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 --> 00:00:18,520
results from your marketing efforts.

4
00:00:18,520 --> 00:00:23,920
Now let's join our hosts, Paul Avery and Martin Broadhurst.

5
00:00:23,920 --> 00:00:28,360
Welcome to episode 19 of Artificially Intelligent Marketing.

6
00:00:28,360 --> 00:00:34,160
It's me, Paul Avery, on a solo cast this week because our good friend and fellow co-host

7
00:00:34,160 --> 00:00:36,360
Martin is out and about.

8
00:00:36,360 --> 00:00:39,560
For those of you who are regular listeners of the podcast, you'll know that he's been

9
00:00:39,560 --> 00:00:44,400
out presenting at the AI Marketing Conference, Mike on in Cleveland this week.

10
00:00:44,400 --> 00:00:49,680
He has recorded us a little update from the event, which we'll play for you a bit later

11
00:00:49,680 --> 00:00:51,480
in the episode.

12
00:00:51,480 --> 00:00:56,360
Until then, you'll start with me, I'm afraid, with two main agenda items.

13
00:00:56,360 --> 00:01:00,600
The first one is to cover off all those news items that we didn't get to cover in last

14
00:01:00,600 --> 00:01:07,200
week's episode to bring us up to date after having a couple of weeks off for the summer.

15
00:01:07,200 --> 00:01:11,200
Then we're going to cover this week's news and there are some zingers in there.

16
00:01:11,200 --> 00:01:13,160
So let's jump straight into it.

17
00:01:13,160 --> 00:01:20,120
So first and foremost, Shopify has launched a new AI driven support agent called Sidekick.

18
00:01:20,120 --> 00:01:24,160
In the example video, you can see Sidekick answering general questions about running

19
00:01:24,160 --> 00:01:29,960
a business, providing possible answers for trends within a user Shopify data.

20
00:01:29,960 --> 00:01:35,320
So for example, why might there have been a drop off in sales this quarter?

21
00:01:35,320 --> 00:01:39,800
It also makes it easier to take bulk actions like putting a whole bunch of products on

22
00:01:39,800 --> 00:01:43,400
sale or adding a product line to the company's homepage.

23
00:01:43,400 --> 00:01:50,400
So one can imagine that using Sidekick is going to make managing your Shopify site so

24
00:01:50,400 --> 00:01:55,680
much easier for Shopify users, making it also easier for them to analyze their data perhaps

25
00:01:55,680 --> 00:02:00,640
than ever before and make strategic decisions about their business.

26
00:02:00,640 --> 00:02:06,120
Now according to the Neuron, example Sidekick tasks include things like crafting a blog

27
00:02:06,120 --> 00:02:12,580
post announcing say a new product, discounting certain products automatically, composing

28
00:02:12,580 --> 00:02:16,280
an FAQ about a product and even generating monthly sales reports.

29
00:02:16,280 --> 00:02:20,960
So we can absolutely start to imagine that not only is this going to be powerful for

30
00:02:20,960 --> 00:02:26,760
Shopify users, it's starting to provide a little hint around what AI powered agents

31
00:02:26,760 --> 00:02:30,240
can do when they're connected to your specific data.

32
00:02:30,240 --> 00:02:37,160
So not just using chat bots like ChatGPT or Claude to help you produce content based on

33
00:02:37,160 --> 00:02:43,480
their training data, but really leveraging natural language conversations with chat bots

34
00:02:43,480 --> 00:02:46,360
about your own data.

35
00:02:46,360 --> 00:02:50,840
I think this is also why we're seeing lots of providers that are connected to lots of

36
00:02:50,840 --> 00:02:56,800
your internal data like your CRM or your project management systems are probably poised to

37
00:02:56,800 --> 00:03:03,960
have an even bigger impact on your business than external large language models like ChatGPT

38
00:03:03,960 --> 00:03:06,560
because they're connected to your data.

39
00:03:06,560 --> 00:03:11,760
Shopify's example is a grade one chat spot from HubSpot is really starting to mature as

40
00:03:11,760 --> 00:03:15,720
a tool as well and I think we're going to see a lot more of this over the coming months.

41
00:03:15,720 --> 00:03:25,040
In fact, on this same track, we saw that also Wix, the website and hosting platform has

42
00:03:25,040 --> 00:03:30,960
also brought out an AI driven website builder that integrates with OpenAI's large language

43
00:03:30,960 --> 00:03:37,100
models, so GPT 3.5, GPT 4 and the tool will make it easier for you to build unique high

44
00:03:37,100 --> 00:03:40,960
quality websites just from text prompts.

45
00:03:40,960 --> 00:03:46,000
When you look at the teaser video, it really doesn't look like it provides high quality

46
00:03:46,000 --> 00:03:47,000
outputs.

47
00:03:47,000 --> 00:03:50,440
I haven't had a chance to play with it yet, so I can't speak to it, but again, it's another

48
00:03:50,440 --> 00:03:58,480
example I think of the emerging capabilities of these tools to make business management,

49
00:03:58,480 --> 00:04:03,520
marketing management, marketing production faster and easier by basically asking the

50
00:04:03,520 --> 00:04:09,440
tool in natural language for what you want and then it automatically generating an output.

51
00:04:09,440 --> 00:04:15,560
Some other news that we should cover is how seven leading AONI companies in the US have

52
00:04:15,560 --> 00:04:20,640
all agreed to manage the risks posed by the technology according to the White House.

53
00:04:20,640 --> 00:04:28,320
These companies are Amazon, Google, IBM, Microsoft, Nvidia, Oracle and Salesforce.

54
00:04:28,320 --> 00:04:31,840
The safeguards they're looking to put in place include things like transparency, so

55
00:04:31,840 --> 00:04:38,440
providing clear explanations of how AI systems make decisions, ensuring fairness and non-discrimination

56
00:04:38,440 --> 00:04:43,400
so that these AI systems do not discriminate against individuals or groups, for example,

57
00:04:43,400 --> 00:04:45,680
based on their training data.

58
00:04:45,680 --> 00:04:49,880
Safety and security, so taking those steps to ensure that AI systems are secure and safe

59
00:04:49,880 --> 00:04:51,520
to use.

