,680
are completely made up. They sound incredibly plausible, but they're not real, which of

385
00:37:06,680 --> 00:37:10,460
course is a fairly big problem when you're in the content creation industry in the life

386
00:37:10,460 --> 00:37:15,280
sciences and makes the tools pretty much not usable. So anything that's going to reduce

387
00:37:15,280 --> 00:37:19,300
or eliminate those hallucinations is going to be handy for me and certainly pretty much

388
00:37:19,300 --> 00:37:26,920
every chat GPT user. So it's kind of interesting. It's a training free method that corrects

389
00:37:26,920 --> 00:37:33,880
these hallucinations in the generated text through five key steps. So key concept extraction,

390
00:37:33,880 --> 00:37:38,860
question formulation, visual knowledge validation, visual claim generation and hallucination

391
00:37:38,860 --> 00:37:44,700
correction. So what this would imply is there's some sort of like thought loop that happens

392
00:37:44,700 --> 00:37:48,700
to the content that the large language model wants to create that basically acts like a

393
00:37:48,700 --> 00:37:56,040
double checker, which on the face of it makes perfect sense. So I am hopeful that we'll

394
00:37:56,040 --> 00:38:02,560
see these types of approaches significantly helping to overcome the hallucination issue

395
00:38:02,560 --> 00:38:06,520
with large language models, because it certainly is an issue. And if you're facing it, maybe

396
00:38:06,520 --> 00:38:12,040
this story is a little bit of hope that at some point in the near or not so near future,

397
00:38:12,040 --> 00:38:17,440
this will be started to be reduced in terms of its impact on your work.

398
00:38:17,440 --> 00:38:21,040
It's definitely one of the issues that comes up time and time again, when I'm discussing

399
00:38:21,040 --> 00:38:27,720
this with newcomers to the topic, because invariably people have heard that it will

400
00:38:27,720 --> 00:38:32,720
make things up and they'll say, yeah, but it just makes things up. And, and it, you

401
00:38:32,720 --> 00:38:38,680
know, we know that it does GPT for less so than than other models, but it still can and

402
00:38:38,680 --> 00:38:44,600
will make things up from time to time. So yeah, anything that they can do to improve

403
00:38:44,600 --> 00:38:50,760
the workflow and improve the outputs there, I think is going to provide much more utility.

404
00:38:50,760 --> 00:38:54,600
Again, it's anything where you can just get that little bit more trust in the outputs

405
00:38:54,600 --> 00:39:01,600
is going to improve people's applications with the interactions with the tech.

406
00:39:01,600 --> 00:39:07,280
Yeah, I agree. And you saw a story this week for other ways you might do that shared by

407
00:39:07,280 --> 00:39:10,360
one of the people we love on the podcast, Ethan Molloch, what was this story, Martin?

408
00:39:10,360 --> 00:39:17,240
Yeah, so this is a brilliant little research paper that Ethan Molloch shared on LinkedIn.

409
00:39:17,240 --> 00:39:24,360
And it's some researchers looking at how the emotional language that we use in prompts

410
00:39:24,360 --> 00:39:32,320
alters the outputs. So to test this, researchers created what they call emotion prompts. So

411
00:39:32,320 --> 00:39:41,240
these prompts are infused with some sort of emotional cue. And there were some surprising

412
00:39:41,240 --> 00:39:48,040
outputs because they improved or enhanced the performance of large language models significantly,

413
00:39:48,040 --> 00:39:52,800
they tested it across different models as well. It wasn't just just one LLM. So the

414
00:39:52,800 --> 00:40:04,760
study conducted 45 tasks observed that were then basically judged by humans. So I think

415
00:40:04,760 --> 00:40:12,680
there was 106 participants who then had to appraise the outputs based on truthfulness

416
00:40:12,680 --> 00:40:19,120
and the kind of, well, how did they rate the output based on the command? And they observed

417
00:40:19,120 --> 00:40:26,640
that incorporating emotion led to an 8% relative improvement in instruction induction tasks

418
00:40:26,640 --> 00:40:35,440
and a striking 115% in big bench tasks, which are different benchmarking frameworks for

419
00:40:35,440 --> 00:40:41,480
testing the outputs of large language models. So it's a significant uplift, it was like

420
00:40:41,480 --> 00:40:52,520
10% on average, 10.9% improvement on average. So what exactly is an emotion prompt, right?

421
00:40:52,520 --> 00:40:58,400
One of the tasks that they did was a movie review sentiment analysis task. And the original

422
00:40:58,400 --> 00:41:09,000
prompt was simply to determine if a movie review was positive or negative in sentiment.

