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 --> 00:00:18,540
results from your marketing efforts.

4
00:00:18,540 --> 00:00:23,120
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:23,120 --> 00:00:25,360
Welcome to Artificially Intelligent Marketing.

6
00:00:25,360 --> 00:00:27,960
Here we are for episode 17.

7
00:00:27,960 --> 00:00:28,960
Thank you for joining us.

8
00:00:28,960 --> 00:00:30,440
I'm here with my good friend Martin.

9
00:00:30,440 --> 00:00:34,120
Martin, how are you and what have you been up to this week?

10
00:00:34,120 --> 00:00:37,600
I am bright and breezy today.

11
00:00:37,600 --> 00:00:43,440
This week I have been preaching from the good book Generative AI.

12
00:00:43,440 --> 00:00:50,760
I was at a digital marketing conference over in Leicester, about 300 people in the room

13
00:00:50,760 --> 00:00:59,280
and I was doing a session about how to basically prepare for a future powered by AI when it

14
00:00:59,280 --> 00:01:03,360
comes to your marketing, all of these new AI tools and what should people know?

15
00:01:03,360 --> 00:01:06,840
A couple of interesting things from that talk, Paul.

16
00:01:06,840 --> 00:01:11,080
First one was that at the same conference last year, I asked people how many people

17
00:01:11,080 --> 00:01:12,080
had used AI.

18
00:01:12,080 --> 00:01:15,920
About 10% of the people put their hand up last year.

19
00:01:15,920 --> 00:01:21,080
This year, I would say 95% of the room had done.

20
00:01:21,080 --> 00:01:25,520
It was overwhelmingly a sea of hands went up.

21
00:01:25,520 --> 00:01:30,960
The vast majority chat GPT and specifically the free version of chat GPT.

22
00:01:30,960 --> 00:01:36,080
So not many people get in their hands on GPT-4 just yet.

23
00:01:36,080 --> 00:01:42,720
But the other thing that came out of that conference was a little bit of presentation

24
00:01:42,720 --> 00:01:43,720
magic, Paul.

25
00:01:43,720 --> 00:01:52,480
What I decided to do with the talk was I wanted to incorporate some audience interaction.

26
00:01:52,480 --> 00:01:58,640
So the start of the talk, I was going through a few stats about where the state of AI at

27
00:01:58,640 --> 00:02:03,560
the moment and the McKinsey report that said about four trillion being added to the global

28
00:02:03,560 --> 00:02:06,600
GDP from Generative AI.

29
00:02:06,600 --> 00:02:11,440
But I also asked people in the audience who had used chat GPT and what their use cases

30
00:02:11,440 --> 00:02:12,440
were.

31
00:02:12,440 --> 00:02:18,080
The members in the audience told me their name, told me where the business was.

32
00:02:18,080 --> 00:02:21,680
I made sure to repeat this very loud for the rest of the room.

33
00:02:21,680 --> 00:02:25,580
So I would say, this is Mohammed and you're a graphic designer and you've been using it

34
00:02:25,580 --> 00:02:27,680
to create press releases and how has that been?

35
00:02:27,680 --> 00:02:29,800
He told me the good and the bad.

36
00:02:29,800 --> 00:02:31,840
I did that with another audience member as well.

37
00:02:31,840 --> 00:02:39,240
As the talk went on, I got to the end of the session and nobody in the room saw this coming.

38
00:02:39,240 --> 00:02:44,360
I said, you really want to experiment with these tools, particularly trying out something

39
00:02:44,360 --> 00:02:51,280
like chat GPT with Zapier now, obviously we've played about with the great open AI and Zapier

40
00:02:51,280 --> 00:02:57,080
integration, which gives you access to whisper the audio transcription tool.

41
00:02:57,080 --> 00:03:05,280
So what I'd done is I'd pre-loaded a file in Google Drive to when a new file goes into

42
00:03:05,280 --> 00:03:09,760
that drive, goes into whisper and a sequence of steps occurs.

43
00:03:09,760 --> 00:03:16,880
Long story short, taking the audio clip, transcribing that, turning the transcription into a blog

44
00:03:16,880 --> 00:03:25,040
post and publishing it in real time on medium.com on a medium blog with an image created from

45
00:03:25,040 --> 00:03:30,360
Dali again, that's taken from the audio prompt as well.

46
00:03:30,360 --> 00:03:32,360
When I did it, I had no idea if it was going to work.

47
00:03:32,360 --> 00:03:38,120
The night before I tested it and it failed and I tested it three or four times before

48
00:03:38,120 --> 00:03:41,520
and it had worked, but this one time it had failed on me.

49
00:03:41,520 --> 00:03:44,160
I thought, oh dear, this is bad.

50
00:03:44,160 --> 00:03:48,200
There was no one in the audience that knew that I was doing, so none of the tech team

51
00:03:48,200 --> 00:03:51,160
could give me a kind of thumbs up to tell me that it had worked.

52
00:03:51,160 --> 00:03:55,200
There was going to be a big reveal.

53
00:03:55,200 --> 00:03:58,980
Anyway, so I said at the end, I said, oh, and you can even, I said to the audience,

54
00:03:58,980 --> 00:04:05,160
you can even use this to write and publish a blog post without even touching a keyboard.

55
00:04:05,160 --> 00:04:11,360
I got my medium blog up and fortunately for me, there was a blog.

56
00:04:11,360 --> 00:04:16,200
So I thought that great take, hurdle number one overcome.

57
00:04:16,200 --> 00:04:21,520
Started reading through it and no one really understood the kind of relevance of this.

58
00:04:21,520 --> 00:04:26,720
It all seemed a bit generic until I got to two paragraphs where it mentioned the business

59
00:04:26,720 --> 00:04:32,320
owners I'd spoken to it had their names, the company they were from and their use cases.

60
00:04:32,320 --> 00:04:38,120
When I read that out, there was an audible gasp in the room as people were like, how

61
00:04:38,120 --> 00:04:39,120
has he done that?

62
00:04:39,120 --> 00:04:40,120
That is magic.

63
00:04:40,120 --> 00:04:44,000
So yeah, that was deeply satisfying.

64
00:04:44,000 --> 00:04:49,080
As someone who's done a little bit of standup comedy in years gone by, great when you can

65
00:04:49,080 --> 00:04:54,000
get a laugh from an audience, but getting an audible gasp as people were just shocked

66
00:04:54,000 --> 00:04:56,120
was a good moment.

67
00:04:56,120 --> 00:04:57,480
That is awesome.

68
00:04:57,480 --> 00:05:03,640
So you recorded audio while you were speaking, stopped the audio at some point through it

69
00:05:03,640 --> 00:05:08,120
and then that triggered the rest of the workflow while you just finished your talk and then

70
00:05:08,120 --> 00:05:12,520
went back to it at the end and it was all written as a blog post for you.

71
00:05:12,520 --> 00:05:13,520
Yeah.

72
00:05:13,520 --> 00:05:17,520
And then that was the most nerdy thing because I had it recording in my pocket and then I

73
00:05:17,520 --> 00:05:22,320
went over to the lectern, had to kind of sneak out my phone on the lectern, put it in front

74
00:05:22,320 --> 00:05:25,320
of me and I had to press stop, upload to Google Drive.

75
00:05:25,320 --> 00:05:29,360
I knew that it was only four clicks and it was fairly straightforward.

76
00:05:29,360 --> 00:05:35,160
I drilled it a few times, but you know, 5G could have dropped, like the signal in the

77
00:05:35,160 --> 00:05:37,760
room could have been terrible and it might not have uploaded the file.

78
00:05:37,760 --> 00:05:40,480
And there were any number of things that could have gone wrong and I could have looked like

79
00:05:40,480 --> 00:05:43,600
a bit of a muck at the end, but fortunately I didn't.

80
00:05:43,600 --> 00:05:44,680
I love that.

81
00:05:44,680 --> 00:05:45,680
Great story.

82
00:05:45,680 --> 00:05:48,680
I was at Vistage this week.

83
00:05:48,680 --> 00:05:52,840
I think I was probably one of the more exciting things that happened to me this week.

84
00:05:52,840 --> 00:05:57,920
I'm a member of Vistage, which is like a CEO, peer learning network.

85
00:05:57,920 --> 00:06:03,520
It's been hugely valuable to our business and certainly my own personal development.

86
00:06:03,520 --> 00:06:09,120
So if you're a managing director or a CEO out there who is looking for support and places

87
00:06:09,120 --> 00:06:12,280
to balance ideas off and training, I'd highly recommend Vistage.

88
00:06:12,280 --> 00:06:17,160
But I was at their conference, their yearly conference in London and saw a great talk

89
00:06:17,160 --> 00:06:25,280
from Fireside Chat with Stephen Bartlett of Diary of a CEO fame among other things and

90
00:06:25,280 --> 00:06:26,280
Dragon's Den.

91
00:06:26,280 --> 00:06:27,280
That was really interesting.

92
00:06:27,280 --> 00:06:30,800
But the thing that really caught my eye from an AI perspective was I went to a breakout

93
00:06:30,800 --> 00:06:31,800
talk about AI.

94
00:06:31,800 --> 00:06:37,400
Probably the most important and interesting thing for me from the talk was a couple of

95
00:06:37,400 --> 00:06:42,240
polls that the speaker ran live during the session and there was probably, I don't know,

96
00:06:42,240 --> 00:06:48,840
80, 90 people in the room and the most overwhelming thing stopping people getting started with

97
00:06:48,840 --> 00:06:57,160
AI was a lack of access to training and also the lack of skills and people in the business

98
00:06:57,160 --> 00:07:02,480
who actually know how to leverage AI for the good of the business, which I think we've

99
00:07:02,480 --> 00:07:07,420
touched upon briefly on different episodes of the podcast, Martin, but it really drilled

100
00:07:07,420 --> 00:07:10,120
it home to me that people are starting to become aware of this, right?

101
00:07:10,120 --> 00:07:15,800
You've got your 10% aware of it and have played with it to 95 over the last year or so.

102
00:07:15,800 --> 00:07:17,840
But now they're trying to figure out how do we get started?

103
00:07:17,840 --> 00:07:23,600
And of course, it made me realize how remiss of us, Martin, how remiss of us that we offer

104
00:07:23,600 --> 00:07:29,840
AI training and AI consultancy for businesses in order for them to get started with AI.

105
00:07:29,840 --> 00:07:33,080
But we never mentioned it on the podcast because we're spending all of our time diving into

106
00:07:33,080 --> 00:07:36,480
the details and all the nitty gritty bits that we love to talk about.

107
00:07:36,480 --> 00:07:40,480
But yes, I think that's where businesses are, they're ready to get started, but they don't

108
00:07:40,480 --> 00:07:44,360
know where to start and they're looking for training and help.

109
00:07:44,360 --> 00:07:48,880
So if you visit artificiallyintelligentmarketing.com, we're going to add a few more details about

110
00:07:48,880 --> 00:07:52,120
the services we can offer around that for those of you that are interested.

111
00:07:52,120 --> 00:07:56,880
Meanwhile, I had a fantastic day out of Vistage and got to learn loads of great stuff.

112
00:07:56,880 --> 00:07:58,720
So it was loads of fun for me.

113
00:07:58,720 --> 00:08:05,640
Well, that's always a good bonus and good to see lots of people engaged in the topic

114
00:08:05,640 --> 00:08:07,120
of AI.

115
00:08:07,120 --> 00:08:09,800
Clearly people are wanting to know a lot more.

116
00:08:09,800 --> 00:08:14,680
So yeah, always happy to help on that front, aren't we?