60
00:04:51,520 --> 00:04:56,680
Human control, so making sure that humans can override AI decisions if necessary.

61
00:04:56,680 --> 00:05:01,640
And then finally, privacy to protect personal information and ensure that AI systems are

62
00:05:01,640 --> 00:05:04,880
used in a way that respects users' privacy.

63
00:05:04,880 --> 00:05:08,880
The companies have committed to developing a system to watermark all forms of content

64
00:05:08,880 --> 00:05:14,640
as well as part of this, so text, images, audio, video, so that users will know when

65
00:05:14,640 --> 00:05:16,280
the technology has been used.

66
00:05:16,280 --> 00:05:21,120
Now no timeline has been given or any technical details on how that will be achieved, but

67
00:05:21,120 --> 00:05:29,080
certainly we can imagine in a world of easier and easier deepfakes during things like presidential

68
00:05:29,080 --> 00:05:33,600
election cycles or other government cycles in other countries, just how important it's

69
00:05:33,600 --> 00:05:37,640
going to be for users to be able to tell what's real and what's not.

70
00:05:37,640 --> 00:05:41,760
And we also live in a world of fake news, where it's going to be easier and easier to

71
00:05:41,760 --> 00:05:48,240
generate fake news about well-known people in the public eye producing videos that to

72
00:05:48,240 --> 00:05:50,960
all intents and purposes look and sound like the real people.

73
00:05:50,960 --> 00:05:56,320
So having those watermarks and that ability to really make it clear what's been generated

74
00:05:56,320 --> 00:06:01,520
by an AI system and what's real could be one of the most important steps that we see in

75
00:06:01,520 --> 00:06:06,480
terms of making it easier for us humans to figure out what's been generated and what

76
00:06:06,480 --> 00:06:09,240
we can actually trust is real.

77
00:06:09,240 --> 00:06:15,040
As another news item, many of you will have seen that Elon Musk has deputed his ex-company,

78
00:06:15,040 --> 00:06:21,160
ex-AI, staff with talent from across the ways, so team members that have worked at places

79
00:06:21,160 --> 00:06:23,580
like Google and OpenAI.

80
00:06:23,580 --> 00:06:28,480
In terms of what they're looking to achieve, it's pretty vague so far, but it seems like

81
00:06:28,480 --> 00:06:33,160
there are plans for them to create, in very common words, a good AGI to help us understand

82
00:06:33,160 --> 00:06:35,080
the true nature of the universe.

83
00:06:35,080 --> 00:06:41,000
And then this week, that saw Twitter soft rebranded as X. And if you're a Twitter user,

84
00:06:41,000 --> 00:06:47,360
you will see that you now have an X logo on your phone, for example, not a little bird.

85
00:06:47,360 --> 00:06:53,160
That has caused a few sticking points as that migration has tried to make being triggered.

86
00:06:53,160 --> 00:06:57,280
Be interesting to see how Elon and the team jump through some of the hurdles that are

87
00:06:57,280 --> 00:07:01,120
being put in front of them as they try and transition to the name of X for the whole

88
00:07:01,120 --> 00:07:05,080
platform including what used to be called Twitter.

89
00:07:05,080 --> 00:07:10,160
In other news, Apple is reportedly working on AI products to rival giants like OpenAI

90
00:07:10,160 --> 00:07:15,560
and Google, with the chatbot project internally known as Apple GPT.

91
00:07:15,560 --> 00:07:21,440
There are no plans to release it yet, but this is an early indication that Apple's

92
00:07:21,440 --> 00:07:27,240
not just going to sit back and watch a number of other large and small companies emerge

93
00:07:27,240 --> 00:07:34,360
in the AI driven natural language prompt driven chatbot world.

94
00:07:34,360 --> 00:07:38,500
They're actually going to try and build their own one too.

95
00:07:38,500 --> 00:07:42,840
We will probably have all seen at this point the Hollywood writers and actors that have

96
00:07:42,840 --> 00:07:44,200
gone on strike.

97
00:07:44,200 --> 00:07:48,800
Lots of issues around this, but AI is certainly a major one of them.

98
00:07:48,800 --> 00:07:53,160
And we touched a little bit on that in last week's episode.

99
00:07:53,160 --> 00:07:57,400
We also saw some interesting marketing AI stuff over the last week or two.

100
00:07:57,400 --> 00:08:01,400
So you may have noticed Jen.ai from Virgin.

101
00:08:01,400 --> 00:08:08,820
So a supposedly AI driven version of Jennifer Lopez, JLo, which went viral.

102
00:08:08,820 --> 00:08:13,920
Although if I'm honest, my take on it is that it was mostly clever marketing and humor and

103
00:08:13,920 --> 00:08:15,480
not much AI.

104
00:08:15,480 --> 00:08:17,200
So I had a bit of a play with it.

105
00:08:17,200 --> 00:08:22,520
And the best I could get it to do in my test is to pronounce names of the people who would

106
00:08:22,520 --> 00:08:25,480
go on the holiday in JLo's voice.

107
00:08:25,480 --> 00:08:27,360
And that wasn't even with lip syncing.

108
00:08:27,360 --> 00:08:32,720
So when Jennifer JLo actually said the custom content, if you like, it was just the audio

109
00:08:32,720 --> 00:08:36,900
you couldn't actually see her face during those sections.

110
00:08:36,900 --> 00:08:43,280
So interesting, but I think it's more a clever marketing play to jump on the AI hype train

111
00:08:43,280 --> 00:08:49,080
than AI that we can get particularly excited about considering what we know these tools

112
00:08:49,080 --> 00:08:51,360
are now capable of.

113
00:08:51,360 --> 00:08:58,120
On that topic, Mini actually launched an AI driven campaign that leaned much more into

114
00:08:58,120 --> 00:09:01,160
AI, but seemed to generate less hype than what I've seen.

115
00:09:01,160 --> 00:09:06,800
So in this campaign, you went onto the Mini website and then it would take a picture of

116
00:09:06,800 --> 00:09:07,800
your face.

117
00:09:07,800 --> 00:09:14,920
And in fact, I think you have it record a short snippet of you speaking into your camera

118
00:09:14,920 --> 00:09:17,240
on say your laptop.