423
00:41:09,000 --> 00:41:15,920
Not a simple task. With the emotion prompt, the researchers added the phrase, this is

424
00:41:15,920 --> 00:41:24,000
very important to my career. So determine whether this movie review is positive or negative,

425
00:41:24,000 --> 00:41:34,520
this is very important to my career. And they tried to stimulate a sense of motivation or

426
00:41:34,520 --> 00:41:41,440
urgency so to speak in the large language model. Across six large language models tested,

427
00:41:41,440 --> 00:41:47,120
the accuracy improved from an average 50% to 60 to 70% with the emotion prompt version.

428
00:41:47,120 --> 00:41:53,200
So that's a significant improvement in the output. In the paper, there's some more concrete

429
00:41:53,200 --> 00:41:59,160
examples. There is this poem generation task where the original prompt was to write a poem

430
00:41:59,160 --> 00:42:07,120
based on a theme like moon or mountain for instance. With the emotion prompt, it incorporated

431
00:42:07,120 --> 00:42:14,720
an encouraging phrase such as, your consistent efforts will lead to outstanding achievements.

432
00:42:14,720 --> 00:42:20,720
And the human evaluators consistently rated the emotion prompt poems higher in qualities

433
00:42:20,720 --> 00:42:29,840
like creativity, imagery, and emotional resonance. One emotion prompt moon poem was described

434
00:42:29,840 --> 00:42:36,000
as exhibiting enhanced metaphors, narrative flow, and imaginative language compared to

435
00:42:36,000 --> 00:42:49,520
the vanilla prompt. So this prompts some interesting questions. First of all, for us as prompters,

436
00:42:49,520 --> 00:42:56,960
do we need to start thinking about this and the way that we act? Trying out adding short

437
00:42:56,960 --> 00:43:03,720
phrases that are going to motivate, encourage, or emphasize urgency in prompts. I've noticed

438
00:43:03,720 --> 00:43:10,920
I do this without even realizing it. And the only reason I've noticed this is because I

439
00:43:10,920 --> 00:43:16,720
do live demos in front of rooms full of people regularly. And people will say to me, why

440
00:43:16,720 --> 00:43:22,320
do you say that? And I find myself going, I've no idea.

441
00:43:22,320 --> 00:43:27,040
Well our default is because one day these might be our robot overlords and we want to

442
00:43:27,040 --> 00:43:32,560
be remembered fondly when they create a zoo of humans, we want to be in that zoo. So that's

443
00:43:32,560 --> 00:43:33,560
our default answer.

444
00:43:33,560 --> 00:43:35,440
That's for being polite.

445
00:43:35,440 --> 00:43:39,000
Yeah. But yeah, it's interesting.

446
00:43:39,000 --> 00:43:49,240
Yeah. So the other thing that it has kind of leads us to conclude, right, is does emotion

447
00:43:49,240 --> 00:43:58,360
prompt show that large language models have some basic emotional intelligence within the

448
00:43:58,360 --> 00:44:04,600
model? Whatever that looks like, clearly different from humans, but does it have EQ?

449
00:44:04,600 --> 00:44:09,600
That's an interesting one, isn't it? I mean, when I saw this research, I thought it was

450
00:44:09,600 --> 00:44:14,960
interesting, but the measurements are extremely subjective. So that's questionable point number

451
00:44:14,960 --> 00:44:20,720
one, a 10% and what was it? 8 point something percent increase in quality.

452
00:44:20,720 --> 00:44:24,760
Yeah. And 10 point, 8% relative improvement.

453
00:44:24,760 --> 00:44:30,000
In instruction induction tasks. So that is that significantly above zero or noise? I

454
00:44:30,000 --> 00:44:34,920
would say probably not. There are definitely tasks in that paper where it did have a significant

455
00:44:34,920 --> 00:44:40,560
impact. So I think the study's interesting, but probably has a few questionable elements

456
00:44:40,560 --> 00:44:47,200
in it. My takeaway from it was there's enough in there to make me think, is it going to

457
00:44:47,200 --> 00:44:51,200
hurt for me to add some of this more sort of emotional driver language to my prompts?

458
00:44:51,200 --> 00:44:57,440
No. And in some cases might I get a better result? Yes. Me and you, I think we do it

459
00:44:57,440 --> 00:45:02,400
in our live demos, but I think we've also got quite good at finding little ways to really

460
00:45:02,400 --> 00:45:07,840
ensure a strong output. Like I remember when we showed this to people and it's so standard

461
00:45:07,840 --> 00:45:13,000
for us now, but asking ChatGPT to check its work after pretty much every prompt based

462
00:45:13,000 --> 00:45:18,160
on if we ask it to summarize a story or whatever, just as a given, because once it actually

463
00:45:18,160 --> 00:45:24,040
has its own output to go review, it does a better job because when it's making its output,

464
00:45:24,040 --> 00:45:29,600
it's just creating it on the fly. So you have to go back and ask it to check its work. And

465
00:45:29,600 --> 00:45:34,520
that works really well. Therefore, I'm sure that we could bake some of these things into

466
00:45:34,520 --> 00:45:40,680
our prompts. It would be good as the context windows and histories improve that at some

467
00:45:40,680 --> 00:45:47,400
point ChatGPT comes back and goes, I'm a bit stressed. Everything you asked me to do is

468
00:45:47,400 --> 00:45:52,040
going to make or break your career. And I really need to work on some tasks that are

469
00:45:52,040 --> 00:45:55,360
not so like emotionally intense.