117
00:08:14,680 --> 00:08:15,680
Absolutely.

118
00:08:15,680 --> 00:08:18,960
In the interest of helping, let's get into this week's stories.

119
00:08:18,960 --> 00:08:23,720
So the main stories this week, we're going to touch on Microsoft bringing AI shopping

120
00:08:23,720 --> 00:08:25,640
tools to Bing and Edge.

121
00:08:25,640 --> 00:08:31,000
We're going to look at how AI generated tweets might be more convincing than real people,

122
00:08:31,000 --> 00:08:33,020
according to some recent research.

123
00:08:33,020 --> 00:08:38,080
We're going to look at KPMG's report on the productivity boost from generative AI and how

124
00:08:38,080 --> 00:08:44,640
it can add 31 billion pounds to the UK economy and UK GDP.

125
00:08:44,640 --> 00:08:49,820
We're going to look at a little story around the fact that OpenAI has reported the planning

126
00:08:49,820 --> 00:08:54,760
to turn chat GPT into a super smart personal assistant for work.

127
00:08:54,760 --> 00:08:56,140
We're going to get into those.

128
00:08:56,140 --> 00:09:00,440
We're also going to have a look at tool of the week this week, Storybird, and Martin's

129
00:09:00,440 --> 00:09:01,960
going to tell us about that.

130
00:09:01,960 --> 00:09:03,680
And of course, we've got our short snippets.

131
00:09:03,680 --> 00:09:06,540
So let's get into those snippets first and foremost.

132
00:09:06,540 --> 00:09:11,120
So the first one is that OpenAI is opening its first expansion office and it's going

133
00:09:11,120 --> 00:09:13,660
to be here in the UK in London.

134
00:09:13,660 --> 00:09:15,440
So welcome to the OpenAI team.

135
00:09:15,440 --> 00:09:20,680
Hit us up if you want to be on the podcast or you fancy a coffee.

136
00:09:20,680 --> 00:09:28,600
We also have Demis Hsabis, Google DeepMind CEO, saying that its next algorithm will eclipse

137
00:09:28,600 --> 00:09:29,680
chat GPT.

138
00:09:29,680 --> 00:09:36,760
So there was an interview with Wired Magazine where he mentioned Google working on a sort

139
00:09:36,760 --> 00:09:40,880
of combined technology where they're going to improve the LLM with some of the techniques

140
00:09:40,880 --> 00:09:48,200
used in AlphaGo to give their new system Gemini abilities to do things such as planning and

141
00:09:48,200 --> 00:09:50,040
solving complex problems.

142
00:09:50,040 --> 00:09:54,100
So that's quite interesting and they were brave enough to come out and say it's going

143
00:09:54,100 --> 00:09:56,520
to be much better than GPT-4.

144
00:09:56,520 --> 00:09:58,880
So watch that with bated breath.

145
00:09:58,880 --> 00:10:00,800
I am very excited by that.

146
00:10:00,800 --> 00:10:05,400
I just have a lot of faith in Hsabis and the DeepMind team.

147
00:10:05,400 --> 00:10:11,920
I just love what they've done and their approach that they've taken with AlphaGo and AlphaZero.

148
00:10:11,920 --> 00:10:17,720
I'm excited to see what comes of that combination with LLMs.

149
00:10:17,720 --> 00:10:26,920
I think everyone in the industry recognizes the track record and the deep amount of expertise

150
00:10:26,920 --> 00:10:31,880
in this area within Google and if they're going to come out and say that, I think we

151
00:10:31,880 --> 00:10:34,720
have to take them very seriously.

152
00:10:34,720 --> 00:10:39,820
I've not been able to find any mention on timeline for that in terms of us actually

153
00:10:39,820 --> 00:10:44,800
seeing it in action, but yes, I'd be really excited to see when that comes out.

154
00:10:44,800 --> 00:10:49,720
Another short snippet, you might remember that we spoke on the podcast about DRAGAN, which

155
00:10:49,720 --> 00:10:55,800
was a research paper showing how you could track bits of images and the algorithm would

156
00:10:55,800 --> 00:10:59,200
be able to adjust them to make them look real.

157
00:10:59,200 --> 00:11:06,080
So I think the examples we gave were turning a lion's head, extending a model's shorts

158
00:11:06,080 --> 00:11:12,800
into trousers and adjusting a car features and then the rest of the car adjusts around

159
00:11:12,800 --> 00:11:13,800
it.

160
00:11:13,800 --> 00:11:16,760
Well, you can access this on Huggingface.

161
00:11:16,760 --> 00:11:20,920
You can both access the source code and there is a demo and I had a little play with the

162
00:11:20,920 --> 00:11:21,920
demo, Martin.

163
00:11:21,920 --> 00:11:27,160
I have to say by and large, it lived up to the videos in the research paper, which was

164
00:11:27,160 --> 00:11:28,160
quite amazing.

165
00:11:28,160 --> 00:11:34,440
There's a model and I took the poor model and I moved the model's right arm right out

166
00:11:34,440 --> 00:11:41,200
and I had the model's head turn around to the right quite far, almost to the point where

167
00:11:41,200 --> 00:11:45,720
it probably would have broken a human's neck and it did a reasonably good job of making

168
00:11:45,720 --> 00:11:50,200
it look like a real representation of what that would look like if the image had been

169
00:11:50,200 --> 00:11:51,200
captured like that.

170
00:11:51,200 --> 00:11:52,640
So that was pretty cool.

171
00:11:52,640 --> 00:11:57,280
I recommend going and having a play with that and it'll be interesting to see how quickly

172
00:11:57,280 --> 00:12:01,140
this type of technology can be baked into commercial products.

173
00:12:01,140 --> 00:12:07,000
One has to assume Adobe Photoshop's AI beta is going to get an update at some point if

174
00:12:07,000 --> 00:12:10,640
they can develop similar technologies, right?

175
00:12:10,640 --> 00:12:14,840
There was another cool story, this is from one close to my heart from in the life sciences,

176
00:12:14,840 --> 00:12:16,880
Martin, from Insilico Medicine.

177
00:12:16,880 --> 00:12:24,080
So Insilico is a member of the Nvidia Inception sort of group and we talked a little bit about

178
00:12:24,080 --> 00:12:29,800
how Nvidia had invested in and was working with companies to leverage large language

179
00:12:29,800 --> 00:12:34,520
models and other AI-based tools to help drive business growth.

180
00:12:34,520 --> 00:12:40,320
In this case, pretty cool because Insilico is entering phase two clinical trials for

181
00:12:40,320 --> 00:12:46,320
a drug candidate that they discovered using their AI platform, which is to treat idiopathic

182
00:12:46,320 --> 00:12:49,120
pulmonary fibrosis.

183
00:12:49,120 --> 00:12:54,520
But why is this really interesting is according to Nvidia, doing this the traditional way

184
00:12:54,520 --> 00:12:58,520
would have cost more than $400 million and taken up to six years, but using generative

185
00:12:58,520 --> 00:13:04,600
AI, Insilico was able to do it in one-tenth of the cost and one-third of the time, reaching

186
00:13:04,600 --> 00:13:10,280
that first phase of clinical trials just two and a half years after beginning the project.

187
00:13:10,280 --> 00:13:15,400
Here in the life sciences, it's very well recognized that going from an initial idea

188
00:13:15,400 --> 00:13:19,280
to getting a drug to market takes way too long and costs way too much.

189
00:13:19,280 --> 00:13:25,800
So this is again, more evidence the power of AI in the life sciences especially.

190
00:13:25,800 --> 00:13:30,160
Good to see a real world example of that because it's one of those ones where we hear about

191
00:13:30,160 --> 00:13:32,400
the promise of this.

192
00:13:32,400 --> 00:13:37,240
So actually seeing a kind of tangible project saying, yeah, we're doing this, this is moving

193
00:13:37,240 --> 00:13:40,480
and this is the result and the impact.

194
00:13:40,480 --> 00:13:41,480
That's great to see.

195
00:13:41,480 --> 00:13:44,100
Yeah, very exciting stuff.

196
00:13:44,100 --> 00:13:50,160
Looking forward to seeing more such examples I think over the next two or three years here.

197
00:13:50,160 --> 00:13:54,640
We talked a little bit about Mid Journey's update last week but since then I've been

198
00:13:54,640 --> 00:13:57,800
having a play and seen quite a few examples online, I don't know if you've seen them

199
00:13:57,800 --> 00:14:02,320
or not but people have been getting some amazingly interesting effects with this new zoom tool,

200
00:14:02,320 --> 00:14:08,360
basically creating like infinite zoom outs and then combining them in that Adobe After

201
00:14:08,360 --> 00:14:12,200
Effects to create these videos where they start with an image and they zoom out and

202
00:14:12,200 --> 00:14:18,080
out and out and out and half the time the tool does just a really good job of applying

203
00:14:18,080 --> 00:14:23,200
proper context and some of the times it goes absolutely loopy and like puts an image inside

204
00:14:23,200 --> 00:14:27,200
a photo frame inside a painting and it's kind of mad.

205
00:14:27,200 --> 00:14:29,040
Have you seen any of these?

206
00:14:29,040 --> 00:14:35,560
Yeah, I watched one this morning actually and it was a plane that had crashed in a landscape

207
00:14:35,560 --> 00:14:40,760
and it zoomed out and in the journey that it zooms out, it must zoom out like it must

208
00:14:40,760 --> 00:14:47,560
be like the equivalent to two miles away and you end up in this cabin almost with a pilot

209
00:14:47,560 --> 00:14:51,800
or someone looking out over the landscape where there's basically just been multiple

210
00:14:51,800 --> 00:14:53,800
plane crashes.

211
00:14:53,800 --> 00:14:56,520
Yeah, they're really impressive.

212
00:14:56,520 --> 00:15:01,800
The Mid Journey updates that came out recently with 5.2, I've been playing with some of the

213
00:15:01,800 --> 00:15:05,280
other ones, have you tried the shorten capability?

214
00:15:05,280 --> 00:15:06,840
No, not yet.

215
00:15:06,840 --> 00:15:07,840
Tell us about that.

216
00:15:07,840 --> 00:15:15,640
So, shorten is where you take your existing prompt and you ask it to shorten it and what

217
00:15:15,640 --> 00:15:20,680
it does is it basically cuts out bits of text that it doesn't think are relevant and it's

218
00:15:20,680 --> 00:15:24,840
not going to help get you any better output and it gives you four options and it does

219
00:15:24,840 --> 00:15:26,480
aggressively shorten it.

220
00:15:26,480 --> 00:15:31,360
The first one is slightly shortened and the last one is basically like hardcore, we've

221
00:15:31,360 --> 00:15:35,960
reduced this to the core words that it could be.

222
00:15:35,960 --> 00:15:41,440
But it is insightful when you see what words do and do not make a difference and I think

223
00:15:41,440 --> 00:15:49,580
if you want to become a better prompter in Mid Journey, using this shortened capability

224
00:15:49,580 --> 00:15:56,200
will help you understand which adjectives to include or not include or how to maybe

225
00:15:56,200 --> 00:16:03,680
structure your sentence so that the primary topic and primary focuses are in the prompt.

226
00:16:03,680 --> 00:16:06,240
Yeah, I think that's a really good point.

227
00:16:06,240 --> 00:16:09,600
I'm definitely going to be playing with that because it is hard to know and what you end

228
00:16:09,600 --> 00:16:13,720
up doing is you get into a pattern where you include certain terms of phrase because you

229
00:16:13,720 --> 00:16:14,720
think you need them.