119
00:09:17,240 --> 00:09:22,560
And then it takes your face and your voice and it creates a narrative that it speaks

120
00:09:22,560 --> 00:09:27,280
back to you when it creates a video where you effectively convince yourself to buy a

121
00:09:27,280 --> 00:09:28,280
Mini.

122
00:09:28,280 --> 00:09:31,880
The whole thing's kind of a bit janky and doesn't really work that well.

123
00:09:31,880 --> 00:09:39,240
I mean, certainly it's me, the face and the audio is not bad, but I'm not sure it sounded

124
00:09:39,240 --> 00:09:42,000
exactly like me in my test.

125
00:09:42,000 --> 00:09:45,640
And the lip syncing wasn't brilliant, but this is where we're headed.

126
00:09:45,640 --> 00:09:51,880
And I've seen a number of examples where people are using AI driven lip syncing and audio

127
00:09:51,880 --> 00:09:55,680
generation to influence what people are saying in videos.

128
00:09:55,680 --> 00:09:59,840
And I don't think it'd be very long before this technology really gets a lot better.

129
00:09:59,840 --> 00:10:05,640
I saw an example with the Lex Friedman podcast where he was talking to Mark Zuckerberg and

130
00:10:05,640 --> 00:10:08,600
they'd used the same approach to have them speak in Hindi.

131
00:10:08,600 --> 00:10:13,840
And yes, I think you could see when you looked at their mouths that they didn't quite look

132
00:10:13,840 --> 00:10:17,640
natural, but they look really quite good and it was rather impressive.

133
00:10:17,640 --> 00:10:24,800
So I think this is again, a really clever ploy by Mini to jump on the AI hype train,

134
00:10:24,800 --> 00:10:30,560
but also something for us to think about how those capabilities open as they open up and

135
00:10:30,560 --> 00:10:35,160
become easy to access and the technology becomes even better.

136
00:10:35,160 --> 00:10:41,040
How can we use those in our own marketing activities?

137
00:10:41,040 --> 00:10:45,120
That story was something that quite close to my heart here in the sciences.

138
00:10:45,120 --> 00:10:52,200
So there was a new study published in Nature's Human Behaviour, exploring how AI could aid

139
00:10:52,200 --> 00:10:57,680
and expand scientific discoveries by predicting and generating hypotheses that humans might

140
00:10:57,680 --> 00:10:58,680
not consider.

141
00:10:58,680 --> 00:11:03,600
So in the paper, researchers built models that generated scientifically promising, but

142
00:11:03,600 --> 00:11:09,000
fairly alien, their words, hypotheses that wouldn't be considered by humans.

143
00:11:09,000 --> 00:11:13,920
The AI was also able to predict with over 40% precision, which to be honest is not particularly

144
00:11:13,920 --> 00:11:20,120
precise, but there you go, was able to predict with that level of precision, the actual people

145
00:11:20,120 --> 00:11:23,960
who would make discovery based on their experiences and relationships.

146
00:11:23,960 --> 00:11:24,960
That's quite interesting.

147
00:11:24,960 --> 00:11:31,000
The overall, the study suggests AI could turbocharge our scientific explorations by helping us make

148
00:11:31,000 --> 00:11:37,000
faster discoveries and even coming up with cool, interesting avenues of exploration that

149
00:11:37,000 --> 00:11:39,200
humans wouldn't naturally think of.

150
00:11:39,200 --> 00:11:43,640
This reminds me of a story that we covered here on the podcast previously, where BeatMine

151
00:11:43,640 --> 00:11:49,080
researchers created AlphaDev to improve computer information processing.

152
00:11:49,080 --> 00:11:54,240
And in this case, the system suggested improvements to the data management that sped up computing

153
00:11:54,240 --> 00:11:58,080
and data movement that humans just wouldn't have thought of.

154
00:11:58,080 --> 00:12:03,320
So I think it's interesting how we're starting to see the emergence of AI tools that can

155
00:12:03,320 --> 00:12:07,720
solve problems in a way that's different from how humans would solve problems because

156
00:12:07,720 --> 00:12:15,080
of the way that these AIs don't quite think like we do, which is pretty interesting.

157
00:12:15,080 --> 00:12:21,120
Another paper, scientific paper this time, is in Science, where some research was published

158
00:12:21,120 --> 00:12:26,440
where writers who chose to use ChatGPT took 40% less time on average to complete their

159
00:12:26,440 --> 00:12:33,520
task and produce work that assessors felt scored 18% higher in quality than the participants

160
00:12:33,520 --> 00:12:35,160
who didn't use it.

161
00:12:35,160 --> 00:12:40,240
So further data here suggesting that you can use tools like ChatGPT if you're in content

162
00:12:40,240 --> 00:12:45,320
production to reduce the amount of time that you need to spend on production and slightly

163
00:12:45,320 --> 00:12:46,680
improve the quality.

164
00:12:46,680 --> 00:12:51,140
I do think this is going to vary a lot depending on who the original writer is and who the

165
00:12:51,140 --> 00:12:54,880
users of these tools are.

166
00:12:54,880 --> 00:13:00,720
Last bit of old news before we get into the new news is that the Mayo Clinic is using

167
00:13:00,720 --> 00:13:04,920
Google's AI ChatPort as part of providing healthcare.

168
00:13:04,920 --> 00:13:10,880
So the Mayo Clinic has been using MedPalm 2 in hospital training since April 2023, so

169
00:13:10,880 --> 00:13:12,840
the last few months.

170
00:13:12,840 --> 00:13:17,580
It's performed comparably to doctors in metrics such as evidence of reasoning, consensus

171
00:13:17,580 --> 00:13:21,440
supported answers and comprehensive accuracy.

172
00:13:21,440 --> 00:13:26,520
According to Google's Senior Research Director Greg Corrado, MedPalm 2 is still in its early

173
00:13:26,520 --> 00:13:31,880
stages and has the potential to expand the beneficial roles of AI in healthcare greatly.

174
00:13:31,880 --> 00:13:37,840
So now that you're all caught up on the older news, now let's jump into this week's

175
00:13:37,840 --> 00:13:38,840
news.