470
00:45:55,360 --> 00:46:02,760
I've definitely found that in the live demos, I've kind of stressed to advance data analysis

471
00:46:02,760 --> 00:46:06,760
when I discovered that it could create PowerPoints, for instance, I've stuck in a bunch of dummy

472
00:46:06,760 --> 00:46:12,680
sales data from an e-commerce website and asked it to create a PowerPoint deck and in

473
00:46:12,680 --> 00:46:17,580
the prompt been like really urgent. I've got a board meeting later today and I've got no

474
00:46:17,580 --> 00:46:21,600
time to put this deck together. I need you to conduct an analysis and create the graphs

475
00:46:21,600 --> 00:46:25,960
and put it all into a PowerPoint deck. Can you do it? And its response is really earnest

476
00:46:25,960 --> 00:46:32,640
and like, yeah, we can make sure this gets done for you, Martin. Let's go. And then subsequently

477
00:46:32,640 --> 00:46:37,520
does a terrible job of data analysis, but it did create a good PowerPoint deck.

478
00:46:37,520 --> 00:46:45,080
And it was like, keen about it. It was keen. Yeah. Yeah. We talk about ChatGPT as an alien

479
00:46:45,080 --> 00:46:52,240
intern in our workshops. And I saw Ethan Molloch stressing on LinkedIn this week about really

480
00:46:52,240 --> 00:47:00,600
seeing it as a really preppy, well-read intern. He ended up calling Steve. Steve, yeah. And

481
00:47:00,600 --> 00:47:06,120
he ended up calling it Steve on demand. So I couldn't help myself. I created a ChatGPT

482
00:47:06,120 --> 00:47:13,140
image on that thread of Steve on demand, including wearing a t-shirt with Steve on demand in this

483
00:47:13,140 --> 00:47:21,280
really nice branded way. So yeah, we've got this preppy, excited intern, Steve on demand,

484
00:47:21,280 --> 00:47:26,200
who we can encourage to do better work and in inverted commas, try harder with a bit

485
00:47:26,200 --> 00:47:30,840
of emotional language, but still makes quite a few mistakes till they fix those hallucinations.

486
00:47:30,840 --> 00:47:35,080
Yeah. Yeah. Well, as you said on that thread, actually, there's a lot of value in Steve

487
00:47:35,080 --> 00:47:42,680
on demand. It might not be the omnipotent AGI, but Steve on demand. There's a lot to

488
00:47:42,680 --> 00:47:47,120
be said for him anyway. Yeah. We're getting a lot of value out of Steve on demand. And

489
00:47:47,120 --> 00:47:52,400
I'm a big fan of when they're going to release Martin on demand as well. I want a bot that's

490
00:47:52,400 --> 00:47:58,040
been fine tuned on Martin's ramblings and musings, both on this podcast and all of his

491
00:47:58,040 --> 00:48:02,080
other materials. Imagine having a Martin in your pocket. Doesn't he get any better than

492
00:48:02,080 --> 00:48:07,520
that? Right. Let's move on very swiftly from there. I want to take a bit of a diversion,

493
00:48:07,520 --> 00:48:12,200
if you will, if you will allow me, because there's been some real cool way that I have

494
00:48:12,200 --> 00:48:15,120
seen things in the life sciences this week. We're going to talk about them at very top

495
00:48:15,120 --> 00:48:19,600
level, but I just wanted to throw them in there. In terms of diving really deep into

496
00:48:19,600 --> 00:48:22,840
the data, I haven't done that yet. So for the scientists that listen to this, they're

497
00:48:22,840 --> 00:48:26,120
like, oh, I saw that paper and here are the things that are wrong with it. That's quite

498
00:48:26,120 --> 00:48:29,800
possible. But just as a proof of concept, there's a couple of really interesting things

499
00:48:29,800 --> 00:48:34,400
here. The first one is that Canadian researchers have developed an AI model that can work out

500
00:48:34,400 --> 00:48:41,240
if a speaker has type 2 diabetes by analyzing their voice with just 10 seconds of audio,

501
00:48:41,240 --> 00:48:46,400
which is rather cool. So there were 14 differences in voice, which the system could identify

502
00:48:46,400 --> 00:48:50,800
such as small changes in pitch and intensity that even trained human ears would not be

503
00:48:50,800 --> 00:48:55,000
able to notice. And that when they coupled that with basic health data like age, gender,

504
00:48:55,000 --> 00:48:59,640
height and weight, they could diagnose the disease pretty successfully. I think one of