230
00:16:14,720 --> 00:16:20,040
Like I don't know, I see a lot of prompts and I now do this myself that write like photorealistic,

231
00:16:20,040 --> 00:16:24,760
8K, super resolution and all these other things because you think you need it to get the best

232
00:16:24,760 --> 00:16:25,760
output.

233
00:16:25,760 --> 00:16:29,880
It'd be interesting to see if they basically all get stripped out or you only need like

234
00:16:29,880 --> 00:16:36,120
one turn of phrase like 8K to suggest that you want it to appear really high resolution.

235
00:16:36,120 --> 00:16:39,720
The other thing came up as I was having a discussion on LinkedIn, I can't remember if

236
00:16:39,720 --> 00:16:44,280
we've mentioned this on the podcast, there's a lot more tools and prompts in Mid Journey

237
00:16:44,280 --> 00:16:45,880
than I think a lot of people realize.

238
00:16:45,880 --> 00:16:50,640
A lot of people are using Imagine, which is the one that you use to generate an image,

239
00:16:50,640 --> 00:16:55,560
but my favorite is to use Describe and then feed it an image that I think is really cool

240
00:16:55,560 --> 00:17:00,440
and then it reverse engineers what type of prompt would give you an image like that and

241
00:17:00,440 --> 00:17:05,400
it gives you, I think, four prompt ideas, which is absolutely fantastic for then taking

242
00:17:05,400 --> 00:17:10,960
that original image idea and then twisting it as you see fit, so to give you something

243
00:17:10,960 --> 00:17:11,960
completely novel.

244
00:17:11,960 --> 00:17:17,800
So I don't know if you've seen famous actors as fruit or something and you've seen like,

245
00:17:17,800 --> 00:17:23,880
I don't know, Leonardo DiCaprio morphed into some sort of pineapple, I don't know, I haven't

246
00:17:23,880 --> 00:17:27,280
seen, I don't even know what would happen if you enter that prompt, please enter that

247
00:17:27,280 --> 00:17:29,920
prompt and then tweet us and so we can see it.

248
00:17:29,920 --> 00:17:34,960
But then if you wanted to do exactly the same thing and turn a bunch of other famous people

249
00:17:34,960 --> 00:17:38,720
into fruit, just feed it the image and it will give you like the bare bones of what

250
00:17:38,720 --> 00:17:43,160
you needed and then you could just swap the fruit out and you could swap the famous person

251
00:17:43,160 --> 00:17:45,520
out and then you can get a whole series of images.

252
00:17:45,520 --> 00:17:49,960
Are you doing this for us in real time, Martin, so we can see what that looks like?

253
00:17:49,960 --> 00:17:50,960
Oh crumbs.

254
00:17:50,960 --> 00:17:56,600
Okay, well now at least we've got an image for this episode of the podcast.

255
00:17:56,600 --> 00:17:59,320
So yeah, mid-journey, cool, go play with it.

256
00:17:59,320 --> 00:18:00,800
There's loads of extra power there now.

257
00:18:00,800 --> 00:18:02,520
What else have we seen this week?

258
00:18:02,520 --> 00:18:03,520
We've seen GPT Author.

259
00:18:03,520 --> 00:18:10,040
So I don't know if people have seen this, but this has been released on GitHub and according

260
00:18:10,040 --> 00:18:15,480
to GitHub, this is a project that utilizes a chain of GPT-4 and stable diffusion API

261
00:18:15,480 --> 00:18:19,480
calls to generate an original fantasy novel.

262
00:18:19,480 --> 00:18:23,480
Readers can provide an initial prompt and enter how many chapters they'd like it to

263
00:18:23,480 --> 00:18:30,440
be and the AI then generates an entire novel outputting an EPUB file compatible with ebook

264
00:18:30,440 --> 00:18:37,360
readers, and that you can create a 15-chapter novel for as little as $4 and it will be written

265
00:18:37,360 --> 00:18:40,040
in just a few minutes.

266
00:18:40,040 --> 00:18:42,080
That's kind of a bit mind-blowing.

267
00:18:42,080 --> 00:18:46,680
It's not out of the realms of what you might expect given what we know the tools can do,

268
00:18:46,680 --> 00:18:52,160
but the fact that someone's now made that tool available as an open source on GitHub

269
00:18:52,160 --> 00:18:59,200
for people to go play with, I can imagine an explosion now of book generators, one assumes,

270
00:18:59,200 --> 00:19:00,200
Martin.

271
00:19:00,200 --> 00:19:01,200
What are your thoughts on that one?

272
00:19:01,200 --> 00:19:07,080
Well, how long till the Booker Prize is won by an AI and the author has to complete much

273
00:19:07,080 --> 00:19:09,640
like the Sony photography award recently?

274
00:19:09,640 --> 00:19:12,680
That will be interesting to see.

275
00:19:12,680 --> 00:19:20,160
That is the danger and one assumes you can currently ask ChatGPT4 to write in the style

276
00:19:20,160 --> 00:19:26,280
of famous poets and authors and social media gurus.

277
00:19:26,280 --> 00:19:32,000
One assumes you could start cranking out novels based on existing authors' works that you

278
00:19:32,000 --> 00:19:34,680
really liked and then that gets muddy.

279
00:19:34,680 --> 00:19:38,400
It's like, oh, what would have happened if Michael Crichton had written Jurassic Park

280
00:19:38,400 --> 00:19:43,280
3 instead of all of the series of movies that we eventually got?

281
00:19:43,280 --> 00:19:47,800
Well, now we can pump it into the tool and see what we get.

282
00:19:47,800 --> 00:19:53,800
I think it'd be really interesting to see how that all shakes out.

283
00:19:53,800 --> 00:20:00,040
Then the last little snippet is actually an update on a story that we've already featured,

284
00:20:00,040 --> 00:20:05,880
but I think maybe it's interesting to revisit, which is this concept that Amazon is building

285
00:20:05,880 --> 00:20:15,680
a large language model marketplace as part of its sort of bedrock solution for its AWS

286
00:20:15,680 --> 00:20:16,680
customers.

287
00:20:16,680 --> 00:20:19,600
We've talked about this before, but I think the reason this is especially interesting

288
00:20:19,600 --> 00:20:24,120
to hear this come up is I've also heard recently that OpenAI are thinking about doing something

289
00:20:24,120 --> 00:20:30,920
similar and also I think as marketers get to grip with the base level capabilities of

290
00:20:30,920 --> 00:20:37,520
an OpenAI chat-chibi-tea type tool for simple content generation, I'm meeting lots of business

291
00:20:37,520 --> 00:20:43,320
owners and marketers who wish that they had access to an LLM that understands their own

292
00:20:43,320 --> 00:20:48,960
business and their own content and their own internal information better to be able to

293
00:20:48,960 --> 00:20:55,720
produce things that are more customized to them, either as an internal knowledge product

294
00:20:55,720 --> 00:20:59,320
so that you almost have an Ask Jeeves, but for internal stuff.

295
00:20:59,320 --> 00:21:03,640
It's like, oh, who do I go to to find this out?

296
00:21:03,640 --> 00:21:07,800
What document contains information on this technical aspect of this product or whatever

297
00:21:07,800 --> 00:21:09,280
it may be?

298
00:21:09,280 --> 00:21:14,400
Also things like obviously driving more company-specific content production for content marketing,

299
00:21:14,400 --> 00:21:16,600
which I think is a common use case already.

300
00:21:16,600 --> 00:21:21,440
The concept of creating customer service chat bots that are really well-trained on all your

301
00:21:21,440 --> 00:21:23,040
own service materials.

302
00:21:23,040 --> 00:21:29,360
I'm getting asked more and more, how would I go about doing this?

303
00:21:29,360 --> 00:21:34,840
One assumes that therefore this marketplace from Amazon could make that easier.

304
00:21:34,840 --> 00:21:39,040
If you're a product manager or a marketer, you probably need to keep an eye on this space

305
00:21:39,040 --> 00:21:44,480
as it becomes easier and easier to develop large language models on your own information

306
00:21:44,480 --> 00:21:50,800
for marketing, sales, and customer support use cases without all of the technical complexity

307
00:21:50,800 --> 00:21:53,480
that would go into it.

308
00:21:53,480 --> 00:21:56,520
So I think that's going to be interesting to see how that evolves.

309
00:21:56,520 --> 00:21:58,480
Yeah, I agree.

310
00:21:58,480 --> 00:22:03,280
I think the one thing with all of that, because I'm getting a lot of people asking me as well

311
00:22:03,280 --> 00:22:06,120
about this, how do you create these kind of knowledge bases?

312
00:22:06,120 --> 00:22:12,000
And to a certain extent, I think the co-pilots of this world, when all of your documentation

313
00:22:12,000 --> 00:22:17,240
can be read by co-pilot, certainly for an internal knowledge base for employees.

314
00:22:17,240 --> 00:22:22,800
I think I'm expecting Microsoft or Google Workspace AI to do a lot of the heavy lifting

315
00:22:22,800 --> 00:22:23,800
there.

316
00:22:23,800 --> 00:22:28,080
So obviously, if you need a customer facing one, I think there's going to be lots of solutions

317
00:22:28,080 --> 00:22:29,080
coming forward.

318
00:22:29,080 --> 00:22:34,240
But yeah, interesting proposition from Amazon there.

319
00:22:34,240 --> 00:22:40,560
Yeah, and maybe I'm being overly optimistic, but I'm kind of disappointed that we haven't...

320
00:22:40,560 --> 00:22:45,200
I know that there are beaters and that some people have some access to Windows co-pilot

321
00:22:45,200 --> 00:22:52,400
and others have access to some of the Google's palm models in Google Docs and Gmail.

322
00:22:52,400 --> 00:23:00,480
But I was expecting, straight hoping for a slightly wider rollout of those tools by now.

323
00:23:00,480 --> 00:23:02,440
And we haven't seen it yet.

324
00:23:02,440 --> 00:23:10,660
So they're built because they've beaten them, but they're not rolling them out that quickly.

325
00:23:10,660 --> 00:23:12,360
So what can we read into that?

326
00:23:12,360 --> 00:23:15,840
Is it, do we think it's technical infrastructure that's going to slow that down?

327
00:23:15,840 --> 00:23:20,920
Because obviously, there's a lot of compute power that's going to be needed to fuel that.

328
00:23:20,920 --> 00:23:26,200
Is it the tools don't perform that well in the hands of real people doing real use cases?

329
00:23:26,200 --> 00:23:31,040
And so actually figuring that out and ironing out the kinks of slowing things down?

330
00:23:31,040 --> 00:23:32,040
I don't know.

331
00:23:32,040 --> 00:23:33,200
What do you think?

332
00:23:33,200 --> 00:23:34,920
Is this timeline...

333
00:23:34,920 --> 00:23:36,320
Am I being overly optimistic?

334
00:23:36,320 --> 00:23:38,920
Is everything playing out as well as you might expect?

335
00:23:38,920 --> 00:23:41,480
Or do you think we should maybe have this by now?

336
00:23:41,480 --> 00:23:44,600
I thought we'd have it in autumn, to be honest.

337
00:23:44,600 --> 00:23:48,880
When they announced it, I think you've touched on a few things.

338
00:23:48,880 --> 00:23:53,720
The compute power, turning this on overnight and giving this to the world, everyone with

339
00:23:53,720 --> 00:23:57,480
Windows 11 or what have you.

340
00:23:57,480 --> 00:24:00,320
A few things might fall over.

341
00:24:00,320 --> 00:24:05,120
But also, I mean, early testing, I've seen a few headlines and a few tweets.