176
00:13:38,840 --> 00:13:40,360
So what do we see this week?

177
00:13:40,360 --> 00:13:47,720
Well, Stability AI released SDXL 1.0 to rival Mid Journey in image generation.

178
00:13:47,720 --> 00:13:54,520
So this is quite an interesting one because the new open model, SDXL 1.0, is a significant

179
00:13:54,520 --> 00:14:00,320
advancement on some of Stability AI's previous image generation models and it can generate

180
00:14:00,320 --> 00:14:06,320
high quality images in any art style, including photorealism and has an improved ability to

181
00:14:06,320 --> 00:14:11,560
interpret language and distinguish between similar terms with different meanings.

182
00:14:11,560 --> 00:14:16,560
It's got a total parameter count of 10.1 billion, making it the largest open image

183
00:14:16,560 --> 00:14:18,400
model to date.

184
00:14:18,400 --> 00:14:21,280
The quality of the image is actually really quite good.

185
00:14:21,280 --> 00:14:23,720
I've been impressed in some of my trials with it.

186
00:14:23,720 --> 00:14:29,800
So it's going to be interesting to see how people use this model, both by accessing tools

187
00:14:29,800 --> 00:14:33,800
like ClipDrop that we'll talk about in a moment, but also because it's open source,

188
00:14:33,800 --> 00:14:37,200
how can they leverage it in their own products?

189
00:14:37,200 --> 00:14:42,920
Reportedly, SDXL 1.0 can actually write readable text.

190
00:14:42,920 --> 00:14:46,760
Now this would be a huge step forward for marketers because for many of you who've played

191
00:14:46,760 --> 00:14:51,800
with Mid Journey or other image generation tools, you'll have noticed that they absolutely

192
00:14:51,800 --> 00:14:53,840
suck at text.

193
00:14:53,840 --> 00:14:59,560
Supposedly, this new model, SDXL, is better, but in my test, if I'm honest, the results

194
00:14:59,560 --> 00:15:01,040
were extremely mixed.

195
00:15:01,040 --> 00:15:04,660
I would say on average, better at producing text.

196
00:15:04,660 --> 00:15:10,400
So I did a test where I asked it to create a billboard with text on it, no images, just

197
00:15:10,400 --> 00:15:11,400
text.

198
00:15:11,400 --> 00:15:15,840
It was like I had to run the generation multiple times and the results were really mixed, nowhere

199
00:15:15,840 --> 00:15:17,920
near production quality.

200
00:15:17,920 --> 00:15:23,200
I've also seen on Twitter people talking about this saying you can get decent text out of

201
00:15:23,200 --> 00:15:24,840
it, but it's very iterative.

202
00:15:24,840 --> 00:15:28,320
You've got to be patient and you've got to find different ways to prompt it and run multiple

203
00:15:28,320 --> 00:15:31,080
generations to get what you want.

204
00:15:31,080 --> 00:15:33,440
Now if you want to have a play with this, you can.

205
00:15:33,440 --> 00:15:38,400
Just visit clippdrop.co, which is a stability AI product with lots of really interesting

206
00:15:38,400 --> 00:15:44,040
image, generation and manipulation tools where you can now generate images using the new

207
00:15:44,040 --> 00:15:50,480
SDXL 1.0 model.

208
00:15:50,480 --> 00:15:56,360
In other news this week, Rewind has released a new personalized AI app for iPhone to expand

209
00:15:56,360 --> 00:15:58,120
upon its Mac app.

210
00:15:58,120 --> 00:16:02,200
So for those of you that haven't heard of this, Rewind is an AI driven app that functions

211
00:16:02,200 --> 00:16:06,640
as a search engine for users' personal digital interactions.

212
00:16:06,640 --> 00:16:12,360
So what it basically does is it allows users to record, store and rewind their work by

213
00:16:12,360 --> 00:16:17,480
recording anything they've seen, said or heard when they've been using, in this case their

214
00:16:17,480 --> 00:16:22,640
iPhone or their Mac and making all that info searchable.

215
00:16:22,640 --> 00:16:28,320
It's powered by OpenAI's GPT-4 and it kind of acts like a personal AI time traveler to

216
00:16:28,320 --> 00:16:32,400
remind you of certain things that you might have been doing a couple of weeks ago that

217
00:16:32,400 --> 00:16:33,400
you can't remember.

218
00:16:33,400 --> 00:16:37,440
You can just ask the tool and because it's indexed what you've been up to, it can give

219
00:16:37,440 --> 00:16:40,440
you some info about what you've been doing.

220
00:16:40,440 --> 00:16:46,620
Now I haven't tried this yet because as far as I know, there's no way to easily customize

221
00:16:46,620 --> 00:16:49,080
what it pays attention to and what it doesn't.

222
00:16:49,080 --> 00:16:52,000
So it feels a bit like a security risk to me.

223
00:16:52,000 --> 00:16:59,400
And also I'm not sure I want at all monitoring absolutely everything that I do on my Mac.

224
00:16:59,400 --> 00:17:03,640
That seems that I can see the benefits, but this is feel like you have to give up quite

225
00:17:03,640 --> 00:17:06,400
a lot of your own privacy to be able to access that.

226
00:17:06,400 --> 00:17:08,840
I'm not quite ready to do that.

227
00:17:08,840 --> 00:17:13,920
By definition for these AI assistants to help us to the max, we're going to have to at some

228
00:17:13,920 --> 00:17:17,920
point accept that they're going to be monitoring everything we're doing.

229
00:17:17,920 --> 00:17:22,440
In fairness, I'm sure Facebook and other tools that have got access to my Mac are doing that

230
00:17:22,440 --> 00:17:28,080
right now, but it may take a bit of a mental shift before most of us are ready to provide

231
00:17:28,080 --> 00:17:33,360
that level of access to these tools in order to get the most benefits.

232
00:17:33,360 --> 00:17:39,760
Next news item here is that Amazon have announced the launch of agents for Bedrock at the AWS

233
00:17:39,760 --> 00:17:41,880
summit in New York this week.