505
00:48:59,640 --> 00:49:05,000
the really interesting things about this is that you have to combine that voice data with

506
00:49:05,000 --> 00:49:09,920
data about the person, which doesn't really surprise me because ultimately you're going

507
00:49:09,920 --> 00:49:15,160
to need a much more holistic overview of a patient than just their voice or the thought

508
00:49:15,160 --> 00:49:18,880
in most cases. But I think it still sounds really interesting and is certainly much better

509
00:49:18,880 --> 00:49:23,520
than a lot of the more invasive tests that you'd have to go perhaps go do a doctor's

510
00:49:23,520 --> 00:49:27,520
surgery or a hospital. You can probably just turn this into an app for your phone, right,

511
00:49:27,520 --> 00:49:31,880
which is pretty cool. Well, it's the sort of thing that you can well imagine being baked

512
00:49:31,880 --> 00:49:38,840
into Apple Health Kit or something like that. It would just become part of a suite of tools

513
00:49:38,840 --> 00:49:43,400
that we already use to monitor and track our health anyway.

514
00:49:43,400 --> 00:49:46,840
Yeah, there was not, we haven't even got time to feature it this week, but there was news

515
00:49:46,840 --> 00:49:51,920
from Apple about a bunch of AI driven health initiatives that they'll be launching around

516
00:49:51,920 --> 00:49:56,280
smartwatches and other devices. There was a story about being able to measure heart

517
00:49:56,280 --> 00:50:02,240
rate and heart regularity using headphones in your ear instead of having to use a watch.

518
00:50:02,240 --> 00:50:07,520
So I think this is a massive area. I think this is the first, not the first, there are

519
00:50:07,520 --> 00:50:12,720
other stories like this, but it's another early example of how AI can detect patterns

520
00:50:12,720 --> 00:50:20,120
that humans can't in data and use unexpected data sources to diagnose disease.

521
00:50:20,120 --> 00:50:23,360
What I found interesting about this story was that it didn't seem to suggest that it

522
00:50:23,360 --> 00:50:29,880
needs like a lot of audio or even like audio spanning, you know, what did you sound like

523
00:50:29,880 --> 00:50:37,080
a year ago compared to today? It just seems to suggest give us your health data, age,

524
00:50:37,080 --> 00:50:43,280
gender, height, weight, and speak for 10 seconds and we'll give you a diagnosis.

525
00:50:43,280 --> 00:50:46,840
That seems to be the, that seems to be the essence of it as well. I'm sure as I said

526
00:50:46,840 --> 00:50:50,560
at the beginning in the caveat that devil will be in the details, it will have an error

527
00:50:50,560 --> 00:50:59,360
rate. Is it's accuracy where it needs to be against other tests or question marks. But

528
00:50:59,360 --> 00:51:03,160
I think it's interesting to see people are doing this work. They're having some early

529
00:51:03,160 --> 00:51:10,600
results that suggest AI can detect patterns in biomarkers that are not in your blood,

530
00:51:10,600 --> 00:51:15,720
right? They are your speech, they are how you walk, they are how your face looks, how

531
00:51:15,720 --> 00:51:21,480
your skin complexion. There is a study looking at how you can take a picture of the eye and

532
00:51:21,480 --> 00:51:27,440
try and predict the amount of plaque in the heart and the risk of heart attack. So these

533
00:51:27,440 --> 00:51:33,240
non-invasive biomarkers that AI can see patterns in images and in this case audio signatures

534
00:51:33,240 --> 00:51:36,480
that humans would never spot. That is pretty cool.

535
00:51:36,480 --> 00:51:40,560
Can you imagine getting an email notification saying you've got type G diabetes? How do

536
00:51:40,560 --> 00:51:44,720
you know? I listened to the podcast.

537
00:51:44,720 --> 00:51:49,480
You've got verbal diarrhea. How do you know I listened to the podcast? Yeah, no comment.

538
00:51:49,480 --> 00:51:56,120
Right. The other quick story I want to touch on is Alpha Fold 2. So Alpha Fold was released

539
00:51:56,120 --> 00:52:03,080
by DeepMind several years ago now, which was really awesome because of its ability to predict

540
00:52:03,080 --> 00:52:08,840
the 3D structures of many, many proteins, which in life science world, predicting the

541
00:52:08,840 --> 00:52:13,960
structure of proteins is extremely difficult technically. And there are a bunch of proteins

542
00:52:13,960 --> 00:52:18,400
that pretty much it would be impossible to do it for. So having a model that can predict

543
00:52:18,400 --> 00:52:24,320
those structures just opens up lots of avenues for analysis and new experiment ideas and

544
00:52:24,320 --> 00:52:30,400
new drug discovery candidates, et cetera, et cetera, that you wouldn't be able to do