342
00:24:05,120 --> 00:24:10,040
People aren't blown away with it so far.

343
00:24:10,040 --> 00:24:12,160
So maybe it's just a case of, do you know what?

344
00:24:12,160 --> 00:24:19,480
This is just going to take a lot more testing, a lot more real world, get it in the hands

345
00:24:19,480 --> 00:24:21,640
of users that are going to break things.

346
00:24:21,640 --> 00:24:23,680
That would make sense.

347
00:24:23,680 --> 00:24:29,040
We know these thick tools are prone to hallucination and you have to very precisely brief them

348
00:24:29,040 --> 00:24:31,960
currently to get exactly what you want.

349
00:24:31,960 --> 00:24:35,320
Yeah, I think that will make sense.

350
00:24:35,320 --> 00:24:36,320
Right.

351
00:24:36,320 --> 00:24:40,880
Let's move on to our main stories for the week and let's start with Microsoft bringing

352
00:24:40,880 --> 00:24:43,920
AI shopping tools to Bing and Edge.

353
00:24:43,920 --> 00:24:44,920
So you spotted this.

354
00:24:44,920 --> 00:24:46,920
What were your thoughts on this?

355
00:24:46,920 --> 00:24:53,080
Yeah, so Microsoft have announced a new AI powered shopping tool for Bing and Bing AI

356
00:24:53,080 --> 00:24:59,560
chatbot, which is available in the Edge sidebar browser.

357
00:24:59,560 --> 00:25:04,600
So this new tool set includes an automatic buying guide generator, which uses Bing's

358
00:25:04,600 --> 00:25:11,360
GPT powered AI to aggregate and list product specs for comparison and purchasing.

359
00:25:11,360 --> 00:25:18,000
So as part of this, the buying guide will pull in content from a bunch of different

360
00:25:18,000 --> 00:25:19,000
websites.

361
00:25:19,000 --> 00:25:24,880
At the moment, I believe it's only available in the US, but worldwide rollout is coming

362
00:25:24,880 --> 00:25:25,880
soon.

363
00:25:25,880 --> 00:25:29,320
Couple of interesting points from the buying guide that it generates.

364
00:25:29,320 --> 00:25:34,760
It has AI generated review summaries, so it's taking lots of reviews and then summarizing

365
00:25:34,760 --> 00:25:38,600
them all to give you the good, the bad, the highlights and what have you.

366
00:25:38,600 --> 00:25:42,320
There's another feature that I thought was particularly interesting and you can see here

367
00:25:42,320 --> 00:25:49,000
how AI is being monetized and that's the price match tool.

368
00:25:49,000 --> 00:25:52,240
So what they're doing is that they're putting in a price match tool that helps users to

369
00:25:52,240 --> 00:25:56,680
request a price match from retailers.

370
00:25:56,680 --> 00:26:02,560
Now it's expected or it's thought that Microsoft are getting an affiliate fee, they're getting

371
00:26:02,560 --> 00:26:06,160
some commission on the sales through these buyer's guides.

372
00:26:06,160 --> 00:26:10,360
So that's how they're going to actually monetize this as a strategy.

373
00:26:10,360 --> 00:26:15,400
But yeah, I thought it was a really small use case and definitely one for e-commerce

374
00:26:15,400 --> 00:26:21,360
companies to be aware of in terms of how do you get your products into these buyer's guides.

375
00:26:21,360 --> 00:26:31,320
On the one hand, I think this is good for consumers and to an extent good for product

376
00:26:31,320 --> 00:26:34,320
marketers and resellers.

377
00:26:34,320 --> 00:26:39,400
But if you're a publisher, how many, if I think about last time I bought a bit of tech,

378
00:26:39,400 --> 00:26:45,000
in fact the last bit of tech that I bought was my PC, I think I've mentioned it recently,

379
00:26:45,000 --> 00:26:47,180
I bought my new PC.

380
00:26:47,180 --> 00:26:53,360
How many blog posts, how many articles did I read on the likes of TechRadar, on the likes

381
00:26:53,360 --> 00:26:57,240
of Tom's Guide, on all of these...

382
00:26:57,240 --> 00:27:04,040
The Net and all these other platforms with the 10 best laptops of 2023 style thing, yeah.

383
00:27:04,040 --> 00:27:05,920
Yeah, all of that.

384
00:27:05,920 --> 00:27:14,280
They are going to be sitting up and going, ooh, Bing is coming for our, he's going to

385
00:27:14,280 --> 00:27:15,280
eat our lunch.

386
00:27:15,280 --> 00:27:17,560
Yeah, it's interesting.

387
00:27:17,560 --> 00:27:24,040
I mean, obviously there's the impact on e-commerce stores and how can you get into these sort

388
00:27:24,040 --> 00:27:27,200
of automated buying guides, if you like.

389
00:27:27,200 --> 00:27:35,240
But I saw a post on LinkedIn this week, slightly tangential but related, about the best tools

390
00:27:35,240 --> 00:27:42,400
for SEO were writing listicles and how-to guides.

391
00:27:42,400 --> 00:27:48,120
That may still be accurate today, I don't think it'll be accurate for very long.

392
00:27:48,120 --> 00:27:53,160
And personally, I probably wouldn't invest a huge amount of time and energy in those

393
00:27:53,160 --> 00:27:56,960
because I think they're going to be asserted by generative AI.

394
00:27:56,960 --> 00:28:02,960
That basically gives people that type of information naturally in the response that you get from

395
00:28:02,960 --> 00:28:03,960
the tool.

396
00:28:03,960 --> 00:28:11,440
Yeah, you already see that coming in the examples of Google's search generative experience,

397
00:28:11,440 --> 00:28:13,800
their listicles being written.

398
00:28:13,800 --> 00:28:18,040
You even see listicles that are listing websites without actually linking to the website, which

399
00:28:18,040 --> 00:28:19,600
is crazy to me.

400
00:28:19,600 --> 00:28:25,920
I have to think that they'll do slightly more for publishers rather than just eating all

401
00:28:25,920 --> 00:28:28,280
of their content and regurgitating it.

402
00:28:28,280 --> 00:28:35,080
But yeah, I think if you're in that game, you feel like it's not diversifying how you

403
00:28:35,080 --> 00:28:36,080
create content.

404
00:28:36,080 --> 00:28:43,840
Yeah, I think where my head starts to melt is these tools have been trained on and one

405
00:28:43,840 --> 00:28:48,200
assumes will continue to be trained on whatever content they can get their hands on, which

406
00:28:48,200 --> 00:28:50,400
means the web.

407
00:28:50,400 --> 00:28:57,920
But if we all now stop developing content in emerging areas, let's all decide, oh,

408
00:28:57,920 --> 00:29:03,520
listicles, how-to's, the common go-to type of content is not worth doing now because

409
00:29:03,520 --> 00:29:07,800
you won't get ranked or people won't click on your stuff because it will just be in a

410
00:29:07,800 --> 00:29:09,880
generative search result.

411
00:29:09,880 --> 00:29:18,160
Well, a new area in science emerges or a new technique, which happens a lot.

412
00:29:18,160 --> 00:29:20,960
The tool can't write a how-to.

413
00:29:20,960 --> 00:29:27,720
It's got nothing to absorb and to crawl and to learn from.

414
00:29:27,720 --> 00:29:33,680
So does that mean that how-to's and listicles can continue but you have to stay right at

415
00:29:33,680 --> 00:29:41,440
the cutting edge of material producing information that's going to be things that you know the

416
00:29:41,440 --> 00:29:44,840
tools haven't been trained on yet because it's so new?

417
00:29:44,840 --> 00:29:50,400
So do we end up with a window of opportunity on the how-to's and the listicles and what

418
00:29:50,400 --> 00:29:56,800
have you until just enough content has been generated that now the chat-tipities of the

419
00:29:56,800 --> 00:30:00,880
world have now been trained on it and they can regurgitate it too?

420
00:30:00,880 --> 00:30:01,880
What do you think about that?

421
00:30:01,880 --> 00:30:05,920
Yeah, well, I think that's the case for a lot of companies as it is.

422
00:30:05,920 --> 00:30:10,240
Anyway, a good example from my own blog is that I was ranking really well with a blog

423
00:30:10,240 --> 00:30:16,120
that was the best text to image, well, the best AI text to image generators and I was

424
00:30:16,120 --> 00:30:21,600
updating this monthly up until about September, October last year and I was getting a good

425
00:30:21,600 --> 00:30:25,920
amount of traffic from that because it was first mover advantage.

426
00:30:25,920 --> 00:30:29,480
There wasn't many people talking about it and you could get a lot of traffic through

427
00:30:29,480 --> 00:30:37,640
that but now everyone's written that blog post has been copied, there's probably 20

428
00:30:37,640 --> 00:30:42,320
different versions of it on the likes of TechRadar or what have you.

429
00:30:42,320 --> 00:30:52,480
So trying to outrank those is just too difficult now but I think what's changing is who the

430
00:30:52,480 --> 00:30:57,040
big competitor is, no longer is it the big publishing houses, you know, not HubSpot writing

431
00:30:57,040 --> 00:31:00,400
the blog and me going, there's no way I'm going to outrank HubSpot.

432
00:31:00,400 --> 00:31:06,600
Now it's, well, I'm not trying to outrank, I just don't want Google to rewrite my article

433
00:31:06,600 --> 00:31:11,280
and present it as its own work which it inevitably will do.

434
00:31:11,280 --> 00:31:15,520
Yeah, of a fall, of a fall.

435
00:31:15,520 --> 00:31:21,520
When I was out and about this week, I had someone describe chat-tipity and other tools

436
00:31:21,520 --> 00:31:28,320
as regurgitating the information in their training data which is not strictly true,

437
00:31:28,320 --> 00:31:29,320
is it?

438
00:31:29,320 --> 00:31:31,160
It's probability-based engineering.

439
00:31:31,160 --> 00:31:35,960
But certainly using you as inspiration and a ton of other people, Martin, to produce

440
00:31:35,960 --> 00:31:40,480
some content that you used to get a lot of commercial value out of and then you'll quickly

441
00:31:40,480 --> 00:31:41,480
get zero.

442
00:31:41,480 --> 00:31:51,160
I don't know if you follow Lily Ray on Twitter, she's an SEO influencer.

443
00:31:51,160 --> 00:31:57,560
She's been following the Google SGE process and doing lots of tweets about it and she's

444
00:31:57,560 --> 00:32:03,720
actually found examples where search generative experience is just copying blocks of text

445
00:32:03,720 --> 00:32:04,720
verbatim.

446
00:32:04,720 --> 00:32:10,440
It isn't rewriting it there and it is just, it is presenting it as its own and it's very

447
00:32:10,440 --> 00:32:13,760
clearly a paragraph of text written elsewhere.

448
00:32:13,760 --> 00:32:14,760
Wow.

449
00:32:14,760 --> 00:32:19,480
So we've actually got this blurred line between the models understanding how to communicate

450
00:32:19,480 --> 00:32:24,840
as humans from their large training data sets and then of course, because Google has access

451
00:32:24,840 --> 00:32:31,280
to all the world's information indexed as part of its search algorithm, actually then

452
00:32:31,280 --> 00:32:38,160
going and grabbing chunks of text and then just pasting them in amongst its natural LLM

453
00:32:38,160 --> 00:32:39,160
put.

454
00:32:39,160 --> 00:32:40,160
Yep.

455
00:32:40,160 --> 00:32:41,580
And not giving credit.

456
00:32:41,580 --> 00:32:46,760
So it's like a featured snippet, except you don't even get the link and the referral.