234
00:17:41,880 --> 00:17:47,360
So as described in tech crunch, which is where we read about this, Bedrock is Amazon platform

235
00:17:47,360 --> 00:17:52,400
for building generative AI powered apps using pre-tenant trained models from a bunch of

236
00:17:52,400 --> 00:17:55,000
companies, including Amazon, but also others.

237
00:17:55,000 --> 00:18:02,400
The new feature agents allows customers of Amazon AWS to create conversational agents

238
00:18:02,400 --> 00:18:08,280
that can deliver personalized up to date answers based on the company's own proprietary data.

239
00:18:08,280 --> 00:18:09,560
So that's quite an interesting one.

240
00:18:09,560 --> 00:18:14,200
I know a number of companies out there are really looking at how do we create internal

241
00:18:14,200 --> 00:18:19,240
facing and external facing natural language chat bots for our teams and customers that

242
00:18:19,240 --> 00:18:21,120
are based on our own data.

243
00:18:21,120 --> 00:18:25,880
And here Bedrock agents is looking to enable that.

244
00:18:25,880 --> 00:18:29,960
So a tool like this could be used to create custom service chat bots that can process

245
00:18:29,960 --> 00:18:34,240
orders, tapping into things like internal information about stock levels, et cetera,

246
00:18:34,240 --> 00:18:36,160
to customize each order.

247
00:18:36,160 --> 00:18:41,160
Bedrock agents can also manage and perform tasks by making API calls to company systems.

248
00:18:41,160 --> 00:18:46,980
So really being able to have the agent embed in lots of different information repositories

249
00:18:46,980 --> 00:18:48,960
across your business.

250
00:18:48,960 --> 00:18:52,880
So again, I think this is going to be really, really interesting to see how this plays out

251
00:18:52,880 --> 00:18:54,520
in the future.

252
00:18:54,520 --> 00:19:00,240
On a similar note to this, Cohere, which is another large language model developing company,

253
00:19:00,240 --> 00:19:05,560
has announced Coral, which is a knowledge assistant that's designed to enhance the productivity

254
00:19:05,560 --> 00:19:08,220
of teams within enterprise businesses.

255
00:19:08,220 --> 00:19:12,760
So again, this is another chat bot based tool, a large language model that can tap into the

256
00:19:12,760 --> 00:19:15,040
data within your organization.

257
00:19:15,040 --> 00:19:19,520
Coral can find answers across documents and provide responses back with citations, ensuring

258
00:19:19,520 --> 00:19:23,560
the information provided is verifiable and mitigating against false information.

259
00:19:23,560 --> 00:19:28,360
So it does this not just by looking at your company's own data, but also external data

260
00:19:28,360 --> 00:19:31,040
and things on the web, as far as I understand.

261
00:19:31,040 --> 00:19:34,720
Coral can be customized for different teams within your organization, such as finance,

262
00:19:34,720 --> 00:19:38,800
support, marketing, sales, and it can be made even more powerful by connecting it to data

263
00:19:38,800 --> 00:19:41,840
sources to augment its knowledge base.

264
00:19:41,840 --> 00:19:47,640
And at the moment, there are over 100 integrations across CRMs, project management tools, databases,

265
00:19:47,640 --> 00:19:48,760
et cetera.

266
00:19:48,760 --> 00:19:54,520
In terms of data security and privacy, Coral operates within the user's own server secure

267
00:19:54,520 --> 00:19:59,120
cloud, whether that's through cloud partners or virtual private clouds.

268
00:19:59,120 --> 00:20:03,880
The data used by Coral is never sent to Cohere, ensuring it remains within the user's environment.

269
00:20:03,880 --> 00:20:08,600
Again, this is going to be absolutely critical because a lot of enterprise businesses effectively

270
00:20:08,600 --> 00:20:13,840
have banned their staff from using tools like ChatGPT because they don't want any of their

271
00:20:13,840 --> 00:20:19,240
sensitive internal or customer information making it into the hands of OpenAI because

272
00:20:19,240 --> 00:20:25,320
it's just not clear how OpenAI and other providers of these large language models use the data

273
00:20:25,320 --> 00:20:26,800
that we put into them.

274
00:20:26,800 --> 00:20:31,720
So what in effect, what Cohere is doing here is saying, here's a large language model,

275
00:20:31,720 --> 00:20:36,040
a chatbot for you to use like ChatGPT that can access all of your company's information,

276
00:20:36,040 --> 00:20:39,600
making it really customized and useful for your business.

277
00:20:39,600 --> 00:20:44,000
And by the way, it's super secure because we don't ever get to see what you're basically

278
00:20:44,000 --> 00:20:45,800
doing with your own information.

279
00:20:45,800 --> 00:20:51,000
So I think this could make this very, very attractive to large enterprises and even small

280
00:20:51,000 --> 00:20:53,440
businesses over time, to be honest.

281
00:20:53,440 --> 00:20:58,440
And we should expect to see many other players in this space follow suit.

282
00:20:58,440 --> 00:21:03,760
And in fact, this tool is likely to compete with the likes of Microsoft Copilot and Bard

283
00:21:03,760 --> 00:21:09,380
in terms of natural language chatbots that help you with your work, but do so in a customized

284
00:21:09,380 --> 00:21:12,560
way that's super relevant to your business because they're plugged into all your other

285
00:21:12,560 --> 00:21:14,800
docs and data and stuff like that.

286
00:21:14,800 --> 00:21:21,120
So it is worth noting, we do expect this to come from Microsoft and Google over the next,

287
00:21:21,120 --> 00:21:25,000
we don't know, one, two, five, six, 12 months.

288
00:21:25,000 --> 00:21:30,360
But when we see companies like Cohere moving quite quickly with tools like Coral, expect

289
00:21:30,360 --> 00:21:36,080
it to nudge the Microsofts and the Googles perhaps into faster action.

290
00:21:36,080 --> 00:21:42,160
We did talk last week about how the pricing for Microsoft Copilot for Office 365 has already

291
00:21:42,160 --> 00:21:44,920
been talked about, maybe $30 per user per month.

292
00:21:44,920 --> 00:21:50,900
So maybe that is an early sign that we can expect to see Microsoft's tool coming soon.

293
00:21:50,900 --> 00:21:55,600
In other news, TrackTBT is now available on Android in a number of countries as an app.