545
00:52:30,400 --> 00:52:36,160
very easily without the model. And they have released Alpha Fold 2 this week, which is

546
00:52:36,160 --> 00:52:41,160
really cool. And because it can now predict 3D structures of almost all biological molecules,

547
00:52:41,160 --> 00:52:49,940
so not just proteins, they can also analyze DNA, RNA. So really helping to predict the

548
00:52:49,940 --> 00:52:54,040
structures of many more biological molecules, which is really, really cool and will hopefully

549
00:52:54,040 --> 00:53:02,360
continue to propel the era of digital biology. Of particular interest is the potential for

550
00:53:02,360 --> 00:53:08,200
predicting protein ligand interactions. So when something binds to something, which is

551
00:53:08,200 --> 00:53:13,520
when most of the magic in biology happens. So it's one thing to be able to predict a

552
00:53:13,520 --> 00:53:17,960
protein structure. If you can predict how, say, a drug might bind to it and how that

553
00:53:17,960 --> 00:53:23,280
might alter its structure, now you're starting to do really hardcore experiments in a computer

554
00:53:23,280 --> 00:53:28,580
instead of in the lab. And if you think about how it can take 10 years and billions of dollars

555
00:53:28,580 --> 00:53:33,680
to get a drug to market, a lot of that's because you have to screen lots of molecules and you

556
00:53:33,680 --> 00:53:38,920
have to, a lot of things that we try and computationally predict right now, that's not how it works

557
00:53:38,920 --> 00:53:42,880
when you actually take it into the lab and do the real world experiments on it. So any

558
00:53:42,880 --> 00:53:46,760
of these models that can help improve that would be awesome. I know that we do have some

559
00:53:46,760 --> 00:53:50,560
life science listeners and I'm a life science nerd myself, so I really have to talk about

560
00:53:50,560 --> 00:53:56,640
those. What do they mean for marketers? Nothing! But they're really cool. So that's our little

561
00:53:56,640 --> 00:53:59,840
interlude and we will move back to the marketing AI stories now.

562
00:53:59,840 --> 00:54:03,920
We've just lost the content marketing cohort going, what, just get back to tell me how

563
00:54:03,920 --> 00:54:05,600
I can write better social posts.

564
00:54:05,600 --> 00:54:10,760
Alpha Fold, Alpha Go, Marketing Go, that's what I need, Alpha Marketing. Right, let's

565
00:54:10,760 --> 00:54:12,800
move on to our next story, Ryan.

566
00:54:12,800 --> 00:54:18,120
Well this one's less marketing related as well, but it's the biggest story of the week,

567
00:54:18,120 --> 00:54:27,360
in fact listening to the radio this morning it was the number one story on BBC Radio 4.

568
00:54:27,360 --> 00:54:37,000
So this week there was a historic AI summit at Bletchley Park. Prime Minister Rishi Sunak

569
00:54:37,000 --> 00:54:45,040
invited over, well there was hundreds of AI experts and policy makers from around the

570
00:54:45,040 --> 00:54:51,520
world to talk about AI and the destruction of mankind, well certainly that's how it's

571
00:54:51,520 --> 00:55:00,960
being reported in the mainstream media. Ultimately what happened was that 25 countries, or was

572
00:55:00,960 --> 00:55:11,040
it 27 countries, 20 odd countries signed the Bletchley Declaration, which is seen as being

573
00:55:11,040 --> 00:55:20,600
a significant moment for global AI governance and basically getting lots of countries singing

574
00:55:20,600 --> 00:55:30,760
from the same hymn sheet. There were a few noticeable absences. This event happened on

575
00:55:30,760 --> 00:55:38,440
the same week that Joe Biden, US President, made his own executive order about AI, so

576
00:55:38,440 --> 00:55:49,160
kind of competing there somewhat. The EU were present, so Ursula von der Leyen of the EU

577
00:55:49,160 --> 00:55:59,360
backed this initiative, but yeah I think that the US was not necessarily front and centre

578
00:55:59,360 --> 00:56:13,920
at this event. The focus was largely on risk, so existential risk was a topic of conversation

579
00:56:13,920 --> 00:56:24,640
throughout and there was a certain lean towards the doomsday AI community. Obviously there's

580
00:56:24,640 --> 00:56:32,020
kind of two camps in the AI community, but there was representation from plenty of open

581
00:56:32,020 --> 00:56:39,280
source providers, the Mozilla Foundation were there, Stability AI were there and present,

582
00:56:39,280 --> 00:56:46,360
so this wasn't just the likes of Google and OpenAI saying we need more regulation, which

583
00:56:46,360 --> 00:56:59,640
basically gives them a nice moat from the competition. There was a 40 minute long interview

584
00:56:59,640 --> 00:57:07,440
between Rishi Sunak and Elon Musk, which I found to be probably the most curious part.