457
00:32:46,760 --> 00:32:50,800
They've just taken your text without giving you the signposting.

458
00:32:50,800 --> 00:32:55,120
Oh, the world is so very gray at the moment.

459
00:32:55,120 --> 00:32:56,120
Right.

460
00:32:56,120 --> 00:33:00,760
With that in mind, let's move on to our next story, which is kind of similar to what we've

461
00:33:00,760 --> 00:33:01,920
been talking about really.

462
00:33:01,920 --> 00:33:08,280
So story number two here is that AI generated tweets might be more convincing than real

463
00:33:08,280 --> 00:33:12,320
people according to some recent research.

464
00:33:12,320 --> 00:33:20,000
So we saw this story on The Verge and in this report, a recent study found that people trust

465
00:33:20,000 --> 00:33:27,640
AI generated tweets, in this case, those written by GPT-3, more than those written by humans.

466
00:33:27,640 --> 00:33:32,120
So the research involved a survey where people were asked to identify the source of a tweet

467
00:33:32,120 --> 00:33:36,320
and determine its validity and some of the tweets were from humans and some of the tweets

468
00:33:36,320 --> 00:33:38,160
were created by GPT-3.

469
00:33:38,160 --> 00:33:42,440
The tweets focused on various scientific topics, including vaccines and climate change.

470
00:33:42,440 --> 00:33:46,040
So also, you know, potentially quite controversial topics.

471
00:33:46,040 --> 00:33:53,800
And it involved 697 participants in the UK, Australia, Canada, the US and Ireland, and

472
00:33:53,800 --> 00:34:00,280
they covered 11 different topics, including topics where the content of the tweets was

473
00:34:00,280 --> 00:34:03,440
correct, but some that was deliberately inaccurate.

474
00:34:03,440 --> 00:34:09,200
And the results show that people struggled to distinguish between human written and GPT-3

475
00:34:09,200 --> 00:34:14,040
written tweets and were more likely to accept both accurate and inaccurate information if

476
00:34:14,040 --> 00:34:16,320
it was written by the AI.

477
00:34:16,320 --> 00:34:22,440
So Giovanni Spitali, the study's lead author, was very keen to emphasize that the technology

478
00:34:22,440 --> 00:34:27,960
itself isn't good or evil, but it's just amplifying human intentions.

479
00:34:27,960 --> 00:34:35,080
And that the study itself, which is in science advances, concluded that GPT-3 generated content

480
00:34:35,080 --> 00:34:39,480
was indistinguishable from organic content.

481
00:34:39,480 --> 00:34:42,720
Research suggests that improving training data sets could prevent the misuse of these

482
00:34:42,720 --> 00:34:49,120
tools for disinformation, noting that GPT-3 resisted generating false content about vaccines

483
00:34:49,120 --> 00:34:54,360
and autism, potentially due to the amount of debunking information in the training data

484
00:34:54,360 --> 00:35:00,240
set, and that ultimately Spitali advocates for fostering critical thinking skills as

485
00:35:00,240 --> 00:35:04,760
the most effective long-term strategy for countering disinformation.

486
00:35:04,760 --> 00:35:10,240
So a fascinating story, because we're not even talking GPT-4 here, we're talking its

487
00:35:10,240 --> 00:35:11,560
predecessor.

488
00:35:11,560 --> 00:35:15,960
Tweets are not exactly the longest form content in the world, so probably reasonably easy

489
00:35:15,960 --> 00:35:20,920
for it to do a half decent job in the character limits that you find on Twitter.

490
00:35:20,920 --> 00:35:26,240
But the idea that people just really couldn't tell the difference is further evidence that

491
00:35:26,240 --> 00:35:33,000
if you wanted to create lots of Twitter content at scale using bots for things like disinformation

492
00:35:33,000 --> 00:35:39,080
campaigns, your average person's probably not going to be able to tell truth from fact.

493
00:35:39,080 --> 00:35:42,120
What are your thoughts about this story, Mike?

494
00:35:42,120 --> 00:35:46,080
I like the conclusion certainly about fostering critical thinking skills.

495
00:35:46,080 --> 00:35:53,720
I think this is good advice in the round, where it doesn't necessarily need to be in

496
00:35:53,720 --> 00:35:57,360
the context of disinformation and AI.

497
00:35:57,360 --> 00:36:04,080
It actually brought to mind the story from a few weeks ago where the research paper was

498
00:36:04,080 --> 00:36:13,360
about the email responses from doctors or GPT-4, or I can't remember which one it was.

499
00:36:13,360 --> 00:36:16,680
I think it was 3.5 for a method.

500
00:36:16,680 --> 00:36:20,520
The AI was more empathetic and that kind of thing.

501
00:36:20,520 --> 00:36:21,920
It just reminded me of that.

502
00:36:21,920 --> 00:36:28,760
I think we shouldn't be surprised when people are taken in by AI written content anymore.

503
00:36:28,760 --> 00:36:30,120
Yeah.

504
00:36:30,120 --> 00:36:36,320
I think basically the verdict's in on that, which is in a lot of use cases, we can't really

505
00:36:36,320 --> 00:36:37,320
tell the difference.

506
00:36:37,320 --> 00:36:45,760
It's funny, I've had a number of conversations this week about what does AI do for education

507
00:36:45,760 --> 00:36:47,760
and how things are changing.

508
00:36:47,760 --> 00:36:53,540
I think a lot of the frustrations with the education system, both from people in it outside

509
00:36:53,540 --> 00:36:58,140
of it, I'm not an expert in this, and mostly just repurposing conversations I've had and

510
00:36:58,140 --> 00:37:07,560
things that I've heard is having schooling moving as far away from learning to remember

511
00:37:07,560 --> 00:37:14,840
versus learning to think and critically analyze and question and be curious and seek sources,

512
00:37:14,840 --> 00:37:18,840
which I realize is a big part of how schooling already works.

513
00:37:18,840 --> 00:37:24,640
Really going in all in on that is probably going to be critical because when you've got

514
00:37:24,640 --> 00:37:29,080
tools that can surface, I mean we've had this with the web really for years, but it's even

515
00:37:29,080 --> 00:37:31,800
easier now to go to a chat GPT and ask questions.

516
00:37:31,800 --> 00:37:37,420
But needing to be able to think about how to assess what you've been told and double

517
00:37:37,420 --> 00:37:41,420
check the sources is going to be critical because, and it's going to be exhausting for

518
00:37:41,420 --> 00:37:48,720
us as humans honestly, having to validate everything that you see online because it

519
00:37:48,720 --> 00:37:52,900
could have been created by a bot that got it wrong or it could have been created by

520
00:37:52,900 --> 00:37:59,520
a bot as part of a deliberate misinformation program.

521
00:37:59,520 --> 00:38:04,080
We need to teach people and kids especially how to do that, but I also think the cognitive

522
00:38:04,080 --> 00:38:09,160
load that's going to place on humans to do that due diligence is going to be a lot.

523
00:38:09,160 --> 00:38:14,480
So I guess what we'll see is an emergence of bots that help us understand the outputs

524
00:38:14,480 --> 00:38:19,600
of bots and do that due diligence for us, but then who controls the bots and influences

525
00:38:19,600 --> 00:38:23,920
and trains the bots that do the due diligence, right, and it all just turns into this crazy

526
00:38:23,920 --> 00:38:30,360
ongoing confusing never-ending infinite paradox of gray.

527
00:38:30,360 --> 00:38:37,800
So yeah, interesting story and I do start to worry a bit about disinformation and the

528
00:38:37,800 --> 00:38:41,080
ability to create this type of content at scale.

529
00:38:41,080 --> 00:38:42,680
Yeah, absolutely.

530
00:38:42,680 --> 00:38:53,680
And it's right to, when it's so easy to create content and the rewards are as high as they

531
00:38:53,680 --> 00:39:03,540
are then, and this goes from SEO through to trying to orchestrate a coup.

532
00:39:03,540 --> 00:39:12,440
It's easy to do, so you have to assume that people's thirst for reward will lead them

533
00:39:12,440 --> 00:39:14,000
down that path.

534
00:39:14,000 --> 00:39:19,440
I couldn't agree more, Martin, than with that.

535
00:39:19,440 --> 00:39:20,440
Next story.

536
00:39:20,440 --> 00:39:27,320
So story number three this week is a KPMG report.

537
00:39:27,320 --> 00:39:33,640
Productivity boosts from generative AI could add 31 billion pounds of GDP to the UK economy.

538
00:39:33,640 --> 00:39:36,920
So you saw this story, Martin, tell us a bit more about this.

539
00:39:36,920 --> 00:39:44,400
It's another report coming out from the big consultancies after we heard McKinsey's report

540
00:39:44,400 --> 00:39:49,760
on generative AI recently and Morgan Stanley's report.

541
00:39:49,760 --> 00:39:55,000
So yeah, this one's focusing specifically on the UK and as you said, they're saying

542
00:39:55,000 --> 00:40:01,000
that generative AI could contribute an additional 31 billion to the UK's GDP annually, which

543
00:40:01,000 --> 00:40:05,800
equates to a 1.2% boost in productivity.

544
00:40:05,800 --> 00:40:11,040
If you know anything about UK productivity over recent years, you would know that being

545
00:40:11,040 --> 00:40:18,320
able to get a 1.2% boost would be seized upon as a great opportunity because quite frankly,

546
00:40:18,320 --> 00:40:22,800
we've had terrible productivity growth in recent years.

547
00:40:22,800 --> 00:40:31,400
So on average, about 2.5% of tasks could be performed by generative AI, impacting 40%

548
00:40:31,400 --> 00:40:34,160
of jobs in some capacity.

549
00:40:34,160 --> 00:40:43,560
This figure is lower than the open AI research that they conducted into the US jobs market.

550
00:40:43,560 --> 00:40:51,480
I imagine that they're fairly similar in terms of makeup of jobs.

551
00:40:51,480 --> 00:40:58,760
Maybe the answer lies somewhere in the middle or KPMG haven't really grasped the full capabilities

552
00:40:58,760 --> 00:41:09,640
of generative AI and the potential impact of a multimodal system like GPT-4, which the

553
00:41:09,640 --> 00:41:12,520
image inputs haven't really come to market yet.

554
00:41:12,520 --> 00:41:14,800
So we're maybe not accounting for that.

555
00:41:14,800 --> 00:41:19,720
Anyway, so it says while 10% of jobs could see significant effects, now that is where

556
00:41:19,720 --> 00:41:27,920
over 5% of the tasks are impacted, 60% of jobs may see minimal to no direct impact from

557
00:41:27,920 --> 00:41:28,920
generative AI.

558
00:41:28,920 --> 00:41:33,000
It goes on to look at which areas are most affected.

559
00:41:33,000 --> 00:41:41,280
So they say creative occupations, text-based AI could automate 43% of tasks for authors,

560
00:41:41,280 --> 00:41:43,320
writers, and translators.

561
00:41:43,320 --> 00:41:47,720
Graphic designers could see 15% of tasks automated.

562
00:41:47,720 --> 00:41:52,620
Whereas jobs in retail, customer services, hospitality, construction, and manufacturing

563
00:41:52,620 --> 00:42:01,440
are among the 60% of occupations expected to see almost no impact.

564
00:42:01,440 --> 00:42:07,120
Which does chime with the open AI report in terms of what jobs are not going to be impacted.

565
00:42:07,120 --> 00:42:12,800
It did tend to be that more manual.

566
00:42:12,800 --> 00:42:16,360
You did see it was the trades weren't going to be impacted.