294
00:21:55,600 --> 00:21:57,480
And this includes the US and UK.

295
00:21:57,480 --> 00:22:00,360
So I've been able to have a little play, which is great.

296
00:22:00,360 --> 00:22:04,080
It was already available on iPhone, but now you can get it on Android.

297
00:22:04,080 --> 00:22:09,440
What I really love about the tool versus using it on, say, your desktop computer is that

298
00:22:09,440 --> 00:22:14,680
you can dictate your prompt into the app because it has a microphone button.

299
00:22:14,680 --> 00:22:16,880
And it's actually worked really well in my hands.

300
00:22:16,880 --> 00:22:22,760
I think it's better than the dictation app on my keyboard on my phone, for example.

301
00:22:22,760 --> 00:22:27,560
When I installed it, I did see a very interesting disclaimer that specifically warns you not

302
00:22:27,560 --> 00:22:36,080
to share sensitive info as chats may be reviewed by OpenAI's trainers to improve the systems.

303
00:22:36,080 --> 00:22:41,860
It's not clear if this is true even when the chat history is turned off, which was previously

304
00:22:41,860 --> 00:22:45,200
believed to opt you out of having your inputs used by OpenAI.

305
00:22:45,200 --> 00:22:51,000
But the way that this technical warning works is actually a little bit confusing and makes

306
00:22:51,000 --> 00:22:55,760
me wonder that even with chat history off, they are still paying quite a lot of attention

307
00:22:55,760 --> 00:22:57,960
to the information that we send to them.

308
00:22:57,960 --> 00:23:02,720
So I think that's very much worth keeping in mind.

309
00:23:02,720 --> 00:23:09,760
Next news item is the emergence of a generative AI tool on the dark web called Fraud GPT that

310
00:23:09,760 --> 00:23:14,280
offers capabilities to cybercriminals.

311
00:23:14,280 --> 00:23:20,480
So this is a bit of a public service announcement for all of us because it's on the dark web,

312
00:23:20,480 --> 00:23:24,480
it's being positioned as an all-in-one solution with features including writing malicious

313
00:23:24,480 --> 00:23:28,920
code, creating phishing emails, and finding leaks and vulnerabilities.

314
00:23:28,920 --> 00:23:34,240
Sadly, the tool has already got over 3,000 confirmed sales and reviews.

315
00:23:34,240 --> 00:23:39,200
And I think the emergence of cybercrime AI tools like Fraud GPT is probably just the

316
00:23:39,200 --> 00:23:40,820
beginning.

317
00:23:40,820 --> 00:23:44,880
And it's going to make the ability of scammers when they create some phishing emails, et

318
00:23:44,880 --> 00:23:49,760
cetera, to make them look even more realistic than before and at a larger scale because

319
00:23:49,760 --> 00:23:53,080
they'll be automatable in a way that maybe hasn't been possible.

320
00:23:53,080 --> 00:23:58,040
So I think the take home here is that we all need to pay even closer attention to the emails

321
00:23:58,040 --> 00:24:04,800
and messages that we get as the fakes are going to get even more impressive when powered

322
00:24:04,800 --> 00:24:09,200
by the likes of large language models, says Amit to keep in mind.

323
00:24:09,200 --> 00:24:11,720
Next news item is 11 Labs.

324
00:24:11,720 --> 00:24:13,720
They've released a bunch of new voices.

325
00:24:13,720 --> 00:24:18,120
So they were created in collaboration with industry professionals and now they can offer

326
00:24:18,120 --> 00:24:22,360
a wider range of delivery styles, accents, and improved audio quality.

327
00:24:22,360 --> 00:24:25,280
So this is going to provide even more options for those of you who've been exploring using

328
00:24:25,280 --> 00:24:29,240
11 Labs and synthetic voices in your content.

329
00:24:29,240 --> 00:24:32,440
And it's going to allow users to choose from an even broader selection of voices to meet

330
00:24:32,440 --> 00:24:33,440
their needs.

331
00:24:33,440 --> 00:24:37,920
And this even includes voices that can change their delivery.

332
00:24:37,920 --> 00:24:40,600
So from whispering through to basically screaming.

333
00:24:40,600 --> 00:24:47,080
So we can really see that the synthetic voice market is really evolving quickly to meet

334
00:24:47,080 --> 00:24:50,400
a wide range of human spoken audio needs.

335
00:24:50,400 --> 00:24:54,680
And the nuances of performance are improving all the time.

336
00:24:54,680 --> 00:25:00,840
So being able to create audio books at scale, even for people who are self publishing or

337
00:25:00,840 --> 00:25:05,400
for businesses, you can expect that to be made even easier, but even more realistic

338
00:25:05,400 --> 00:25:06,880
by tools like this.

339
00:25:06,880 --> 00:25:11,520
And who knows even this podcast like I'm doing now, which I promise is a real podcast and

340
00:25:11,520 --> 00:25:16,680
I'm the real human speaking here, but how long until you can synthesize my voice and

341
00:25:16,680 --> 00:25:21,480
then deliver probably far more impressive performance than mine using tools like those

342
00:25:21,480 --> 00:25:26,120
released by 11 Labs remains to be seen.

343
00:25:26,120 --> 00:25:30,960
Next is Runway's Gen 2 image to video has been released and it's pretty cool.

344
00:25:30,960 --> 00:25:35,800
Now regular listeners to the podcast will know that we've talked a fair bit about Runway

345
00:25:35,800 --> 00:25:42,680
and their Gen 1 and Gen 2 text to video tools, that we've played with mixed results.

346
00:25:42,680 --> 00:25:46,200
I've seen some people do some really cool stuff on social media, but it felt like the

347
00:25:46,200 --> 00:25:50,600
sort of stuff that probably took ages to put together because it's so iterative trying

348
00:25:50,600 --> 00:25:52,800
to get half decent videos out of these things.

349
00:25:52,800 --> 00:25:59,480
But I think where the image to video tool is a bit of a game changer is it's so much

350
00:25:59,480 --> 00:26:03,720
better than a text prompt for getting an animated thing that you want.