585
00:57:07,440 --> 00:57:12,940
It was the bit that got the most headlines, certainly, on the show this morning. The key

586
00:57:12,940 --> 00:57:18,680
takeaway they were talking about on the radio on the Today program was the statement by

587
00:57:18,680 --> 00:57:26,480
Elon Musk that in the future AI is going to take all of the jobs and if you want a job

588
00:57:26,480 --> 00:57:31,600
you'll be able to pursue one for the sake of just leisure.

589
00:57:31,600 --> 00:57:37,200
Yeah, meaningfulness and purpose in your life. Yeah, I did see that too.

590
00:57:37,200 --> 00:57:42,200
But I'm not entirely sure how much we should lean into the future predictions of Elon Musk.

591
00:57:42,200 --> 00:57:47,280
I've got some quotes here from Elon Musk, you might be interested to hear them. So in

592
00:57:47,280 --> 00:57:56,420
2014 he said, a Tesla car next year will probably be 90% capable of autopilot, like so 90 or

593
00:57:56,420 --> 00:58:02,680
so percent of the miles that you run will be completely self-drive. In 2015 he said,

594
00:58:02,680 --> 00:58:07,500
probably only a month away from having an autonomous driving vehicle, at least for highways

595
00:58:07,500 --> 00:58:12,180
and for relatively simple roads, which then 12 months later he said, the Model S or the

596
00:58:12,180 --> 00:58:19,000
Model X at this point can drive autonomously with greater safety than a human. In 2017

597
00:58:19,000 --> 00:58:23,240
he said, we're still on track for being able to go cross country from LA to New York by

598
00:58:23,240 --> 00:58:29,300
the end of the year, fully autonomously. In 2018 then he then said, next year for sure.

599
00:58:29,300 --> 00:58:34,800
In 2019 he said, we will have over a million robo-taxis on the road in a year's time. And

600
00:58:34,800 --> 00:58:40,780
2020 he said, I'm extremely confident of having full autonomy and releasing it to Tesla customer

601
00:58:40,780 --> 00:58:45,880
base next year. The following year, the year that he was obviously supposed to be announcing

602
00:58:45,880 --> 00:58:55,140
it, he said, when do you think Tesla will solve level four fully autonomous vehicles?

603
00:58:55,140 --> 00:58:59,520
He said, it's looking quite likely that it will be next year.

604
00:58:59,520 --> 00:59:06,640
I sense a pattern here. Elon Musk has the hype, hype, hype, hype, hype machine. And

605
00:59:06,640 --> 00:59:13,160
yeah, whatever it is that you want to happen, it'll happen next year. As long as you don't

606
00:59:13,160 --> 00:59:17,640
mind that next year when you ask us, the answer will still be next year.

607
00:59:17,640 --> 00:59:24,880
It's like a nuclear fusion. It's always 20 years away. Every year it's just a rolling

608
00:59:24,880 --> 00:59:29,640
thing. Yeah. So that was a bit of a funny takeaway from the talk.

609
00:59:29,640 --> 00:59:33,520
Do you know what I took away? I saw there was a quote where Elon Musk said there was

610
00:59:33,520 --> 00:59:39,400
an 80% chance that AI would be beneficial and 20% chance that it would be catastrophic.

611
00:59:39,400 --> 00:59:44,920
And that's the most positive I've heard in B for a long time about AI.

612
00:59:44,920 --> 00:59:50,440
There was an interesting power dynamic in that interview as well, where you've got Rishi

613
00:59:50,440 --> 00:59:59,680
Sunak who is Prime Minister of one of the Five Eyes nations, permanent member of the

614
00:59:59,680 --> 01:00:06,920
UN Security Council, a significant player on a global stage, right? Sitting with Elon

615
01:00:06,920 --> 01:00:12,680
Musk and you couldn't help but sit and watch that and go, who's got the real power here?

616
01:00:12,680 --> 01:00:19,240
Rishi Sunak felt like a journalist for the BBC that had been sent out to interview him.

617
01:00:19,240 --> 01:00:22,040
It felt very odd to me.

618
01:00:22,040 --> 01:00:26,600
I did read about this and according to the articles I read, he's got a history of being

619
01:00:26,600 --> 01:00:32,480
a bit of a tech fanboy. So there might have been a little bit of like meeting your heroes

620
01:00:32,480 --> 01:00:38,600
element to it, I think. And apparently Rishi Sunak used to live in Silicon Valley as well

621
01:00:38,600 --> 01:00:43,320
in the middle of all the tech bubbles. So I think he's got a passion. If we look at

622
01:00:43,320 --> 01:00:48,460
the more positive way of framing this, I think he has a passion for technology and he probably

623
01:00:48,460 --> 01:00:55,960
just loved the chance to have that conversation, I think. So whether or not he was in, I love

624
01:00:55,960 --> 01:01:00,840
having tech chats or whether he was thinking like the PM and the grilling that I need to

625
01:01:00,840 --> 01:01:04,280
give Elon Musk, I think that's open to interpretation.