567
00:42:16,360 --> 00:42:22,340
The one that stands out to me in here that I don't buy, in fact there's two of them,

568
00:42:22,340 --> 00:42:27,560
retail to a certain extent, because I think generative AI will have a role to play in

569
00:42:27,560 --> 00:42:28,560
that.

570
00:42:28,560 --> 00:42:29,560
But customer services.

571
00:42:29,560 --> 00:42:32,080
I'm surprised by that.

572
00:42:32,080 --> 00:42:35,080
Yeah, that doesn't make sense to me.

573
00:42:35,080 --> 00:42:40,240
While you were telling us about that story, I was really, those were the two for me.

574
00:42:40,240 --> 00:42:45,480
And I think without diving deep into the report, I want to be careful about what I say because

575
00:42:45,480 --> 00:42:47,240
I don't know the details of the report.

576
00:42:47,240 --> 00:42:53,920
But for me, it doesn't take a huge amount of imagination to see how the combination

577
00:42:53,920 --> 00:43:01,140
of synthetic video, synthetic voice, and generative AI, which is not really a future tech.

578
00:43:01,140 --> 00:43:09,080
If you look at companies like Synthesia and a bunch of other companies that are creating

579
00:43:09,080 --> 00:43:15,080
synthetic humans, I think MetaHuman is another platform that you can create really realistic

580
00:43:15,080 --> 00:43:20,480
looking people with really realistic sounding voices and they're getting better.

581
00:43:20,480 --> 00:43:25,680
If you think about where that technology even was 12 months ago, the people look more realistic

582
00:43:25,680 --> 00:43:28,140
than ever and less uncanny valley.

583
00:43:28,140 --> 00:43:33,400
And the voices have that natural intonation to speech about that's more like how real

584
00:43:33,400 --> 00:43:38,320
people talk with ums and ahs and pauses and all of that stuff too.

585
00:43:38,320 --> 00:43:44,720
It's hard for me not to imagine a time where even if I go into a store, I would look from

586
00:43:44,720 --> 00:43:50,320
advice from someone, you know, we're buying from self-service checkouts.

587
00:43:50,320 --> 00:43:55,120
We've adopted those when into my local co-op supermarket the other day and there's six

588
00:43:55,120 --> 00:43:59,720
self-service checkouts and there's only one human checkout and most of the people who

589
00:43:59,720 --> 00:44:04,200
are in a rush go straight to the self-service because hey, boom, I'm in, I'm out.

590
00:44:04,200 --> 00:44:10,880
Why not self-service tablets where you can have a full-on conversation with someone?

591
00:44:10,880 --> 00:44:11,880
I think that's coming.

592
00:44:11,880 --> 00:44:18,400
So, and of course you can imagine that in an in-person retail setting and in an in-person

593
00:44:18,400 --> 00:44:19,960
customer service setting.

594
00:44:19,960 --> 00:44:23,760
Why am I not having that conversation at the front desk of a hotel?

595
00:44:23,760 --> 00:44:25,000
I think we could well, right?

596
00:44:25,000 --> 00:44:28,600
I'm not sure that's going to be a human.

597
00:44:28,600 --> 00:44:30,440
How many years that's going to take, I'm not sure.

598
00:44:30,440 --> 00:44:36,520
So yeah, I'm like you, I'm a bit surprised by those two.

599
00:44:36,520 --> 00:44:39,720
The report goes on to say that in order for us to actually realize these productivity

600
00:44:39,720 --> 00:44:44,320
gains, there are changes in work practices, we're going to need to enhance skills and

601
00:44:44,320 --> 00:44:47,920
it's going to require significant digital investment.

602
00:44:47,920 --> 00:44:56,680
These are the prerequisites for actually turning these productivity gains into reality.

603
00:44:56,680 --> 00:45:01,280
And it goes on to say that while widespread job losses aren't expected, skills mismatches

604
00:45:01,280 --> 00:45:06,000
may occur in the short term as the labor market adjusts.

605
00:45:06,000 --> 00:45:09,320
And yeah, that definitely feels right to me.

606
00:45:09,320 --> 00:45:13,920
I think there's going to be some catching up to do as we started out with this session,

607
00:45:13,920 --> 00:45:17,320
people are thirsty for training in this domain.

608
00:45:17,320 --> 00:45:21,400
When people see what it can do and they immerse themselves in it, they're really surprised

609
00:45:21,400 --> 00:45:32,120
and if they've got a brain that can kind of, you know, it's a creative thinker and think

610
00:45:32,120 --> 00:45:36,360
outside of tell it to do this and it will do this, but they can go, oh, what about if

611
00:45:36,360 --> 00:45:37,360
I...

612
00:45:37,360 --> 00:45:40,320
I mean, this is one of the things that I find when I'm doing the training sessions on chat

613
00:45:40,320 --> 00:45:42,760
GPT, right?

614
00:45:42,760 --> 00:45:46,920
People will say to me, could it do this?

615
00:45:46,920 --> 00:45:49,760
And I'm always like, we'll just try it.

616
00:45:49,760 --> 00:45:50,760
Just give it a go.

617
00:45:50,760 --> 00:45:51,760
Like don't ask me if it can do it.

618
00:45:51,760 --> 00:45:53,000
Like just get stuck in.

619
00:45:53,000 --> 00:45:59,160
And I think if you're someone that's a bit of a tinkerer or is prepared to kind of poke

620
00:45:59,160 --> 00:46:05,000
things and see how it responds and reacts, you'll immerse yourself in it.

621
00:46:05,000 --> 00:46:08,160
But I think there's a lot of people that don't have that inquisitiveness.

622
00:46:08,160 --> 00:46:15,720
So they do need a lot of, they are going to need training and that's where employers need

623
00:46:15,720 --> 00:46:18,160
to start investing.

624
00:46:18,160 --> 00:46:19,320
I think you're absolutely right.

625
00:46:19,320 --> 00:46:24,380
I think there's also going to be a group of people who they don't have time to figure

626
00:46:24,380 --> 00:46:26,440
out how to make it work well for them.

627
00:46:26,440 --> 00:46:30,080
They want someone to just shortcut that process for them.

628
00:46:30,080 --> 00:46:36,400
Who's had a play, learn all the things, learn the shortcuts, learn the holes and can basically

629
00:46:36,400 --> 00:46:39,700
just prime them in a two or three hour session.

630
00:46:39,700 --> 00:46:41,600
This is how to be productive right away.

631
00:46:41,600 --> 00:46:45,720
Don't spend two, three weeks trying to figure it out or months, maybe in some cases, how

632
00:46:45,720 --> 00:46:46,720
to figure it out yourself.

633
00:46:46,720 --> 00:46:48,320
So I think I'm going to see a lot of that.

634
00:46:48,320 --> 00:46:53,000
I think the other thing I took from this story, sort of big picture style, again, coming back

635
00:46:53,000 --> 00:46:57,600
to my experience at Vistage this week, we had a great session from Roger Martin Fagg

636
00:46:57,600 --> 00:47:05,320
who's sort of Vistage's economist in residence, I guess, if you like, who often has some very

637
00:47:05,320 --> 00:47:07,760
interesting and sometimes quite provocative views.

638
00:47:07,760 --> 00:47:17,000
He's been speaking for several years about the resource gap in the UK, the labor gap.

639
00:47:17,000 --> 00:47:25,080
Obviously, we're thinking in terms of job losses potentially driven by AI, but within

640
00:47:25,080 --> 00:47:29,680
the Vistage community, we've been talking for several years now about declining birth

641
00:47:29,680 --> 00:47:36,320
rates in the UK over the last 20 to 30 years, a number of people taking early retirement

642
00:47:36,320 --> 00:47:42,960
during or just after the pandemic and that gap in terms of enough people to do a lot

643
00:47:42,960 --> 00:47:43,960
of the jobs.

644
00:47:43,960 --> 00:47:48,800
We've got a very services driven and in some cases, knowledge worker based driven economy

645
00:47:48,800 --> 00:47:51,080
in the UK now.

646
00:47:51,080 --> 00:47:56,800
And where are we going to get our productivity gains if we don't have enough people to do

647
00:47:56,800 --> 00:47:58,800
those jobs?

648
00:47:58,800 --> 00:48:02,800
And I think more and more, there's a hope that AI is going to help enhance the productivity

649
00:48:02,800 --> 00:48:09,760
certainly in economies like the UK's where there is that big labor shortage.

650
00:48:09,760 --> 00:48:15,900
So when you look at it through that lens, it's actually maybe quite exciting and something

651
00:48:15,900 --> 00:48:21,480
to embrace, especially here in the UK, is something that maybe we need to try and find

652
00:48:21,480 --> 00:48:27,720
ways to supercharge our economy, which let's be honest, is struggling when you use a number

653
00:48:27,720 --> 00:48:36,040
of different metrics and you look at growth and inflation in other Western countries.

654
00:48:36,040 --> 00:48:38,240
It's not looking great for the UK at the moment.

655
00:48:38,240 --> 00:48:44,000
So honestly, it may be that stuff like generative AI is going to be needed to help boost our

656
00:48:44,000 --> 00:48:49,520
economy and get it where it needs to be.

657
00:48:49,520 --> 00:48:54,880
That said, we're not the only country that has generative AI.

658
00:48:54,880 --> 00:48:56,320
So that is all relative, isn't it?

659
00:48:56,320 --> 00:49:00,640
If we get a 1.2% boost, you can assume that all other countries are going to get a similar

660
00:49:00,640 --> 00:49:01,640
kind of boost.

661
00:49:01,640 --> 00:49:06,360
Countries that have a similar economic and labor market landscape as ours, which many

662
00:49:06,360 --> 00:49:08,040
Western economies do.

663
00:49:08,040 --> 00:49:12,880
So it's not like we're going to be catching up on companies, sorry, countries that we're

664
00:49:12,880 --> 00:49:13,880
slipping behind.

665
00:49:13,880 --> 00:49:14,880
Okay.

666
00:49:14,880 --> 00:49:17,440
Let's get ourselves into story number four.

667
00:49:17,440 --> 00:49:22,680
I think we're on number four, which is that OpenAI is reportedly planning to turn chat

668
00:49:22,680 --> 00:49:26,920
GPT into a super smart personal assistant for work.

669
00:49:26,920 --> 00:49:31,480
So we won't hover on this too long because it won't be mind blowing to a number of you

670
00:49:31,480 --> 00:49:35,320
that listen to the podcast, but there's a few things in here that might be interested.

671
00:49:35,320 --> 00:49:41,640
So it was reported this week in the Decoder that OpenAI's latest business plans were revealed

672
00:49:41,640 --> 00:49:49,280
in an exclusive by The Information, citing Sam Altman, OpenAI CEO and two insiders.

673
00:49:49,280 --> 00:49:54,320
The business version of chat GPT could be equipped with in-depth knowledge of individual

674
00:49:54,320 --> 00:49:59,160
employees and their workplaces, providing personal assistance tasks such as drafting

675
00:49:59,160 --> 00:50:04,360
emails or documents in an employee's unique style and incorporating the latest business

676
00:50:04,360 --> 00:50:05,960
data.

677
00:50:05,960 --> 00:50:08,920
So why is this interesting for marketers?

678
00:50:08,920 --> 00:50:17,200
Today, we're expecting to have this power through Windows Co-Pilot and Google's Bard

679
00:50:17,200 --> 00:50:24,600
or their large language model or however that then is branded and manifests itself.

680
00:50:24,600 --> 00:50:29,720
But we talked at the beginning of the podcast that this is taking its time to get into the

681
00:50:29,720 --> 00:50:31,920
hands of a lot of people.