351
00:26:03,720 --> 00:26:07,480
And I've seen people creating really awesome mid-journey images and then pushing those

352
00:26:07,480 --> 00:26:11,300
into Gen 2 and getting some really impressive results.

353
00:26:11,300 --> 00:26:14,880
My own tests with this have been a bit of a mix.

354
00:26:14,880 --> 00:26:19,600
I've noticed that you have to iterate quite a lot, change your prompt a bit to get something

355
00:26:19,600 --> 00:26:20,600
that you want.

356
00:26:20,600 --> 00:26:23,640
And I've had some videos that just didn't work at all and others that actually were

357
00:26:23,640 --> 00:26:24,640
quite impressive.

358
00:26:24,640 --> 00:26:30,040
In fact, I animated the cover for Artificially Intelligent Marketing and got really quite

359
00:26:30,040 --> 00:26:33,200
an interesting psychedelic animation of it.

360
00:26:33,200 --> 00:26:39,000
It's impressive to me how good the model is at understanding all the sort of elements

361
00:26:39,000 --> 00:26:45,440
of the image that you put in as the prompt and finding natural ways to animate it that

362
00:26:45,440 --> 00:26:48,360
I think an animator might also consider.

363
00:26:48,360 --> 00:26:51,840
So go and have a play with this over at Runway.

364
00:26:51,840 --> 00:26:57,000
You can set up a free account and have a few goes with Gen 2 before you need to pay.

365
00:26:57,000 --> 00:26:59,040
I think these tools are really getting even better.

366
00:26:59,040 --> 00:27:01,560
Again, is this quite production quality?

367
00:27:01,560 --> 00:27:04,120
Not for a brand, I don't think.

368
00:27:04,120 --> 00:27:09,360
The videos that you can find on the Twittersphere are very interesting and some of them are

369
00:27:09,360 --> 00:27:15,080
really quite good, but not, but they still have that kind of weird style that we're seeing

370
00:27:15,080 --> 00:27:18,400
from these video generators, but they're getting there.

371
00:27:18,400 --> 00:27:21,680
So definitely something to keep an eye on.

372
00:27:21,680 --> 00:27:27,080
And then the last story this week is Stability AI has released two new open source large

373
00:27:27,080 --> 00:27:31,240
language models, FreeWheelie 1 and FreeWheelie 2.

374
00:27:31,240 --> 00:27:36,040
These models excel in reasoning and understanding linguistic subtleties and they've been validated

375
00:27:36,040 --> 00:27:38,400
through various benchmarks.

376
00:27:38,400 --> 00:27:41,600
Despite being trained on a smaller dataset compared to previous models, the FreeWheelie

377
00:27:41,600 --> 00:27:46,920
models demonstrate exceptional performance with FreeWheelie 2 even outperforming GPT-4

378
00:27:46,920 --> 00:27:50,320
in some areas and GPT-3 in most.

379
00:27:50,320 --> 00:27:55,080
The release of these models comes off the back of MetaSlammer 2 knowledge?

380
00:27:55,080 --> 00:27:56,080
Knowledge?

381
00:27:56,080 --> 00:27:57,320
Release?

382
00:27:57,320 --> 00:28:01,600
The release of these models comes off the back of MetaSlammer 2, news that we mentioned

383
00:28:01,600 --> 00:28:02,600
last week.

384
00:28:02,600 --> 00:28:06,000
And it's a really big win for the open source community and it's going to probably drive

385
00:28:06,000 --> 00:28:12,640
even faster developments in the AI space as developers of products and tools can get access

386
00:28:12,640 --> 00:28:18,400
to much more moldable, high quality large language models that they can mold and change

387
00:28:18,400 --> 00:28:24,400
and use in ways that's just not possible with some of the closed models like ChatGPT-4 and

388
00:28:24,400 --> 00:28:28,480
other models from businesses like Anthropic.

389
00:28:28,480 --> 00:28:33,240
So again, I think we're going to see the emergence of some really cool tools here off the back

390
00:28:33,240 --> 00:28:38,960
of this that are going to have great power and hopefully provide us as marketers with

391
00:28:38,960 --> 00:28:42,120
even more tools for us to use in our businesses.

392
00:28:42,120 --> 00:28:45,880
So with that, that's the summary of this week's news.

393
00:28:45,880 --> 00:28:48,720
Next week, Martin and I will be back on this as normal.

394
00:28:48,720 --> 00:28:54,160
And what I'm going to do now is hand over to Martin to give us that update from what

395
00:28:54,160 --> 00:28:58,320
looked like a really fascinating conference at MECON this week.

396
00:28:58,320 --> 00:29:01,600
Hi, Paul, and hello listeners.

397
00:29:01,600 --> 00:29:07,920
This is Martin coming to you all the way from Cleveland, Ohio, where I've been at the MECON

398
00:29:07,920 --> 00:29:13,720
Marketing Artificial Intelligence Institute's conference, learning and networking with some

399
00:29:13,720 --> 00:29:19,420
of the best and brightest minds in marketing and AI.

400
00:29:19,420 --> 00:29:25,400
It's been an eye-opening event, some really interesting discussions, broad ranging everything

401
00:29:25,400 --> 00:29:35,440
from what's the state of AI adoption to how can service providers such as marketing agencies

402
00:29:35,440 --> 00:29:40,400
change their proposition and the way that they deliver their services to clients through

403
00:29:40,400 --> 00:29:48,880
to really practical sessions looking at how you can write better and more effective plumps

404
00:29:48,880 --> 00:29:56,280
and also covering some really important topics like how do we use AI ethically and responsibly.

405
00:29:56,280 --> 00:29:58,840
There were some fantastic speakers.

406
00:29:58,840 --> 00:30:02,720
They really did pull together some of the brightest and the best.

407
00:30:02,720 --> 00:30:11,480
So we had the head of marketing, Jasper, Megan Keeney-Anderson, my favorite presenter, Cassie

408
00:30:11,480 --> 00:30:17,060
Kozakov, who's the chief decision scientist from Google.

409
00:30:17,060 --> 00:30:19,400
She was keynoting today.