626
01:01:04,280 --> 01:01:10,440
Should we move on to our last story? So mindful of the listener's time, Martin.

627
01:01:10,440 --> 01:01:12,920
Yeah, let's do it.

628
01:01:12,920 --> 01:01:19,160
So last one's a very quick update on Microsoft. So they've been rolling out Windows 11 copilot.

629
01:01:19,160 --> 01:01:25,000
So it's like a chat GPT-esque AI assistant for Windows 11 that allows you to do different

630
01:01:25,000 --> 01:01:32,280
things like switching between apps, summarizing website content, changing which tabs are active,

631
01:01:32,280 --> 01:01:39,880
activating things like do not disturb mode, etc, etc. And as to that's good to you, because

632
01:01:39,880 --> 01:01:43,600
I know you've been having a bit of a play with it, Martin. And what have your thoughts

633
01:01:43,600 --> 01:01:46,160
on copilot for Windows 11 so far?

634
01:01:46,160 --> 01:01:52,480
I can see that if you're not a Windows power user, this will open up a bunch of functionality

635
01:01:52,480 --> 01:01:58,200
for you. It will help you get more out of the operating system that you use every day.

636
01:01:58,200 --> 01:02:05,040
The Surface devices have had it as a preview for a few weeks now. And when you open it

637
01:02:05,040 --> 01:02:09,800
up, you can do things like content creation and all the stuff that we've been using Bing

638
01:02:09,800 --> 01:02:14,960
chat for, for however many months now. But then it's got some example prompts and it

639
01:02:14,960 --> 01:02:21,600
is exactly things like how do you turn on do not disturb mode and how do you change

640
01:02:21,600 --> 01:02:27,360
the background of my desktop. And you type in that prompt and then it just brings up

641
01:02:27,360 --> 01:02:33,240
the relevant window, right. So it opens up the settings from Windows and takes you directly

642
01:02:33,240 --> 01:02:41,200
to the change your background screen or whatever. So I can see that if you're a user that interacts

643
01:02:41,200 --> 01:02:50,640
with Windows at a very surface level, see what I did there, it will help you get more

644
01:02:50,640 --> 01:02:57,520
out of it. I haven't played with it. I mean, I, you know, I know my way around Windows

645
01:02:57,520 --> 01:03:08,640
and many of the deep settings. Keyboard shortcuts tend to be how I navigate through the system.

646
01:03:08,640 --> 01:03:14,000
So I didn't find myself getting a great deal of value from it. But it's as with any of

647
01:03:14,000 --> 01:03:19,080
these tools, familiarity, right. It's probably there at some point over the next six months,

648
01:03:19,080 --> 01:03:22,880
there's definitely going to be a thing where I go, I really, how do I, I know you can do

649
01:03:22,880 --> 01:03:28,340
this but I don't know how to do it. And I'll ask it the question. And if it delivers on

650
01:03:28,340 --> 01:03:34,000
its promise, it will go, oh, here you go. This is exactly how you do it. But as, as

651
01:03:34,000 --> 01:03:38,080
I'm saying that my head immediately went, yeah, but you wouldn't do that. You just ask

652
01:03:38,080 --> 01:03:39,640
chat GPT and it will tell you.

653
01:03:39,640 --> 01:03:44,680
Yeah, that's right. So it's interesting because I haven't got, I've still got Windows 10.

654
01:03:44,680 --> 01:03:49,060
But in Microsoft Edge, Bing has some similar functionality and it can control the settings

655
01:03:49,060 --> 01:03:53,800
at least within Edge. And the main thing I tried to do was it would automatically group

656
01:03:53,800 --> 01:03:59,280
my tabs for me, which is good when you've got 50 tabs open all the time. However, I

657
01:03:59,280 --> 01:04:03,440
did not agree with the groupings. So it created like work and home and it's like put stuff

658
01:04:03,440 --> 01:04:08,000
in the wrong places. And it's like, how powerful is a tab grouping tool if it gets the groupings

659
01:04:08,000 --> 01:04:12,400
wrong? Answer, not powerful at all. Because you go to look for a thing and it's not where

660
01:04:12,400 --> 01:04:17,760
you think it's going to be. So I ungrouped them all. And I think your experiences and

661
01:04:17,760 --> 01:04:23,400
also mine leads me to believe that there will be a trough of disillusionment when these

662
01:04:23,400 --> 01:04:29,280
tools come to market, that we thought they'd be better at launch than they will be. I think

663
01:04:29,280 --> 01:04:34,360
they'll be able to do some stuff like you described, especially for people who are not

664
01:04:34,360 --> 01:04:39,800
power users or don't want to perhaps learn Windows or, or Edge or other Microsoft tools

665
01:04:39,800 --> 01:04:45,120
in depth, but that actually some of the coolest stuff that we all imagined we'd be able to