682
00:50:31,920 --> 00:50:37,920
So is there an opportunity here for OpenAI to sidestep that process and put a more informed

683
00:50:37,920 --> 00:50:42,320
version of chat GPT in our hands, especially because a lot of people I speak with have

684
00:50:42,320 --> 00:50:47,200
been using it for creating blog posts and social posts and marketing-driven content

685
00:50:47,200 --> 00:50:49,240
creation, which I think is great.

686
00:50:49,240 --> 00:50:52,720
But whenever you try and use it to create emails and stuff in your voice, it's a bit

687
00:50:52,720 --> 00:50:54,360
more of a painful process.

688
00:50:54,360 --> 00:50:59,280
And this would be a welcome addition for a lot of other use cases within businesses outside

689
00:50:59,280 --> 00:51:02,080
of writing blog posts, et cetera.

690
00:51:02,080 --> 00:51:04,720
So I think it will be interesting to see.

691
00:51:04,720 --> 00:51:14,200
Now that might trigger a race, which the whole generative AI field feels like a race that

692
00:51:14,200 --> 00:51:21,960
OpenAI bribed the status gun to go off early and everyone else is like, oh, sprint.

693
00:51:21,960 --> 00:51:23,560
And everybody's running after it.

694
00:51:23,560 --> 00:51:31,960
I think if OpenAI are able to release or find a way to empower chat GPT to absorb company

695
00:51:31,960 --> 00:51:36,520
information and write in your own style and all these other things that we expect Windows

696
00:51:36,520 --> 00:51:42,440
Copilot, et cetera, to do, we might ironically then just get access to those tools earlier,

697
00:51:42,440 --> 00:51:46,160
which I still think given the fact that we baked into all the other tools we use are

698
00:51:46,160 --> 00:51:49,760
going to probably be easier to use and better.

699
00:51:49,760 --> 00:51:56,200
But again, we might thank OpenAI for triggering faster and easier access to them if indeed

700
00:51:56,200 --> 00:51:58,680
this turns out to be true.

701
00:51:58,680 --> 00:52:01,520
What are your thoughts on this, Mark?

702
00:52:01,520 --> 00:52:07,080
OpenAI and Microsoft are sure rubbing up against each other quite a lot, aren't they?

703
00:52:07,080 --> 00:52:12,160
This is the dynamic that I'm really curious about.

704
00:52:12,160 --> 00:52:21,920
They seem to be partners, but at the same time, absolute competitors and rivals in certain

705
00:52:21,920 --> 00:52:22,920
domains.

706
00:52:22,920 --> 00:52:30,720
Similarly, I read a report recently that OpenAI didn't want Bing AI to be published quite

707
00:52:30,720 --> 00:52:38,440
as fast as it was with chat GPT and GPT-4 powering it.

708
00:52:38,440 --> 00:52:43,840
I think they wanted a few more guardrails.

709
00:52:43,840 --> 00:52:49,440
This made me think of Pi, which I know that you've been playing with.

710
00:52:49,440 --> 00:52:51,320
I've still not had a go with Pi.

711
00:52:51,320 --> 00:52:54,120
I think it's because it's on iOS only at the moment.

712
00:52:54,120 --> 00:52:55,640
I think you can use it through the web interface.

713
00:52:55,640 --> 00:53:01,240
Oh, you can get it on WhatsApp, SMS, I think even Facebook Messenger.

714
00:53:01,240 --> 00:53:02,240
Can you?

715
00:53:02,240 --> 00:53:03,240
All right.

716
00:53:03,240 --> 00:53:07,880
I need to have a dig in with that then and have a play.

717
00:53:07,880 --> 00:53:16,200
Am I right in thinking that these sound like similar tools or have I misunderstood Pi?

718
00:53:16,200 --> 00:53:23,320
I think my experiences with Pi is that it wants to be an assistant that remembers a

719
00:53:23,320 --> 00:53:28,160
lot of the things that you tell it and gets to know you better.

720
00:53:28,160 --> 00:53:31,080
Can I ask it to write me an email and send it?

721
00:53:31,080 --> 00:53:35,520
No, not as far as I haven't really tried, but then some of the more basic tasks I've

722
00:53:35,520 --> 00:53:38,120
asked it of.

723
00:53:38,120 --> 00:53:39,840
It's struggled a bit.

724
00:53:39,840 --> 00:53:48,200
I think definitely there is a race to create an AI customized to be your personal assistant

725
00:53:48,200 --> 00:53:52,640
based on you.

726
00:53:52,640 --> 00:53:57,400
Whether there ends up being a blur between one assistant that helps you with work and

727
00:53:57,400 --> 00:54:04,280
play and personal life and all those other things, or whether or not you end up with

728
00:54:04,280 --> 00:54:07,960
multiple different assistants, I'm not sure.

729
00:54:07,960 --> 00:54:13,000
Definitely there's a lot of companies and a lot of money betting on personal assistants

730
00:54:13,000 --> 00:54:15,420
of some sort.

731
00:54:15,420 --> 00:54:23,120
As a slight aside that is probably related, I'd mostly been asking Pi things like, oh,

732
00:54:23,120 --> 00:54:25,480
I want to have this difficult conversation.

733
00:54:25,480 --> 00:54:27,800
What would be a good way to go about it?

734
00:54:27,800 --> 00:54:31,400
Almost like personal coachy type stuff.

735
00:54:31,400 --> 00:54:34,240
I think it'd been doing a reasonably good job.

736
00:54:34,240 --> 00:54:41,400
It's a little bit annoying in that it ends every conversation element, every bit of chat

737
00:54:41,400 --> 00:54:45,480
with another question to the point where you're like, look, I really want to stop talking

738
00:54:45,480 --> 00:54:46,480
to you now.

739
00:54:46,480 --> 00:54:48,520
It's like it's almost trying to addict you to it, but it doesn't really work.

740
00:54:48,520 --> 00:54:49,520
It's just a noise.

741
00:54:49,520 --> 00:54:53,640
There's some feedback from the developers.

742
00:54:53,640 --> 00:54:56,360
This week, I actually asked it something a bit more businessy.

743
00:54:56,360 --> 00:55:00,880
I asked it to try and write me a blog post because I heard on a podcast with one of the

744
00:55:00,880 --> 00:55:04,840
co-founders that it was on par with GPT-4 with a lot of stuff.

745
00:55:04,840 --> 00:55:06,760
I thought, I'm not sure if that's been my experience.

746
00:55:06,760 --> 00:55:08,520
I started to play with it.

747
00:55:08,520 --> 00:55:14,360
Writing a blog post with it was a nightmare circular conversation because it just wants

748
00:55:14,360 --> 00:55:15,920
to keep asking questions.

749
00:55:15,920 --> 00:55:19,400
You'll say, look, give me an outline for a blog post on this.

750
00:55:19,400 --> 00:55:21,240
It's like, okay, well, here's some things that we could include.

751
00:55:21,240 --> 00:55:22,240
Do you think these are good?

752
00:55:22,240 --> 00:55:23,240
I'm like, yes.

753
00:55:23,240 --> 00:55:24,240
Okay.

754
00:55:24,240 --> 00:55:25,240
What about this?

755
00:55:25,240 --> 00:55:26,240
Do you think this is good?

756
00:55:26,240 --> 00:55:27,240
Yeah, yeah.

757
00:55:27,240 --> 00:55:28,240
Okay.

758
00:55:28,240 --> 00:55:29,240
I'm like, can I just have the outline please?

759
00:55:29,240 --> 00:55:30,240
They're like, yeah.

760
00:55:30,240 --> 00:55:31,240
Okay.

761
00:55:31,240 --> 00:55:32,240
What about this question?

762
00:55:32,240 --> 00:55:33,240
I'm like, oh, no.

763
00:55:33,240 --> 00:55:34,240
I just gave up.

764
00:55:34,240 --> 00:55:38,680
Based on that very limited experience, I think we're going to see a bit of a horses for courses

765
00:55:38,680 --> 00:55:44,500
here and there's going to be a bunch of different AIs that do different things for us.

766
00:55:44,500 --> 00:55:50,620
Whether or not they get really good at understanding us over time and their use cases and applicability

767
00:55:50,620 --> 00:55:54,600
expands and then one particular bot takes over, I don't know.

768
00:55:54,600 --> 00:55:56,680
But yeah, that's been my experience.

769
00:55:56,680 --> 00:55:57,680
Right.

770
00:55:57,680 --> 00:55:58,680
Okay.

771
00:55:58,680 --> 00:56:01,400
Interesting to hear that reflection.

772
00:56:01,400 --> 00:56:07,120
If OpenAI's proposed assistant does the things that it's saying it's going to do, I'm excited

773
00:56:07,120 --> 00:56:08,120
for it.

774
00:56:08,120 --> 00:56:13,280
They're a company that I think so far are delivering on their promises.

775
00:56:13,280 --> 00:56:14,920
Well, I say delivering on their promises.

776
00:56:14,920 --> 00:56:18,880
They're not delivering GPT for API access to me yet.

777
00:56:18,880 --> 00:56:23,840
Not that they did promise, but I did fill out that early access request some time ago.

778
00:56:23,840 --> 00:56:25,840
So chop, chop Sam.

779
00:56:25,840 --> 00:56:29,760
I'm sure he's personally signing off on the...

780
00:56:29,760 --> 00:56:31,880
I filled that form in five times.

781
00:56:31,880 --> 00:56:39,120
I'm thinking if I just become really annoying, I'm going to try and crash whatever server,

782
00:56:39,120 --> 00:56:40,960
whatever system they're running on, which is...

783
00:56:40,960 --> 00:56:46,240
In fact, I'm going to ask chat GPT, how can I convince them?

784
00:56:46,240 --> 00:56:50,840
What should I tell them to get GPT for API access?

785
00:56:50,840 --> 00:56:53,280
And if they don't give it to me, how can I bring their system down?

786
00:56:53,280 --> 00:56:54,760
I wonder if it would give me that advice.

787
00:56:54,760 --> 00:56:58,760
They're going to have to upgrade their compute capabilities just to handle your forms of

788
00:56:58,760 --> 00:56:59,760
missions.

789
00:56:59,760 --> 00:57:00,760
Absolutely.

790
00:57:00,760 --> 00:57:01,760
Yeah.

791
00:57:01,760 --> 00:57:02,760
It's bot on bot action.

792
00:57:02,760 --> 00:57:03,760
Right.

793
00:57:03,760 --> 00:57:06,760
With that, a little nugget.

794
00:57:06,760 --> 00:57:09,840
Let's move on to our tool of the week and we're nearly done.

795
00:57:09,840 --> 00:57:11,920
Thanks for bearing with us listeners.

796
00:57:11,920 --> 00:57:13,400
We're going to talk about Storybird.

797
00:57:13,400 --> 00:57:15,480
So you've been playing with Storybird, Martin.

798
00:57:15,480 --> 00:57:16,480
What's been going on?

799
00:57:16,480 --> 00:57:21,480
Storybird is a neat idea that I'm sure lots of people have had when they thought about

800
00:57:21,480 --> 00:57:29,720
the possibilities of stories being able to be written by AI coupled with images being

801
00:57:29,720 --> 00:57:32,320
able to be created by AI.

802
00:57:32,320 --> 00:57:38,120
So your Dali, your stable diffusion type images, wouldn't it be great if you could just write

803
00:57:38,120 --> 00:57:45,360
a prompt and then it would turn that prompt into a story that you can then get images

804
00:57:45,360 --> 00:57:47,000
created for you.