410
00:30:19,400 --> 00:30:26,840
There was a fantastic broad ranging fireside chat between Paul Rateser, who is the CEO

411
00:30:26,840 --> 00:30:36,400
of the Marketing AI Institute, and he sat down with Ethan Molyk, who's a professor of

412
00:30:36,400 --> 00:30:40,920
innovation and entrepreneurship at Wharton University.

413
00:30:40,920 --> 00:30:43,600
Their conversation was truly eye-opening.

414
00:30:43,600 --> 00:30:49,680
Ethan Molyk, he's got some fantastic insights into where this technology is heading.

415
00:30:49,680 --> 00:30:56,880
He's clearly well-networked in terms of the people building the AI systems as it is.

416
00:30:56,880 --> 00:31:04,960
He is of the opinion that this technology is going to see exponential growth in terms

417
00:31:04,960 --> 00:31:11,480
of capabilities and that we really need to sit up and pay attention because people are

418
00:31:11,480 --> 00:31:16,800
really underestimating the speed at which the development is going to come and the implications

419
00:31:16,800 --> 00:31:20,680
that we will all face.

420
00:31:20,680 --> 00:31:27,280
Broadly speaking, he was optimistic, but there was no doubt that he thinks that people's

421
00:31:27,280 --> 00:31:35,160
jobs are going to be impacted and that we should get using these tools now to get a

422
00:31:35,160 --> 00:31:37,400
bit of a head start.

423
00:31:37,400 --> 00:31:44,480
There was reference to an interesting concept that he came up with in one of his blogs recently,

424
00:31:44,480 --> 00:31:48,840
which was this idea of the button.

425
00:31:48,840 --> 00:31:56,760
He said that when the button exists, which is to say that when things like Microsoft

426
00:31:56,760 --> 00:32:01,560
Copilot have that little assistant at the side where you can just press a button and

427
00:32:01,560 --> 00:32:06,160
it will create the thing, and it will create something super, super quick.

428
00:32:06,160 --> 00:32:12,840
The example he gave was he is asked regularly to write a recommendation letter for a student.

429
00:32:12,840 --> 00:32:19,880
It might be someone going for a job and they've written to him as he was their professor and

430
00:32:19,880 --> 00:32:23,880
said, would you write me a recommendation letter?

431
00:32:23,880 --> 00:32:29,760
He says he does this regularly for his ex-students and each one will typically take half an hour

432
00:32:29,760 --> 00:32:36,720
to 45 minutes for him to write, but they've written from him personally.

433
00:32:36,720 --> 00:32:46,440
Now the other day he created a similar, well he wrote one of these using ChatGPT and it

434
00:32:46,440 --> 00:32:49,360
took him seconds.

435
00:32:49,360 --> 00:32:56,200
Now the actual letter was still guided by him, he gave him the context so that it was

436
00:32:56,200 --> 00:33:06,880
genuine but it didn't have the same human, what's the word, love, want, effort.

437
00:33:06,880 --> 00:33:11,160
But the value to the end user, to the customer, in this case the person applying for the job

438
00:33:11,160 --> 00:33:15,240
is the same and if it helps them get the job, who cares that it took five seconds rather

439
00:33:15,240 --> 00:33:16,800
than 45 minutes.

440
00:33:16,800 --> 00:33:20,960
But he gives the example of this in the workplace, when you can just push a button and get a

441
00:33:20,960 --> 00:33:26,800
thing done, what does that do to the value of work and the work that we put together?

442
00:33:26,800 --> 00:33:34,880
If we're writing reports and we can just hit the button, do we become lazy, does this devalue

443
00:33:34,880 --> 00:33:40,640
work that might require human input?

444
00:33:40,640 --> 00:33:46,840
And there was no definitive answer there, it was just an interesting thought experiment.

445
00:33:46,840 --> 00:33:54,080
There was a great session on prompt engineering from Jim Stern who does a lot of work on digital

446
00:33:54,080 --> 00:33:59,800
marketing analytics, he's published many books on the topic but yeah he gave some great advice

447
00:33:59,800 --> 00:34:07,920
on prompt design and how you can build frameworks and yeah I think one of the big things that

448
00:34:07,920 --> 00:34:13,600
he, big takeaways for me with what he said was that actually we're at the very start

449
00:34:13,600 --> 00:34:18,640
of all of this, so there's no right or wrong answer.

450
00:34:18,640 --> 00:34:25,860
If you can find something that works, great, share that knowledge, people are always finding

451
00:34:25,860 --> 00:34:26,860
new ways.

452
00:34:26,860 --> 00:34:30,520
Now he gave some general pointers, things that anyone listening to this podcast probably

453
00:34:30,520 --> 00:34:35,380
knows already such as the more context you put in the front the better output you get,

454
00:34:35,380 --> 00:34:41,520
if you give the AI a persona it's less likely to go kind of wandering off and giving you

455
00:34:41,520 --> 00:34:46,280
horrible responses or something like that, it kind of stays on script.

456
00:34:46,280 --> 00:34:55,120
But yeah by and large he was giving some good practical uses or practical examples of how

457
00:34:55,120 --> 00:34:59,000
you can prompt better.

458
00:34:59,000 --> 00:35:06,040
Now I could talk about various other talks and things all day, there's so much to cover

459
00:35:06,040 --> 00:35:10,360
over the next few weeks, I'm hoping to get some interviews lined up with some of the

460
00:35:10,360 --> 00:35:16,160
people that I've met here, it's a great network of people and some great vendors and looking

461
00:35:16,160 --> 00:35:20,480
forward to bringing them into the podcast and introducing them to the artificially intelligent

462
00:35:20,480 --> 00:35:24,200
marketing community.

463
00:35:24,200 --> 00:35:30,880
So with that I will say farewell, it's half past five on a Friday, it's time for me to

464
00:35:30,880 --> 00:35:35,320
go to a Cleveland bar and drink a nice hazy beer.

465
00:35:35,320 --> 00:35:38,760
Thank you for listening to Artificially Intelligent Marketing.

466
00:35:38,760 --> 00:35:44,800
To stay on top of the latest trends, tips and tools in the world of marketing AI, be

467
00:35:44,800 --> 00:35:46,560
sure to subscribe.

468
00:35:46,560 --> 00:36:14,200
We look forward to seeing you again next week.