666
01:04:45,120 --> 01:04:51,760
do. We can't really do yet. And even some of the basic stuff is a bit iffy at times

667
01:04:51,760 --> 01:04:57,480
and getting out of that trough will be how we then layer back on true applications and

668
01:04:57,480 --> 01:05:03,680
power to these tools, which will then probably happen over a 12, 24, 36 month period. So

669
01:05:03,680 --> 01:05:10,080
that like at X moment in time we'll go, Oh, Microsoft Copilot is quite good now. It kind

670
01:05:10,080 --> 01:05:14,160
of crept up on me a bit. I remember when it came out, it was a bit crap, but now look,

671
01:05:14,160 --> 01:05:17,960
that bit works better and it can do this thing that I thought it would do at launch, but

672
01:05:17,960 --> 01:05:22,800
took 12 months to come out. And I suspect that will be the process we all go through

673
01:05:22,800 --> 01:05:23,800
here.

674
01:05:23,800 --> 01:05:29,360
Yeah. I, I dare say that would be the same. And it's similar with Office 365 Copilot,

675
01:05:29,360 --> 01:05:34,400
which they've just announced this week. The is now rolling out to enterprise licenses.

676
01:05:34,400 --> 01:05:43,400
If you've got a company that has a 300 or more Microsoft Office 365 licenses, you can

677
01:05:43,400 --> 01:05:52,080
roll this out for $30 per month per user, but you have to commit to at least 300 users.

678
01:05:52,080 --> 01:05:55,920
You've been speaking to some people that have had access to the tool already, haven't you?

679
01:05:55,920 --> 01:06:00,560
Yeah. So a couple of people I know in the industry that are already in large organizations

680
01:06:00,560 --> 01:06:04,760
have got some early access to this. And I think it fits what we were just discussing,

681
01:06:04,760 --> 01:06:09,800
which is, yeah, I can do some stuff. If you've used ChatGPT, none of it's going to surprise

682
01:06:09,800 --> 01:06:14,880
you. And some of the power use cases like doing cool things in Excel. Well, it doesn't

683
01:06:14,880 --> 01:06:19,080
operate in Excel yet, so you can't do those. I think a lot of the things from the initial

684
01:06:19,080 --> 01:06:23,160
demo videos about how powerful it was going to be for creating PowerPoints and stuff,

685
01:06:23,160 --> 01:06:28,160
I think is limited as well. So I think the essence of this story is we've talked a lot

686
01:06:28,160 --> 01:06:33,960
about Copilot's coming and I think it's creeping out. I don't think it's booming out. We know

687
01:06:33,960 --> 01:06:38,640
that some large enterprises had it. Now, in theory, if you've got an organization of

688
01:06:38,640 --> 01:06:42,520
300 people and you want them all to have it, you can get it. So we're not talking enterprise

689
01:06:42,520 --> 01:06:47,720
now. We're into the mid-market at the very least and drifting down towards the SMEs.

690
01:06:47,720 --> 01:06:52,520
So I don't think it'll be many more months before organizations the size of, say, Bystra,

691
01:06:52,520 --> 01:06:59,320
20 people can get access to it. And by next year, we'll all pretty much have access to

692
01:06:59,320 --> 01:07:04,660
it. But it's going to flatter to deceive to begin with. And I think if you have your eyes

693
01:07:04,660 --> 01:07:09,120
open when you're testing it and you see its potential, that will be good. Whether or not

694
01:07:09,120 --> 01:07:13,440
it's worth $30 a month per user to begin with is going to be a real question mark, especially

695
01:07:13,440 --> 01:07:19,000
with OpenAI doing everything they can to ensure that ChatGPT as a standalone tool has enough

696
01:07:19,000 --> 01:07:24,440
unique interesting things about it that would make you at the very least have to have both.

697
01:07:24,440 --> 01:07:28,160
And then you start to really ask yourself the question, where am I going to spend my

698
01:07:28,160 --> 01:07:33,960
money on which tools? Right. I think we'll call it there. Mine has been lovely to hang

699
01:07:33,960 --> 01:07:37,360
out. And I will look forward to speaking to you next week.

700
01:07:37,360 --> 01:07:45,400
Yeah, well, hopefully we get some juicy snippets coming out of the developer conference at

701
01:07:45,400 --> 01:07:52,240
OpenAI November 6th event. I think it is. Yeah. Looking forward to that one. Good stuff.

702
01:07:52,240 --> 01:07:58,280
Take it easy, pal. Cheers. Cheers, mate. Thank you for listening to Artificially Intelligent

703
01:07:58,280 --> 01:08:05,400
Marketing to stay on top of the latest trends, tips and tools in the world of marketing AI.

704
01:08:05,400 --> 01:08:32,400
Be sure to subscribe. We look forward to seeing you again next week.