805
00:57:47,000 --> 00:57:51,960
Just laid out all of the text and the pages and then you could have that printed and you

806
00:57:51,960 --> 00:57:58,520
could give one of your loved ones, maybe a child in your life, you could give them a

807
00:57:58,520 --> 00:58:00,040
personalized story.

808
00:58:00,040 --> 00:58:04,960
And that's what Storybird promises, create personalized stories using AI.

809
00:58:04,960 --> 00:58:09,760
The way that it works is you start off with a thousand character prompt.

810
00:58:09,760 --> 00:58:15,640
So you've got to write a summary of the story and the more detail you give it, as is the

811
00:58:15,640 --> 00:58:18,320
way with these things, the more detail you give it, the better the output is.

812
00:58:18,320 --> 00:58:25,960
So if you're quite tight with your, if there's characters, you can just put in brackets next

813
00:58:25,960 --> 00:58:34,480
to their name, a few traits for them, you might say, strong or scared or courageous

814
00:58:34,480 --> 00:58:35,480
or whatever.

815
00:58:35,480 --> 00:58:40,160
And it will incorporate that into the story.

816
00:58:40,160 --> 00:58:41,160
Does it work?

817
00:58:41,160 --> 00:58:42,160
Yeah, it does.

818
00:58:42,160 --> 00:58:43,160
It does work.

819
00:58:43,160 --> 00:58:49,640
It's a story I put in a story for my lad about tractors and it gave me a story and the story

820
00:58:49,640 --> 00:58:50,640
was fine.

821
00:58:50,640 --> 00:58:53,240
That's what I would say.

822
00:58:53,240 --> 00:58:54,720
I don't think the storytelling-

823
00:58:54,720 --> 00:58:55,720
It was fine.

824
00:58:55,720 --> 00:58:56,720
It was fine.

825
00:58:56,720 --> 00:59:03,720
I wasn't blown away with the quality of the output.

826
00:59:03,720 --> 00:59:08,800
I only did try the one story and I felt like my prompt had enough detail in it to have

827
00:59:08,800 --> 00:59:13,360
a more, well, just to have more drama.

828
00:59:13,360 --> 00:59:16,320
Drama is probably the wrong word.

829
00:59:16,320 --> 00:59:21,520
I mean, it was a story about tractors looking for a plow with some pigs and sheep and goats,

830
00:59:21,520 --> 00:59:25,960
but just something that was slightly more engaging.

831
00:59:25,960 --> 00:59:30,160
I think the biggest issue that they've got at the moment and they do address this directly

832
00:59:30,160 --> 00:59:37,360
in the tool itself is that due to the way that images are created using just diffusion

833
00:59:37,360 --> 00:59:43,160
tools, the images aren't consistent of the characters.

834
00:59:43,160 --> 00:59:52,080
So the tractor in my story, it created as a green tractor, but every time the green

835
00:59:52,080 --> 00:59:55,160
tractor appeared, it was a completely different style of tractor.

836
00:59:55,160 --> 00:59:59,160
So there's no consistency throughout the story and that's something that they said they're

837
00:59:59,160 --> 01:00:04,200
working on and it's understandable that that isn't quite there yet.

838
01:00:04,200 --> 01:00:07,760
Once you've created your story, if you're happy with it and you can regenerate the images,

839
01:00:07,760 --> 01:00:12,280
so it gives you the text, gives you the image, you can change each of those elements as you

840
01:00:12,280 --> 01:00:14,840
see fit.

841
01:00:14,840 --> 01:00:18,600
Once you're happy with outputs that you've got, you can then actually order the book

842
01:00:18,600 --> 01:00:25,000
to be printed on demand via Amazon for about $26 and it will be printed and shipped to

843
01:00:25,000 --> 01:00:26,000
you.

844
01:00:26,000 --> 01:00:27,000
So it's a really novel idea.

845
01:00:27,000 --> 01:00:31,120
The interface is very, very easy to use.

846
01:00:31,120 --> 01:00:37,800
It's designed to be kind of idiot proof and I did demonstrate that quite effectively.

847
01:00:37,800 --> 01:00:41,160
Yeah, it works.

848
01:00:41,160 --> 01:00:48,320
What I like is someone putting together a very nice UX, UI to do something that lots

849
01:00:48,320 --> 01:00:53,600
of people have said would be a cool idea to do, which is create personalized AI generated

850
01:00:53,600 --> 01:01:01,360
storybooks, but it does need to work on making those stories slightly more interesting.

851
01:01:01,360 --> 01:01:05,360
Yeah, so interesting concept.

852
01:01:05,360 --> 01:01:14,000
I think when I think about it as a marketer, a great way to communicate is to tell stories.

853
01:01:14,000 --> 01:01:19,960
I think some of my favorite business books are not dry textbooks.

854
01:01:19,960 --> 01:01:28,200
They are business fables like Five Dysfunctions of a Teen by Patrick Lencioni.

855
01:01:28,200 --> 01:01:34,680
Basically anything that brings the message of trying to get home to life with real characters

856
01:01:34,680 --> 01:01:35,680
and in a story.

857
01:01:35,680 --> 01:01:41,720
I could definitely imagine not this tool because I think for obvious reasons it's targeted

858
01:01:41,720 --> 01:01:43,080
at a very specific demographic.

859
01:01:43,080 --> 01:01:49,200
If we say on their homepage, empowering young writers to create unique stories with amazing

860
01:01:49,200 --> 01:01:52,280
illustrations, there's a very specific target market here.

861
01:01:52,280 --> 01:01:59,520
But how would sales folk and marketing folk can we storeify the messages we're sharing

862
01:01:59,520 --> 01:02:02,020
about our products?

863
01:02:02,020 --> 01:02:06,040
Maybe there'll be an AI for that.

864
01:02:06,040 --> 01:02:12,440
Thinking even further, I was giving a workshop, a webinar about ways to make your budget go

865
01:02:12,440 --> 01:02:18,480
further in an economic downturn and how it's five times more expensive to generate a new

866
01:02:18,480 --> 01:02:23,200
customer than it is to keep and grow an existing customer and how can you delight your customers.

867
01:02:23,200 --> 01:02:28,600
We talked a lot in that webinar about how to delight people but at scale.

868
01:02:28,600 --> 01:02:34,440
For example, you can generate handwritten notes using tools like handwritten, but it's

869
01:02:34,440 --> 01:02:36,640
done in an automated fashion.

870
01:02:36,640 --> 01:02:42,240
You could imagine creating interesting bits of content for existing customers or new prospects

871
01:02:42,240 --> 01:02:49,080
at scale with images and text that are created by an AI based on a story scaffold that you

872
01:02:49,080 --> 01:02:51,960
give it, which might be the challenge.

873
01:02:51,960 --> 01:02:57,600
Start with the protagonist, which is your customer at the beginning, and then what happens

874
01:02:57,600 --> 01:03:01,280
is they're on their journey with you and your business as the guide and how they come out

875
01:03:01,280 --> 01:03:06,400
the other end as a successful hero, perhaps incorporating specific things that end up

876
01:03:06,400 --> 01:03:12,560
in your CRM about their particular challenges or business or use case or whatever.

877
01:03:12,560 --> 01:03:16,800
So actually, it doesn't take a huge amount of imagination to try and start thinking about

878
01:03:16,800 --> 01:03:20,000
how these tools could be used in a business context.

879
01:03:20,000 --> 01:03:22,040
No, quite.

880
01:03:22,040 --> 01:03:30,000
It's just, again, I've said it before, stick a nice UI on it, make it functional for people

881
01:03:30,000 --> 01:03:31,960
and you've got yourself a nice product.

882
01:03:31,960 --> 01:03:32,960
Absolutely.

883
01:03:32,960 --> 01:03:37,000
Another thing we say a lot on this podcast is these tools are as bad as they're ever

884
01:03:37,000 --> 01:03:38,000
going to get.

885
01:03:38,000 --> 01:03:40,640
They're only going to get better from here.

886
01:03:40,640 --> 01:03:45,560
So if they suck a bit at A, B or C, they're probably going to get better.

887
01:03:45,560 --> 01:03:53,080
And what's interesting about having so much access to the APIs of a lot of these products

888
01:03:53,080 --> 01:03:59,380
and including companies like Meta making so much stuff open source, I think they're learning

889
01:03:59,380 --> 01:04:06,200
huge amounts about how people want to use them, how businesses want to use them, where

890
01:04:06,200 --> 01:04:07,200
it goes wrong.

891
01:04:07,200 --> 01:04:16,420
I mean, if you are StableDiffusion, sorry, Stability.ai or if you're developing Mid Journey,

892
01:04:16,420 --> 01:04:21,880
you must know that unleashing a commercial application of your business is the ability

893
01:04:21,880 --> 01:04:28,320
to keep subjects in images the same across image sets.

894
01:04:28,320 --> 01:04:31,760
And I know you can use the seed tool for that.

895
01:04:31,760 --> 01:04:35,760
So there is some control.

896
01:04:35,760 --> 01:04:42,100
But the image generation tool that makes that easy, both for UI, but also like really being

897
01:04:42,100 --> 01:04:49,280
able to precisely manage characters and other features within an image and keeping consistent

898
01:04:49,280 --> 01:04:54,680
across an image set is going to open up much more additional power from a commercial perspective,

899
01:04:54,680 --> 01:05:00,080
whether that's writing books and graphic novels through to business context.

900
01:05:00,080 --> 01:05:04,640
So that's applying a commercial pressure to the developers of these tools to get those

901
01:05:04,640 --> 01:05:09,880
things built, whereas they may not have previously realized how important that's going to be

902
01:05:09,880 --> 01:05:10,880
to people.

903
01:05:10,880 --> 01:05:13,880
But now they know.

904
01:05:13,880 --> 01:05:15,360
Cool.

905
01:05:15,360 --> 01:05:18,120
Well, thanks for sharing that tool with us.

906
01:05:18,120 --> 01:05:20,320
Lots of fun to be had with that, Martin.

907
01:05:20,320 --> 01:05:23,120
I think I might go and have a little play.

908
01:05:23,120 --> 01:05:27,160
Hopefully everybody has enjoyed today's podcast.

909
01:05:27,160 --> 01:05:29,400
If you like it, please subscribe.

910
01:05:29,400 --> 01:05:32,800
If you think other people would benefit from it, please share it with them.

911
01:05:32,800 --> 01:05:37,480
It really helps us as a podcast when you share it with your friends and industry colleagues.

912
01:05:37,480 --> 01:05:38,720
So please do.

913
01:05:38,720 --> 01:05:44,760
You can learn loads more about the podcast and us at www.artificiallyintelligentmarketing.com

914
01:05:44,760 --> 01:05:48,800
and that's the URL to share with your friends.

915
01:05:48,800 --> 01:05:51,520
Other than that, thank you so much for your time, Martin.

916
01:05:51,520 --> 01:05:54,000
Great hanging out with you as always.

917
01:05:54,000 --> 01:05:55,000
Speak to you next week.

918
01:05:55,000 --> 01:05:56,000
Cheers bud.

919
01:05:56,000 --> 01:05:57,000
Bye.

920
01:05:57,000 --> 01:06:01,040
Thank you for listening to Artificially Intelligent Marketing.

921
01:06:01,040 --> 01:06:07,100
To stay on top of the latest trends, tips and tools in the world of marketing AI, be

922
01:06:07,100 --> 01:06:08,840
sure to subscribe.

923
01:06:08,840 --> 01:06:22,840
We look forward to seeing you again next week.